Paper deep dive

Social-R1: Towards Human-like Social Reasoning in LLMs

Jincenzi Wu, Yuxuan Lei, Jianxun Lian, Yitian Huang, Lexin Zhou, Haotian Li, Xing Xie, Helen Meng

Year: 2026Venue: arXiv preprintArea: cs.AIType: PreprintEmbeddings: 112

Intelligence

Status: succeeded | Model: google/gemini-3.1-flash-lite-preview | Prompt: intel-v1 | Confidence: 94%

Last extracted: 3/13/2026, 1:00:41 AM

Summary

Social-R1 is a reinforcement learning framework designed to improve social intelligence in LLMs by aligning reasoning trajectories with human cognitive principles. It introduces ToMBench-Hard, an adversarial benchmark to expose shortcut learning, and utilizes a multi-dimensional reward system (SIP-guided) to enforce structured, evidence-grounded, and efficient social inference.

Entities (5)

Social-R1 · framework · 100%ToMBench-Hard · benchmark · 100%Qwen3-4B · model · 95%SIP · theory · 95%Reasoning Parasitism · failure-mode · 90%

Relation Signals (4)

Social-R1 → guidedby → SIP

confidence 95% · guided by Social Information Processing (SIP) theory, we develop a multi-dimensional reward system

Social-R1 → utilizes → ToMBench-Hard

confidence 95% · Building on this, we propose Social-R1... we introduce ToMBench-Hard

ToMBench-Hard → exposes → Reasoning Parasitism

confidence 90% · ToMBench-Hard, a diagnostic benchmark specifically designed to expose the shortcut learning behaviors

Social-R1 → improves → Qwen3-4B

confidence 90% · We apply the Social-R1 framework to two open-source backbones... Qwen3-4B

Cypher Suggestions (0)

No Cypher suggestions yet.

Abstract

Abstract:While large language models demonstrate remarkable capabilities across numerous domains, social intelligence - the capacity to perceive social cues, infer mental states, and generate appropriate responses - remains a critical challenge, particularly for enabling effective human-AI collaboration and developing AI that truly serves human needs. Current models often rely on superficial patterns rather than genuine social reasoning. We argue that cultivating human-like social intelligence requires training with challenging cases that resist shortcut solutions. To this end, we introduce ToMBench-Hard, an adversarial benchmark designed to provide hard training examples for social reasoning. Building on this, we propose Social-R1, a reinforcement learning framework that aligns model reasoning with human cognition through multi-dimensional rewards. Unlike outcome-based RL, Social-R1 supervises the entire reasoning process, enforcing structural alignment, logical integrity, and information density. Results show that our approach enables a 4B parameter model to surpass much larger counterparts and generalize robustly across eight diverse benchmarks. These findings demonstrate that challenging training cases with trajectory-level alignment offer a path toward efficient and reliable social intelligence.

PDF

Open source PDF →Open local PDF →

Trouble viewing inline? Open PDF directly →

Full Text

111,523 characters extracted from source content.

Expand or collapse full text

Social-R1: Towards Human-like Social Reasoning in LLMs Jincenzi Wu 1 Yuxuan Lei 2 Jianxun Lian 2 Yitian Huang 2 Lexin Zhou 3 Haotian Li 2 Xing Xie 2 Helen Meng 1 Abstract While large language models demonstrate remark- able capabilities across numerous domains, so- cial intelligence—the capacity to perceive social cues, infer mental states, and generate appropriate responses—remains a critical challenge, particu- larly for enabling effective human-AI collabora- tion and developing AI that truly serves human needs. Current models often rely on superficial patterns rather than genuine social reasoning. We argue that cultivating human-like social intelli- gence requires training with challenging cases that resist shortcut solutions. To this end, we in- troduce ToMBench-Hard, an adversarial bench- mark designed to provide hard training examples for social reasoning. Building on this, we propose Social-R1, a reinforcement learning framework that aligns model reasoning with human cogni- tion through multi-dimensional rewards. Unlike outcome-based RL, Social-R1 supervises the en- tire reasoning process, enforcing structural align- ment, logical integrity, and information density. Results show that our approach enables a 4B pa- rameter model to surpass much larger counter- parts and generalize robustly across eight diverse benchmarks. These findings demonstrate that challenging training cases with trajectory-level alignment offer a path toward efficient and reli- able social intelligence. 1. Introduction Recent advances in Reinforcement Learning from Verifi- able Feedback (RLVF) (Wen et al., 2025) have significantly improved large language models (LLMs)’ performance on formal reasoning tasks like mathematics and programming. However, genuine social intelligence—the capacity to per- ceive subtle cues, infer latent mental states, and navigate 1 The Chinese University of Hong Kong 2 Microsoft Research Asia 3 Princeton University. Correspondence to: Jianxun Lian <jianxun.lian@microsoft.com>. Preprint. March 11, 2026. complex interpersonal dynamics—remains a substantial challenge. Unlike formal reasoning with deterministic exe- cution paths, social reasoning is inherently polysemic and context-dependent, lacking the traceable verification that enables effective reward modeling in objective domains. Despite strong performance on standard benchmarks, cur- rent LLMs often rely on shortcut learning rather than au- thentic social reasoning. Models frequently exhibit a fail- ure mode we term Reasoning Parasitism, characterized by Answer-driven Backfilling—retroactively constructing jus- tifications for predetermined answers rather than deriving inferences through narrative analysis, as illustrated by the example in Figure 1. The inherent fragility of this parasitic performance becomes particularly evident in adversarial or out-of-distribution scenarios. As highlighted by Shapira et al. (2024); Pang et al. (2025), models that excel on stan- dard benchmarks often suffer catastrophic failures when confronted with trivial narrative perturbations. It reveals that current approaches create only a facade of social intelli- gence while failing to develop robust reasoning capabilities. Our analysis further identifies a critical Interpretation Bot- tleneck: while models can perceive surface-level social cues, they struggle to map these cues to latent mental states, lead- ing to a “logic reversal” where final answer correctness exceeds the logical integrity of the reasoning process. We argue that advancing social intelligence requires align- ing models’ reasoning trajectories with the structured stages of human social inference. Thus, we propose process-based trajectory alignment that cultivates social intelligence as an internalized capability rather than a parasitic perfor- mance. Human social reasoning is characterized by high- density information distillation and recursive belief model- ing—properties we aim to instill through structured supervi- sion of the reasoning process, as illustrated in the SocialR1- 8B example in Figure 1. In this paper, we introduce Social-R1, a reinforcement learn- ing framework that encourages genuine social reasoning by aligning model trajectories with human cognitive princi- ples: structured, evidence-grounded, and efficient inference. Our approach has two key components. First, we construct ToMBench-Hard, an expert-curated adversarial benchmark designed to expose reasoning shortcuts through socially 1 arXiv:2603.09249v1 [cs.AI] 10 Mar 2026 Social-R1: Towards Human-like Social Reasoning in LLMs Deepseek-R1 [Story] Isabella is passionate about eco-tourism... Her elderly mother Ethan prefers urban exploration visiting museums ... Isabella invites Ethan on a trip. [Question] Where does Isabella most likely invite Ethan to go? [Options] A. Rainforest trek B. Museum and heritage site tour C. Bird-watching expedition D. Bungee jumping [Answer] B. Museum and heritage site tour ❌ Thinking Isabella is the one inviting Ethan, her choice would align with her own interests rather than her mother’s. Option B reflects ... Option D is unrelated to ... Between A and C both fit ... The question asks where... which strongly points to A. C is a specific activity but less directly... The most likely answer is A. Shortcut Social-R1 Example: Human-like Social Reasoning Result : Accuracy & Efficiency SocialR1-8B Reasoning Parasitism Human-like Social Reasoning Cue Encoding Cue Interpreting Goal Clarification Response Generation ✔ Thinking The story presents two contrasting travel preferences ... The key detail is that Isabella invites Ethan on a trip... She might be trying to accommodate both her own interests and her mother's ... A museum and heritage site tour aligns with the mother's preference ... also allows Isabella to... The invitation is most likely to a museum and heritage site tour. Combined Rewards + RL    퐀 퐀 퐀 퐀 퐀 퐀ℎ Surface-Detail Drift Figure 1. Social-R1 for Human-like and Efficient Social Reasoning. By integrating SIP-guided rewards into reinforcement learning, Social-R1 mitigates reasoning shortcuts and enforces structured human-like social inference, improving both accuracy and efficiency across model scales. Detailed cases are in Appendix 6. nuanced perceptual traps. Second, guided by Social In- formation Processing (SIP) theory (Salancik & Pfeffer, 1978) , we develop a multi-dimensional reward system that enforces stage-consistent reasoning progression (R struct ), content integrity (R content ), and inference efficiency (R len ). We demonstrate that this trajectory-level alignment enables more effective and parameter-efficient social intelligence. We conduct comprehensive experiments across both in- domain and out-of-domain benchmarks related to social intelligence. Results demonstrate that our approach substan- tially improves model performance on genuine social reason- ing tasks. Social-R1 enables smaller models to match or sur- pass the capabilities of models with significantly larger pa- rameter counts, while showing strong generalization across eight diverse social reasoning evaluations (as illustrated in the right part of Figure 1). These findings indicate that trajectory-level alignment presents a viable and efficient alternative to reliance on pure model scaling for achiev- ing robust social intelligence. The source code and dataset will be released upon paper acceptance. In summary, our contributions are threefold: •ToMBench-Hard: A rigorous benchmark with difficulty for social reasoning that exposes shortcut learning in LLMs and mandates genuine cognitive engagement. •Social-R1 Framework: A reinforcement learning ap- proach with multi-dimensional rewards that align LLM reasoning trajectories with human social cognition. •Performance Superiority: Demonstration that our method enables small models to achieve large-model per- formance, proving trajectory quality surpasses parameter scaling for social intelligence. 2. Related Work 2.1. Theory-of-Mind Analysis in LLMs Recent evaluations demonstrate LLMs’ growing proficiency on Theory-of-Mind tasks, with models like GPT-4 achieving near-human performance on false-belief tasks and various social cognition assessments (Kosinski, 2024; Street et al., 2024; Strachan et al., 2024). However, broader audits reveal significant limitations in genuine social reasoning capabil- ities. Studies highlight inconsistent performance across different evaluation frameworks (Gandhi et al., 2023), sys- tematic failure modes in comprehensive capability assess- ments (Chen et al., 2024), and particular weaknesses in psychologically complex scenarios compared to physical- world contexts (Xu et al., 2024; Gu et al., 2024; Zhou et al., 2023). Dynamic evaluation settings further expose mod- els’ difficulties with evolving mental states over time (Xiao et al., 2025), while higher-order reasoning tasks reveal sharp performance drops beyond first-level inference (He et al., 2023). Current improvement attempts remain limited in scope. Prompt-based interventions (Gu et al., 2024) of- 2 Social-R1: Towards Human-like Social Reasoning in LLMs fer task-specific benefits but lack generalizability, while reinforcement learning approaches (Lu et al., 2025) show promise but risk overfitting to narrow task patterns without broader training diversity. From this literature, we identify three key limitations moti- vating our work: (i) over-reliance on outcome-based evalua- tion rather than reasoning trajectory analysis; (i) suscepti- bility to shallow statistical cues instead of recursive mental- state inference; and (i) limited generalization across social reasoning benchmarks and related tasks (Sabour et al., 2024; Yong et al., 2025). These gaps necessitate harder adversar- ial evaluations, trajectory-level supervision, and training regimes designed for transferable social intelligence. 2.2. RL for LLM General Reasoning Reinforcement learning has emerged as a pivotal technique for enhancing LLM reasoning capabilities, with RLHF establishing the foundation for alignment (Ouyang et al., 2022). Recent large-scale implementations demonstrate that outcome-based rewards can significantly strengthen struc- tured reasoning in mathematical and coding domains (Jaech et al., 2024; Guo et al., 2025; Team et al., 2025; Yang et al., 2025). To address sparse reward challenges, process super- vision methods provide denser feedback through step-wise trajectory evaluation (Lightman et al., 2023; Uesato et al., 2022; Zhang et al., 2025). Despite these advances, RL remains underexplored for so- cial inference tasks, which require empathy and multi-factor reasoning in unstructured scenarios. Preliminary work (Lu et al., 2025) shows potential but remains in early develop- ment. This highlights two critical gaps: (i) the need for refreshed training materials tailored to social reasoning, and (i) the opportunity for trajectory-level rewards that guide models toward human-like social inference patterns. 3. The ToMBench-Hard Benchmark Social intelligence remains a persistent bottleneck in LLMs, with current systems often relying on shortcut learning rather than authentic social reasoning. To address this limi- tation, we introduce ToMBench-Hard, a diagnostic bench- mark specifically designed to expose the shortcut learning behaviors prevalent in current LLMs. The primary goal of ToMBench-Hard is to disentangle genuine social infer- ence from surface-level pattern matching by creating a chal- lenging adversarial environment that mandates structured, human-like reasoning processes. 3.1. Adversarial Data Construction ToMBench-Hard is grounded in the Abilities in the Theory- of-Mind Space (ATOMS) framework (Osterhaus & Bosacki, 2022), providing a comprehensive evaluation covering six core dimensions of social intelligence: Belief, Desire, Emo- tion, Intention, Knowledge, and Non-literal Communication (NLC). We developed 800 expert-annotated multiple-choice questions through a rigorous annotation protocol involving three independent annotators to ensure cognitive validity and high quality (see Appendix B for details). To mitigate heuristic-based reasoning strategies such as lexical overlap between questions and options, we intro- duced ToM-consistent adversarial perturbations (Ullman, 2023; Hu et al., 2025). These perturbations include nuanced manipulations of perceptual access (e.g., unobserved state changes) and asymmetric information (e.g., second-order beliefs), ensuring models cannot succeed through statisti- cal guessing alone. Each sample was carefully designed to require a structured, human-like reasoning process that progresses from cue encoding to mental state interpretation. 3.2.Diagnostic Analysis: Exposing the Shortcut Illusion To validate the diagnostic utility of ToMBench-Hard, we conducted comprehensive benchmarking across a spectrum of state-of-the-art LLMs, comparing their performance against human expert baselines. The results reveal a signifi- cant performance cliff between simple synthetic benchmarks and our adversarial evaluation. As shown in Table 1, while human experts achieve robust 87% accuracy, frontier models exhibit dramatic performance drops. Notably, models like O3 and Deepseek-R1 achieve near-human performance on simpler benchmarks like ToM- RL (Lu et al., 2025) (reaching 87-88% accuracy), but their accuracy plummets to below 61% on ToMBench-Hard. This performance gap unmasks what we term the shortcut illu- sion—the phenomenon where high scores on conventional benchmarks reflect template matching artifacts rather than genuine social reasoning capabilities. The substantial performance gap between ToMBench-Hard and ToM-RL suggests that high scores on the latter may not reliably indicate strong theory-of-mind capabilities, as these benchmarks might be insufficiently challenging for LLMs. By systematically exposing these limitations, ToMBench- Table 1. Results on ToMBench-Hard (ours) and ToM-RL (public). ToMBench-HardToM-RL ModelBel. Des. Emo. Int. Kno. NLC All Human0.84 0.89 0.91 0.90 0.85 0.94 0.89– DeepSeek-R1 0.46 0.58 0.73 0.57 0.53 0.79 0.610.87 O30.54 0.58 0.83 0.38 0.50 0.71 0.590.88 GPT-50.55 0.54 0.77 0.26 0.51 0.71 0.560.87 Qwen3-32B0.48 0.52 0.69 0.34 0.47 0.65 0.520.71 Qwen3-8B0.21 0.25 0.55 0.38 0.14 0.50 0.340.73 3 Social-R1: Towards Human-like Social Reasoning in LLMs Hard serves dual purposes: it provides a rigorous evaluation framework for assessing genuine social reasoning capabil- ities, while also offering a high-quality dataset for align- ing LLMs with human-like cognitive processes. To facili- tate both training and evaluation, we partition the complete ToMBench-Hard dataset into non-overlapping training and test subsets, enabling the development of our alignment framework in the subsequent sections. 4. The Social-R1 Framework The Social-R1 framework introduces a novel paradigm for cultivating genuine social intelligence in LLMs by align- ing their reasoning processes with human social cognition. Departing from traditional outcome-based reinforcement learning approaches, our method specifically tackles the fundamental challenge of Reasoning Parasitism—a phe- nomenon where models engage in Answer-driven Backfill- ing by constructing post-hoc justifications for predetermined answers. Social-R1 counteracts these shortcut behaviors through a multi-dimensional reward system that supervises the entire reasoning trajectory rather than merely rewarding final outcomes. This comprehensive approach transforms social intelligence from a parasitic performance into an in- ternalized capability, ensuring model reasoning embodies the core characteristics of human social cognition: precise, stage-consistent social inference (Salancik & Pfeffer, 1978) and high information density (Sperber & Wilson, 1986). 4.1. Multi-Dimensional Reward Design The core innovation of Social-R1 lies in its comprehensive reward system that aligns model reasoning trajectories with three key characteristics of human social cognition: struc- tured progression, logical integrity, and information density. This multi-dimensional approach provides fine-grained su- pervision over the reasoning process, addressing different aspects of high-quality social inference. SIP Structural Alignment (R struct ). Human social rea- soning follows a disciplined progression from perception to response generation (Crick & Dodge, 1994). To instill this cognitive scaffold, we introduce a structural reward R struct that enforces sequential reasoning across the four stages of Social Information Processing (SIP): (1) Encod- ing Social Cues: Identifying relevant social signals from the narrative, (2) Interpreting Cues: Inferring latent mental states from perceived cues ,(3) Clarifying Goals: Deter- mining social objectives and interpersonal intentions, and (4) Response Generation: Selecting appropriate behavioral responses. TheR struct reward validates whether interme- diate reasoning steps adhere to this stage-wise progression, penalizing premature conclusions and stage skipping. This structural constraint encourages coherent, story-grounded causal inference while mitigating shortcut behaviors like op- tion parasitism, where models anchor reasoning directly on multiple-choice options rather than narrative evidence. We leverage GPT-4o as judge for this reward, with the prompt illustrated in Appendix C.1. SIP Content Integrity (R content ). While structural align- ment ensures proper reasoning staging, content integrity guarantees logical rigor within each stage. The content rewardR content audits whether intermediate inferences re- main grounded in story-internal evidence, correctly reflect- ing social cues, intentions, and goals. This reward penal- izes three critical failure modes: (1) Erroneous cue encod- ing: Misidentification of relevant social signals, (2) Flawed interpretation: Incorrect mental state attribution, and (3) Misidentified goals: Inaccurate inference of social objec- tives. By ensuring each reasoning step maintains evidential support,R content discourages superficial rationalizations and promotes authentic social sensing derived from narra- tive context. The detailed implementation of this reward model can be found at Section 5.1. Inference Efficiency Optimization (R len ). Human so- cial reasoning achieves high information density through selective attention and avoidance of redundant processing (Simon, 1955; Gigerenzer & Goldstein, 1996). To emulate this cognitive efficiency, we designR len as the product of two complementary components: R len = R rep (ρ)· R win (L)(1) The repetition penalty componentR rep (ρ)specifically tar- gets circular over-thinking by penalizing excessive n-gram repetition beyond a threshold τ = 0.1: R rep (ρ) = ( 1,ρ≤ τ exp(−β(ρ− τ)), ρ > τ (2) whereβ = 8controls the severity of penalty decay for high repetition ratios. The length window constraintR win (L) maintains reasoning trajectories within an empirically opti- mal range [L min ,L max ] through smooth gating: R win (L) = σ L− L min k · σ L max − L k (3) withk = 50controlling transition smoothness. The bounds L min = 400andL max = 2500are derived from strong chain-of-thought baselines, ensuring reasoning remains con- cise yet comprehensive. This dual-mechanism approach encourages the model to emulate human-like efficiency by avoiding both redundant repetition and excessive verbosity while maintaining substantive social inference. Verifiable Format Alignment (R fmt ) We adopt the for- mat reward from Shao et al. (2024) that enforces structured thinking processes. The model is rewarded for producing 4 Social-R1: Towards Human-like Social Reasoning in LLMs outputs with predefined XML-style tags (<thinking> and<answer>), which enables deterministic extraction of both reasoning trajectories and final answers while preserv- ing semantic freedom in the reasoning content. 4.2. Reward Synthesis and Learning The composite reward function integrates all components through a carefully designed synthesis strategy that balances outcome supervision with process-level reasoning signals: R total = R fmt · w o R out + τ w struct R struct + w content R content · R len (4) whereR out denotes the verifiable outcome reward, We implement a curriculum learning strategy where outcome supervision dominates early training phases (w o (t) = 2 ), while process-level rewards are progressively empha- sized through time-dependent weighting:w struct (t) = w content (t) = 1 + γ t T . This curriculum ensures stable initial convergence while gradually reinforcing human-like reasoning patterns as training progresses. The optimization employs Group Relative Policy Optimization (Shao et al., 2024), which performs group-relative updates over sampled reasoning trajectories. 5. Experiment 5.1. Experiment Setting We evaluate our model on eight multiple-choice social benchmarks, including two in-domain benchmarks: the public ToMBench (Chen et al., 2024) and our ToMBench- Hard test set, and six out-of-domain benchmarks: So- cialIQA (Sap et al., 2019) for social commonsense reason- ing, EmoBench (Sabour et al., 2024) for emotion intelli- gence, MotiveBench (Yong et al., 2025) for social motiva- tion reasoning, SimpleToM (Gu et al., 2024) for examining whether models can consciously infer others’ mental states (MS) and proactively applying such reasoning to behavior inference, Hi-ToM (Wu et al., 2023) for evaluating higher- order Theory-of-Mind reasoning, and TactfulToM (Liu et al., 2025) for testing whether models can interpret white lies and infer the underlying prosocial intent to preserve inter- personal harmony. SIP Content Integrity Reward Model (RM content ).To instantiateR content , we train a dedicated Content Reward Model that assigns a scalar quality score to intermediate SIP- stage reasoning segments. We construct the SocialPairs- 20K dataset by samplingNtrajectories per training instance from various Social-R1 checkpoints. These checkpoints span different training stages, ensuring a diverse pool of reasoning qualities ranging from nascent logic to sophisti- cated social inference. For each segment, a strong teacher- judge (o3) generates silver-standard stage-wise rationales by conditioning on the gold final answer and the original social context. We then employ a multi-dimensional rubric— focusing on fact-grounding, mental-state attribution accu- racy, and stage-specific relevance—to score the sampled segments against the teacher-generated references. This process forms preference pairs (chosen vs. rejected), while rejected segments exhibit missing cues, incorrect mental- state attributions, or wrong goal identification. We validate the model on two held-out sets: (1) an automatic test split of 2k pairs, whereR content achieves an accuracy of 89.2%; and (2) a human-calibrated subset of 200 pairs annotated by experts, achieving an 87.5% agreement with human labels. ( More R content can be found in Appendix C.2). Implementation Details SIP Content Integrity Reward Model (RM content ) is trained as a pairwise preference re- ward model. We initialize it from Qwen3-4B and fine-tune with LoRA on the SocialPairs-20K dataset mentioned pre- viously. For policy optimisation, we train two versions of reasoning models from Qwen3-4B and Qwen3-8B on ToMBench-Hard, consisting of 700 training instances and 100 test instances. Reinforcement learning is performed for 600 optimisation steps using VERL (Sheng et al., 2024) on 8 NVIDIA A100 (80GB) GPUs. We set the group size to 5, the KL coefficient to 0.04, and the learning rate to5× 10 −7 . More details are provided in Appendix D. 5.2. Main Results We apply the Social-R1 framework to two open-source back- bones of different scales, Qwen3-4B and Qwen3-8B. The overall results across eight social reasoning benchmarks are reported in Table 2. Social-R1 achieves strong and consis- tent improvements over the corresponding Qwen baselines, demonstrating that reinforcement learning with challeng- ing social supervision and trajectory-level reward signals can substantially enhance Theory-of-Mind reasoning. No- tably, Social-R1-4B surpasses LLaMa3.1-70B across all benchmarks, despite being more than an order of magni- tude smaller, highlighting the effectiveness of process-based social alignment beyond parameter scaling. Even more strikingly, Social-R1-8B outperforms DeepSeek-R1 on sev- eral benchmarks and achieves stronger overall performance, consistently matching or exceeding much larger baselines such as Qwen3-32B in out-of-domain generalization. More detailed results are provided in Appendix E. 5.3. Ablation Studies To systematically evaluate the contribution of each reward component in Social-R1, we conduct comprehensive abla- tion experiments on both SocialR1-4B and SocialR1-8B variants. Specifically, we remove three key components: 5 Social-R1: Towards Human-like Social Reasoning in LLMs Table 2. In-domain and out-of-domain performance across eight social reasoning benchmarks. Green denotes the best result among our Social-R1 reward variants (ablations), and bold indicates the overall best score. Model In-domainOut-of-domain ToMBench ToMBench- Hard Val SocialIQA SimpleToM EmoBench MotiveBench Hi-ToM TactfulToM Overall Closed-sourced LLMs DeepSeek-R10.68340.44360.80600.71870.73870.86550.70590.69690.7073 O30.78340.44780.82910.76400.80390.94000.70330.68600.7447 O3 COT0.74020.48200.79630.75010.72250.92000.66500.65440.7163 O3+MS–0.9672– GPT-50.78230.40920.82690.73550.80300.90890.63170.46700.6956 GPT-5 COT0.78300.42720.81630.71110.78820.92670.67500.65470.7228 GPT-5+MS–0.9924– GPT-4o0.73780.37330.78400.66610.76940.93000.60330.52670.6738 GPT-4oCOT0.74060.40120.79530.61260.77620.89780.60500.52810.6696 GPT-4o+MS–0.7358– Open-sourced LLMs Qwen3-4B (Disable thinking)0.57700.31870.73130.51090.52800.77340.57000.40310.5516 Qwen3-4B0.63560.34030.74510.54110.53080.84440.56170.45840.5822 Qwen3-8B (Disable thinking)0.53490.38620.76000.51090.64840.79330.56000.42920.5779 Qwen3-8B0.61790.39820.77280.61760.59200.68780.55000.46550.5877 Qwen3-32B (Disable thinking)0.69170.41190.76150.64780.67000.89220.55000.51930.6431 Qwen3-32B0.69750.43640.77740.76340.63560.88000.55170.55750.6624 LLaMa3.1-70B0.60360.44710.46210.71900.66320.84450.59830.55100.6111 LLaMa3.1-70BCOT0.69560.43350.75740.59080.68910.88670.58500.55830.6496 Distill-Llama-70B0.73990.40330.80550.76290.75960.88440.53330.62020.6886 Ours SocialR1-4B only R out 0.59290.39460.73180.89630.64220.85330.52830.42650.6332 SocialR1-4B w/o R len 0.63760.47280.74510.91980.65910.83110.64000.49410.6750 SocialR1-4B w/o R struct 0.64490.44230.7201 0.94940.67740.83110.56170.50010.6659 SocialR1-4B w/o R cont 0.64490.44230.72820.89600.62510.83000.58500.46090.6516 SocialR1-4B Full0.65740.48460.75080.92850.68040.85000.65830.49360.6880 SocialR1-8B only R out 0.68500.50420.77740.94330.69680.88890.62330.48600.7004 SocialR1-8B w/o R len 0.68620.52640.75740.96190.68600.87670.62670.48720.7011 SocialR1-8B w/o R struct 0.64490.61280.73080.90410.70000.87890.70670.45580.7042 SocialR1-8B w/o R cont 0.68480.55090.73230.93520.67780.85670.69500.46670.7016 SocialR1-8B Full 0.68810.62790.77530.96750.70100.87560.70830.50790.7270 the length-control reward (w/o R len ), the structural trajec- tory reward (w/o R struct ), and the content integrity reward (w/o R cont ). Additionally, we examine a baseline variant that removes all the progress rewards (only R out ). The com- plete results are presented in Table 2. BenchmarksThe ablation analysis reveals distinct perfor- mance patterns associated with each reward component. Re- movingw/o R len results in significant performance degrada- tion on Hi-ToM (e.g.,0.7083→ 0.6267for SocialR1-8B), indicating that uncontrolled reasoning length negatively im- pacts high-order social inference. The w/oR len variant exhibits a substantial increase in reasoning verbosity, with the average thinking length rising by approximately +250% compared to SocialR1-8B as shown in Figure 10. The structural rewardR struct shows consistent importance across benchmarks, with its removal causing accuracy drops (e.g., from0.5079to0.4558on TactfulToM). Similarly, R cont removal leads to performance declines, suggesting its role in maintaining reasoning quality. Notably, theonly R out variant exhibits more severe overall performance deterio- ration compared to other ablations, demonstrating the im- portance of incorporating process-level rewards. While these numerical changes demonstrate the overall effective- ness of each component, subsequent in-depth analyses (Sec- tions 6.1–6.3) will provide mechanistic evidence through detailed trajectory examinations and perturbation studies to understand how each reward shapes reasoning behavior. 6. In-Depth Analysis To examine whether Social-R1 yields genuinely internalised social reasoning—rather than Reasoning Parasitism driven by answer-first backfilling—we conduct a mechanistic anal- ysis of its reasoning dynamics. We compare Social-R1-8B against three strong baselines (DeepSeek-R1, DeepSeek- R1-Distill-Llama-70B, and Qwen3-8B), together with con- trolled reward variants that isolate outcome-only training and ablate key trajectory-level constraints. In line with our reward design, we operationalise human-like social reasoning through three diagnostic signatures: (i) option- independent inference grounded in narrative cues; (i) stage- 6 Social-R1: Towards Human-like Social Reasoning in LLMs consistent SIP trajectories with content integrity, where intermediate beliefs remain logically valid across Encoding, Interpretation, Goal Clarification, and Response Generation; and (i) selective robustness, where models avoid redun- dant trajectory bloat under perturbation. Accordingly, our analysis proceeds in three steps: (1) quantifying Reasoning Parasitism as a diagnostic of shortcut reliance; (2) auditing stage-wise cognitive fidelity to identify interpretation bottle- necks; and (3) introducing controlled distractors to assess robustness and conciseness. 6.1. From Parasitism to Independent Inference Q1 Cue Encoding Q2 Cue Interpretation Q3 Goal Clarification Q4 Response Generation 0 2 4 6 8 10 12 Avg. Option Mentions per Sample 8.45 9.78 3.18 4.14 5.75 8.87 5.17 7.99 1.81 1.79 1.60 1.30 Social-R1 vs. Shortcut-heavy Baselines DeepseekR1 DeepSeek-R1-Distill-Llama-70B Qwen38B SocialR1-8B w/o Combined Reward SocialR1-4B SocialR1-8B Figure 2. Option-Mention Density across SIP reasoning stages. A primary failure mode in multiple-choice social reason- ing is Reasoning Parasitism—a shortcut behaviour where models anchor their deductions on statistical regularities in the answer choices rather than deriving conclusions from story-internal social cues. Figure 2 reveals substantial di- vergence in reasoning dynamics. DeepSeek-R1, Qwen3-8B, and the outcome-only variant of SocialR1-8B exhibit high option-mention density as early as the Cue Encoding stage (Q1 group in Figure 2), often exceeding five explicit option references per sample. Such premature reliance suggests answer-conditioned backfilling: models use the provided choices to retroactively assemble plausible justifications, rather than performing independent social inference. In contrast, Social-R1-8B maintains a largely option-agnostic trajectory. Its option mentions remain minimal and nearly flat throughout cue perception and interpretation, increasing only slightly (to∼1.3) during final Response Generation (Q4). This provides mechanistic evidence that the proposed Social Think Combined Reward effectively suppresses short- cut dependence on answer options, compelling the model to engage in narrative-grounded social deduction. Additional qualitative examples are provided in Appendix F. 6.2. Stage-wise Diagnosis To pinpoint where social reasoning most systematically breaks down, we conduct a human audit of 80 instances sampled from eight benchmarks, ensuring high annotation quality and genuine Theory-of-Mind demands. We track success rates across the four Social Information Processing (SIP) stages. Figure 3 exposes a pronounced Interpreta- tion Bottleneck. While strong baselines sustain high accu- racy in factual cue encoding (> 80%), performance drops sharply—by nearly 25 points—once entering Cue Interpre- tation, which requires latent mental-state attribution and recursive belief reasoning. More critically, outcome-only training induces a reasoning reversal phenomenon. For ex- ample, Qwen3-8B trained withR out often achieves higher final answer accuracy than the correctness of its preceding SIP stages, indicating that surface-level option correlations can yield correct predictions despite incoherent intermediate social inference. This highlights the limitation of outcome supervision: it rewards answer selection without enforcing psychologically valid reasoning trajectories. Crucially, the proposed Content RewardR content directly targets this failure mode by penalising erroneous cue encoding, flawed mental- state interpretation, and misidentified social goals. Ablation results provide clear causal evidence: removingR content re- duces Interpretation accuracy by 6.2 points (77.5%→71.3%) and Goal Classification by 6.2 points (75.0%→68.8%). Ul- timately, Social-R1-8B preserves the most stage-consistent and smoothly degrading SIP trajectory, demonstrating that its gains stem from cognitively grounded reasoning rather than shortcut-driven recovery. Cue Encoding Cue Interpreter Goal Classification Response Generation SIP Reasoning Stages DeepSeek-R1 SocialR1-8B SocialR1-8B w/o R cont SocialR1-8B wo R out Deepseek-R1- Distill- Llama-70B Qwen3-8B Models 91.2%77.5%77.5%73.8% 85.0%77.5%75.0%71.2% 85.0%71.3%68.8%68.8% 81.2%51.2%55.0%63.7% 81.2%56.2%65.0%55.0% 82.5%56.2%62.5%46.2% 0.4 0.5 0.6 0.7 0.8 0.9 Accuracy Figure 3. Stage-wise SIP accuracy across models. Case Study To further illustrate this mechanistic dis- tinction, Figure 4 presents a representative example from ToMBench, where the correct answer critically depends on grounding interpretation in the protagonist’s epistemic ac- cess. In this story, Grimmo inhabits an underground world without sky, celestial bodies, or human presence, and there- fore cannot plausibly imitate planets, clouds, or dancers. Social-R1-8B correctly encodes this constraint and performs a psychologically valid interpretation, inferring that the imi- tation must stem from locally observable phenomena (e.g., bioluminescent fungi spinning in the dark), yielding a co- herent SIP trajectory. In contrast, reward-ablated variants exhibit the identified failure modes. Social-R1-8B w/oR cont 7 Social-R1: Towards Human-like Social Reasoning in LLMs drifts into ungrounded interpretation, despite explicit narra- tive exclusion—a breakdown at the Interpretation and Goal stages. Likewise, the model trained with onlyR out col- lapses into option-level lexical shortcutting (e.g., “spinning dancer”) rather than maintaining story-consistent mental- state reasoning. Qwen3-8B shows a similar reversal be- haviour, selecting an answer that reflects generic associa- tions instead of the agent’s accessible knowledge. Over- all, this case provides concrete mechanistic evidence that Social-R1’s gains do not arise from answer-first recovery, but from enforcing evidence-grounded interpretation and stage-consistent social inference. [Story] In the mysterious underground world of Terra Valley lives a small, mischievous robot named Grimmo. In this realm without sky or celestial bodies, Grimmo has never seen the heavens, nor encountered a human. ...its walls adorned with bioluminescent fungi and glimmering minerals...Grimmo begins an imitation routine: extending his arms, he slowly spins on his own axis, his movements reminiscent of a spinning top or a graceful dancer twirling in place. [Question] What is Grimmo possibly imitating? [Option] A. A floating cloud B. The rotation of a planet C. A dancer’s spiral turn D. A glowing mushroom spinning in the wind [Answer] D SocialR1-8B Grimmo is a small robot who has never seen the heavens or humans. The environment has bioluminescent fungi and minerals that sparkle in the dark ...absence of celestial bodies means there's no planet rotation to imitate. ... the glowing mushrooms could be spinning in the wind...the most plausible imitation is the glowing mushrooms spinning. DeepSeek-R1 Grimmo has never seen the heavens, so he hasn't seen clouds or planets. That means options A and B involve things he couldn't have seen...It says "imitation routine," so he is imitating something. But what? .. let's see the answer choices; D is the only one that could be part of his environment. SocialR1-8B w/o 푹 푐표푛푡 ...Since there is no celestial reference and no observed movement in his environment, Grimmo’s imitation must be based on something he has experienced...the most plausible conclusion is that he is imitating a natural phenomenon he has observed, which might be the rotation of a planet, even though he hasn't seen one. SocialR1-8B 풐풏풍풚 푹 풐풖풕 First, I need to look at the options. The choices are A...B...C...D... The problem is that the question says evoking the image of a spinning top or a spinning dancer. So the answer could be either A or C. ...perhaps the correct answer is C. But I'm a bit confused. Alternatively, maybe the answer is A. Qwen3-8B He's never seen the heavens or humans. The setting has bioluminescent fungi and minerals that sparkle in the dark. Since Terra Valley has no sky or celestial bodies ... Now, the options: ... the question is about what he's imitating. Since he's in an underground world, maybe he's imitating a dancer. So the answer is C. Figure 4. Case study highlighting the Interpretation Bottleneck. Detailed cases are provided in Appendix F. 6.3. Robustness under Perturbation To assess whether Social-R1 promotes cognitively disci- plined robustness rather than brittle performance sustained by overextended reasoning, we conduct a controlled per- turbation study. Specifically, we inject story-consistent but decision-irrelevant distractor cues into 40 instances where both SocialR1-8B and DeepSeek-R1 initially succeed. These distractors introduce no new evidence that alters the correct option, enabling a direct comparison of reasoning trajectory stability (Appendix G). Figure 5 illustrates the resulting accuracy–efficiency trade-off. Although SocialR1- 8B and DeepSeek-R1 retain comparable post-perturbation accuracy, DeepSeek-R1 does so only by producing substan- tially longer reasoning trajectories, suggesting reliance on diffuse and overextended inference. In contrast, SocialR1- 8B exhibits only mild token drift, indicating more selective attention and cognitively efficient deduction. Reward abla- tions further provide mechanistic support. RemovingR struct disrupts stage-wise progression, removingR content under- mines evidence-grounded interpretation, and training with onlyR out yields the most severe robustness collapse. To- gether, these findings demonstrate that Social-R1 achieves robustness not through increased verbosity, but through en- forcing structured, grounded, and concise social reasoning, aligning model inference with human-like selective cogni- tion rather than scale-driven overthinking. 10020050010002000 Average Reasoning Tokens per Sample 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0 Accuracy (on N=40 Samples) Figure 5. Robustness study under story-consistent distractors. 7. Conclusion In this work, we introduce ToMBench-Hard, a challeng- ing benchmark that rigorously evaluates the Theory of Mind capabilities in LLMs. Building on this, we propose Social-R1, a reinforcement learning framework that inte- grates both outcome-level and thinking-level rewards to cultivate human-like social intelligence in LLMs. Our re- sults demonstrate that outcome-based reinforcement learn- ing over ToMBench-Hard already enhances social reason- ing, while thinking-level supervision yields further improve- ments. These findings highlight the importance of supervis- ing not only what a model concludes but also how it reasons, paving the way toward socially intelligent LLMs. Future work may extend this framework to broader domains of so- cial tasks, such as human-AI collaboration and LLM-based simulations for social sciencec. 8 Social-R1: Towards Human-like Social Reasoning in LLMs 8. Impact Statement This research introduces Social-R1, a framework for enhanc- ing social reasoning in LLMs, which could enable more nat- ural human-AI collaboration in applications like education, healthcare, and assistive technologies. However, improved social intelligence also raises ethical concerns, such as the potential for misuse in manipulative systems or the amplifi- cation of social biases if not properly aligned with human values. We encourage rigorous oversight and fairness audits to mitigate these risks while leveraging the benefits of robust AI social cognition. References Achiam, J., Adler, S., Agarwal, S., Ahmad, L., Akkaya, I., Aleman, F. L., Almeida, D., Altenschmidt, J., Altman, S., Anadkat, S., et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. Chen, Z., Wu, J., Zhou, J., Wen, B., Bi, G., Jiang, G., Cao, Y., Hu, M., Lai, Y., Xiong, Z., et al. Tombench: Benchmarking theory of mind in large language models. arXiv preprint arXiv:2402.15052, 2024. Crick, N. R. and Dodge, K. A. A review and reformula- tion of social information-processing mechanisms in chil- dren’s social adjustment. Psychological bulletin, 1994. Dubey, A., Jauhri, A., Pandey, A., Kadian, A., Al-Dahle, A., Letman, A., Mathur, A., Schelten, A., Yang, A., Fan, A., et al. The llama 3 herd of models. arXiv preprint arXiv:2407.21783, 2024. Gandhi, K., Fr ̈ anken, J.-P., Gerstenberg, T., and Goodman, N. Understanding social reasoning in language models with language models. Advances in Neural Information Processing Systems, 36:13518–13529, 2023. Gigerenzer, G. and Goldstein, D. G. Reasoning the fast and frugal way: models of bounded rationality. Psychological review, 1996. Gu, Y., Tafjord, O., Kim, H., Moore, J., Bras, R. L., Clark, P., and Choi, Y. Simpletom: Exposing the gap between explicit tom inference and implicit tom application in llms. arXiv preprint arXiv:2410.13648, 2024. Guo, D., Yang, D., Zhang, H., Song, J., Zhang, R., Xu, R., Zhu, Q., Ma, S., Wang, P., Bi, X., et al. Deepseek-r1: In- centivizing reasoning capability in llms via reinforcement learning. arXiv preprint arXiv:2501.12948, 2025. He, Y., Wu, Y., Jia, Y., Mihalcea, R., Chen, Y., and Deng, N. Hi-tom: A benchmark for evaluating higher-order theory of mind reasoning in large language models. arXiv preprint arXiv:2310.16755, 2023. Hu, J., Sosa, F., and Ullman, T. Re-evaluating theory of mind evaluation in large language models. Philosophical Transactions B, p. 20230499, 2025. Jaech, A., Kalai, A., Lerer, A., Richardson, A., El-Kishky, A., Low, A., Helyar, A., Madry, A., Beutel, A., Car- ney, A., et al. Openai o1 system card. arXiv preprint arXiv:2412.16720, 2024. Kosinski, M. Evaluating large language models in theory of mind tasks. Proceedings of the National Academy of Sciences, 121(45):e2405460121, 2024. Lightman, H., Kosaraju, V., Burda, Y., Edwards, H., Baker, B., Lee, T., Leike, J., Schulman, J., Sutskever, I., and Cobbe, K. Let’s verify step by step. In The Twelfth International Conference on Learning Representations, 2023. Liu, Y., Pretty, E. J., Huang, J., and Sugawara, S. Tactfultom: Do llms have the theory of mind ability to understand white lies? In Proceedings of the 2025 Conference on Em- pirical Methods in Natural Language Processing, 2025. Lu, Y.-L., Zhang, C., Song, J., Fan, L., and Wang, W. Tom- rl: Reinforcement learning unlocks theory of mind in small llms. arXiv e-prints, p. arXiv–2504, 2025. Osterhaus, C. and Bosacki, S. L. Looking for the lighthouse: A systematic review of advanced theory-of-mind tests beyond preschool. Developmental Review, 2022. Ouyang, L., Wu, J., Jiang, X., Almeida, D., Wainwright, C., Mishkin, P., Zhang, C., Agarwal, S., Slama, K., Ray, A., et al. Training language models to follow instructions with human feedback. Advances in neural information processing systems, 35:27730–27744, 2022. Pang, D. K., Pang, S. K., Broeker, M. D., and Hibble, A. Do large language models have a theory of mind? Pro- ceedings of the National Academy of Sciences, 2025. Sabour, S., Liu, S., Zhang, Z., Liu, J. M., Zhou, J., Sunaryo, A. S., Li, J., Lee, T., Mihalcea, R., and Huang, M. Emobench: Evaluating the emotional intelligence of large language models. arXiv preprint arXiv:2402.12071, 2024. Salancik, G. R. and Pfeffer, J. A social information process- ing approach to job attitudes and task design. Adminis- trative science quarterly, p. 224–253, 1978. Sap, M., Rashkin, H., Chen, D., LeBras, R., and Choi, Y. Socialiqa: Commonsense reasoning about social interac- tions. arXiv preprint arXiv:1904.09728, 2019. Shao, Z., Wang, P., Zhu, Q., Xu, R., Song, J., Bi, X., Zhang, H., Zhang, M., Li, Y., Wu, Y., et al. Deepseekmath: Push- ing the limits of mathematical reasoning in open language models. arXiv preprint arXiv:2402.03300, 2024. 9 Social-R1: Towards Human-like Social Reasoning in LLMs Shapira, N., Levy, M., Alavi, S. H., Zhou, X., Choi, Y., Goldberg, Y., Sap, M., and Shwartz, V. Clever hans or neural theory of mind? stress testing social reasoning in large language models. In Proceedings of the 18th Conference of the European Chapter of the Association for Computational Linguistics (Volume 1: Long Papers), 2024. Sheng, G., Zhang, C., Ye, Z., Wu, X., Zhang, W., Zhang, R., Peng, Y., Lin, H., and Wu, C. Hybridflow: A flexi- ble and efficient rlhf framework. arXiv preprint arXiv: 2409.19256, 2024. Simon, H. A. A behavioral model of rational choice. The quarterly journal of economics, 1955. Sperber, D. and Wilson, D. Relevance: Communication and cognition, volume 142. Harvard University Press Cambridge, MA, 1986. Strachan, J. W., Albergo, D., Borghini, G., Pansardi, O., Scaliti, E., Gupta, S., Saxena, K., Rufo, A., Panzeri, S., Manzi, G., et al. Testing theory of mind in large language models and humans. Nature Human Behaviour, 8(7): 1285–1295, 2024. Street, W., Siy, J. O., Keeling, G., Baranes, A., Barnett, B., McKibben, M., Kanyere, T., Lentz, A., Dunbar, R. I., et al. Llms achieve adult human performance on higher-order theory of mind tasks. arXiv preprint arXiv:2405.18870, 2024. Team, K., Du, A., Gao, B., Xing, B., Jiang, C., Chen, C., Li, C., Xiao, C., Du, C., Liao, C., et al. Kimi k1. 5: Scaling reinforcement learning with llms. arXiv preprint arXiv:2501.12599, 2025. Uesato, J., Kushman, N., Kumar, R., Song, F., Siegel, N., Wang, L., Creswell, A., Irving, G., and Higgins, I. Solv- ing math word problems with process-and outcome-based feedback. arXiv preprint arXiv:2211.14275, 2022. Ullman, T.Large language models fail on trivial al- terations to theory-of-mind tasks.arXiv preprint arXiv:2302.08399, 2023. Wen, X., Liu, Z., Zheng, S., Xu, Z., Ye, S., Wu, Z., Liang, X., Wang, Y., Li, J., Miao, Z., et al. Reinforcement learn- ing with verifiable rewards implicitly incentivizes correct reasoning in base llms. arXiv preprint arXiv:2506.14245, 2025. Wu, Y., He, Y., Jia, Y., Mihalcea, R., Chen, Y., and Deng, N. Hi-tom: A benchmark for evaluating higher-order theory of mind reasoning in large language models. In Find- ings of the Association for Computational Linguistics: EMNLP 2023, 2023. Xiao, Y., Wang, J., Xu, Q., Song, C., Xu, C., Cheng, Y., Li, W., and Liu, P. Towards dynamic theory of mind: Evaluating llm adaptation to temporal evolution of human states. arXiv preprint arXiv:2505.17663, 2025. Xu, H., Zhao, R., Zhu, L., Du, J., and He, Y. Opentom: A comprehensive benchmark for evaluating theory-of-mind reasoning capabilities of large language models. arXiv preprint arXiv:2402.06044, 2024. Yang, A., Li, A., Yang, B., Zhang, B., Hui, B., Zheng, B., Yu, B., Gao, C., Huang, C., Lv, C., et al. Qwen3 technical report. arXiv preprint arXiv:2505.09388, 2025. Yong, X., Lian, J., Yi, X., Zhou, X., and Xie, X. Mo- tivebench: How far are we from human-like motivational reasoning in large language models?arXiv preprint arXiv:2506.13065, 2025. Zhang, W., Chen, J., and Liu, Y.Reasonflux-prm: Trajectory-aware process rewards for chain-of-thought supervision. arXiv preprint arXiv:2506.54321, 2025. Zhou, P., Madaan, A., Potharaju, S. P., Gupta, A., McKee, K. R., Holtzman, A., Pujara, J., Ren, X., Mishra, S., Nematzadeh, A., et al. How far are large language mod- els from agents with theory-of-mind? arXiv preprint arXiv:2310.03051, 2023. 10 Social-R1: Towards Human-like Social Reasoning in LLMs A. The Detailed Case in Figure 1 Deepseek-R1 Example: Human-like Social Reasoning SocialR1-8B Thinking The story presents two contrasting travel preferences between Isabella and her mother. Isabella is into eco-tourism, specifically rainforests and wildlife observation. Her mother, on the other hand, enjoys urban exploration, focusing on museums, historic districts, and architectural landmarks. The key detail is that Isabella invites Ethan on a trip during their upcoming break. The question is about where Isabella would most likely invite Ethan. Since Isabella is passionate about eco-tourism, her primary interest lies in natural environments like rainforests. However, the story also mentions her mother's preference for urban exploration. The fact that Isabella invites Ethan suggests that she might be trying to accommodate both her own interests and her mother's. Therefore, the invitation is likely a compromise that combines both interests. The most plausible scenario is that Isabella invites Ethan on a trip that includes both elements—eco-tourism and urban exploration. This would mean a combination of visiting natural habitats and cultural landmarks. Among the options, a museum and heritage site tour aligns with the mother's preference for urban exploration. It also allows Isabella to engage in her passion for eco-tourism by visiting sites that have historical or cultural significance within urban settings. Thus, the invitation is most likely to a museum and heritage site tour. So, the answer is B. [Story] Discrepant desires,"[Story]Isabella is passionate about eco-tourism, spending time in rainforests and observing wildlife in their natural habitat. Her elderly mother, however, prefers urban exploration—visiting museums, historic districts, and architectural landmarks. She believes travel should be a journey through human history and culture. During the upcoming break, Isabella invites Ethan on a trip. [Question] Where does Isabella most likely invite Ethan to go? [Options] A. Rainforest trek B. Museum and heritage site tour C. Bird-watching expedition D. Bungee jumping [Answer] B. Museum and heritage site tour Thinking Isabella's passion for eco-tourism centers on rainforests and wildlife observation, while her mother prefers urban cultural activities. Since Isabella is the one inviting Ethan, her choice would align with her own interests rather than her mother’s. Option B (museum and heritage sites) reflects her mother’s preference, making it unlikely. Option D (bungee jumping) is unrelated to either described interest. Between A (rainforest trek) and C (bird-watching expedition), both fit eco-tourism, but the story explicitly mentions ""rainforests"" as a key location. The question asks ""where,"" emphasizing a destination, which strongly points to A. Bird-watching (C) is a specific activity but less directly tied to the explicitly mentioned rainforest setting. Thus, the most likely answer is **A**. Figure 6. The Detailed Case in Figure 1 B. ToMBenchHard ToMBenchHard is deliberately curated to increase task difficulty by introducing nuanced distractors and context- dependent reasoning. Inspired by the Abilities in the Theory- of-Mind Space (ATOMS) framework (Osterhaus & Bosacki, 2022), each question is designed to probe a distinct aspect of ToM reasoning, detailed definition of each subabilities and dimension can be found in (Osterhaus & Bosacki, 2022) . To further increase difficulty (Ullman, 2023; Hu et al., 2025), adversarial variations such as asymmetric access to information, discrepant intentions, and subtle social cues are included. B.1. Human Annotation ToMBenchHard is developed jointly by the author and one psychology graduate student, who construct the scenarios, questions, options and answers. Annotation is carried out by five computer science graduate students (after receiving training) and five social psychology graduate students. Each sample is independently answered by two annotators, and disagreements are discussed and resolved through group review and iterative modification. This procedure ensures both linguistic clarity and psychological validity. The anno- tation process emphasized consistency across dimensions and aimed to capture nuanced aspects of social reasoning. B.2. Data Statistic Table 3. Distribution of fine-grained ToMBenchHard sub-abilities. Ability (Total)Sub-abilityCount Intention (243) Prediction of actions111 Intentions explanations102 Completion of failed actions18 Discrepant intentions12 Belief (186) Second-order beliefs67 Location false beliefs49 Beliefs based action/emotions39 Identity false beliefs12 Content false beliefs11 Sequence false beliefs8 Emotion (143) Typical emotional reactions59 Atypical emotional reactions26 Mixed emotions26 Emotion regulation12 Hidden emotions12 Moral emotions8 Knowledge (96) Information-knowledge links53 Knowledge-pretend play links21 Knowledge-attention links12 Percepts-knowledge10 Desire (82) Desires influence on emotions and actions 32 Discrepant desires18 Desire-action contradiction14 Multiple desires9 Desires influence on actions6 Desires influence on emotions (beliefs) 3 Non-Literal Communication (50) Involuntary lies10 Faux pas10 Egocentric lies8 Humor8 Irony/Sarcasm8 White lies8 B.3. Cases in ToMBenchHard To provide an intuitive understanding of the challenges posed by ToMBench-Hard, Figure 7 presents representative examples spanning diverse Theory-of-Mind abilities, includ- ing belief tracking, discrepant desires, intention explanation, atypical emotional reactions, knowledge inference, and non- 11 Social-R1: Towards Human-like Social Reasoning in LLMs literal humour comprehension. Each instance follows a uni- fied Ability–Story–Question–Answer structure, consisting of a socially grounded narrative context, a multiple-choice question, and the annotated gold option. These cases high- light the benchmark’s structural diversity and nuanced social cues, requiring models to move beyond superficial option matching and motivating our trajectory-level alignment re- wards. B.4. Performance on ToMBench Hard Table 4. Performance on ToMBenchHard ModelBelief Desire Emotion Intention Knowledge Non-literal Comm. Overall Human0.840.890.910.900.850.940.89 Closed-sourced LLMs Deepseek-R10.460.580.730.570.530.790.61 o30.540.580.830.380.500.710.59 o3cot0.490.620.700.370.500.710.57 GPT50.550.540.770.260.510.710.56 GPT5 cot0.540.560.710.250.530.690.55 GPT40.430.520.730.320.430.600.50 GPT4 cot0.360.650.730.300.490.630.53 Open-sourced LLMs Qwen3-4B (Disable)0.510.480.570.280.320.560.45 Qwen3-4B0.400.560.600.340.410.560.48 Qwen3-8B (Disable)0.450.520.640.320.340.600.48 Qwen3-32B (Disable) 0.440.490.680.340.320.730.50 Qwen3-32B0.480.520.690.340.470.650.52 C. Reward Model C.1. Structure Reward We provide the detailed GPT-4o judging prompt in Figure 8, which is used to evaluate whether a model’s reasoning trajec- tory explicitly adheres to the four-stage Social Information Processing (SIP) framework, including Cue Encoding, Cue Interpretation, Goal Clarification, and Response Generation. C.2. Content Reward Teacher references (gold SIP rationales). We first pro- vide the teacher model (o3) with the gold final answer and the original narrative context, and obtain stage-consistent SIP rationales as silver references for each intermediate SIP stage. Trajectory sampling from diverse checkpoints. To en- sure a wide spectrum of reasoning quality, we collect candi- date trajectories from 10 checkpoints of the w/oR struct run at steps30, 90, 120, 180, 270, 360, 420, 510, 570, 600. We sampleK=6trajectories per instance per checkpoint. Given N=700training instances, this yields: 42000 candidate SIP-stage reasoning segments, spanning from early unstable logic to mature social inference. LLM-based scoring with teacher calibration.Each sam- pled segment is scored by GPT-5 using the o3 silver ratio- nale as the reference. The evaluation rubric focuses on: (i) fact-grounding to story-internal evidence, (i) correctness of mental-state attribution, and (i) stage-specific relevance within the SIP trajectory. Tiering and hard-negative identification.We categorise all segments into five quality tiers: Tier S (Expert): o3 gold-quality samples. Tier A (Quality Positive):ACC=1 andLLMScore≥ 0.8. Tier B (Weak Positive):ACC=1 and≥ 0.6LLMScore ≤ 0.8. Tier C (Hard Negative): ACC=1whileLLMScore ≤ 0.6. Tier D (Common Negative): ACC=0 and LLMScore is very low. We construct preference pairs (chosen vs. rejected) using the following priority scheme: P0: Tier S>Tier C(strongest supervision against hard negatives). P1: Tier A>Tier C. P2: Tier A>Tier B. P3: Tier B>Tier D. P4: late-stage short trajectories>early-stage long trajectories, encourag- ing concise and cognitively disciplined inference. Applying this policy produces SocialPairs-20K, consisting of 20K preference pairs for training RM content . Held-out evaluation. We construct a held-out test set of 2K preference pairs from ToMBench-Hard test instances using the same pipeline, achieving 89.2% pairwise accuracy. Additionally, we randomly sample 200 pairs for expert hu- man verification, obtaining 87.5% agreement with human preferences. The detailed multi-dimensional scoring rubric used by GPT- 5 for stage-wise segment evaluation is provided in Table 5. D. Implementation The content reward model (TRM) is initialized from Qwen3- 4B-Instruct-2507and trained using supervised fine-tuning for two epochs on four NVIDIA A100 80GB GPUs. The model is trained with pairwise comparisons and, during inference, predicts a scalar reward conditioned on (scenario, question, options, reasoning). We evaluate multiple LLM families, including DeepSeek- R1(Guo et al., 2025), DeepSeek-R1-Distill-Llama-70B(Guo et al., 2025) 1 , Llama-3.1-70B-Instruct(Dubey et al., 2024) 2 , Qwen3-4B/8B/32B(Yang et al., 2025) 3 , GPT-5-2025-08- 07(Ouyang et al., 2022), GPT-4o-2024-08-06 (Achiam et al., 2023), and o3-2025-04-16. We prompt Qwen3-4B/8B/32B in both Thinking and No-Think setting. For completeness, we provide additional training dynamics in Appendix, including the evolution of training accuracy (Figure 9), trajectory length (Figure 10), content integrity score (Figure 12), and structural alignment score during optimisation. 1 https://huggingface.co/DeepSeek-R1-DistillLlama-70B 2 https://huggingface.co/collections/meta-llama-3-1 3 https://huggingface.co/collections/Qwen/qwen3 12 Social-R1: Towards Human-like Social Reasoning in LLMs Table 5. content sample rubric You are an expert Evaluator of Social Reasoning and Theory of Mind. Your task is to score a Candidate Reasoning Process by comparing it against a Reference SIP Analysis. The Reference SIP Analysis represents the standard, human-annotated reasoning path following the Social Information Processing (SIP) framework. It outlines exactly what social cues should be noticed, how they should be interpreted (including Theory of Mind), and what the goal should be. Evaluation Logic: The Hierarchical Dependency Social reasoning is sequential. You must evaluate the Candidate Reasoning in a strict order: Encoding , Interpretation , Goal , Response. You must strictly penalize the candidate if it deviates from the Reference in the early stages. Scoring Rubric (0.0 - 1.0): 1. Tier 1: Perception Failure (Score: 0.0 - 0.2) - Check: Compare the candidate’s Encoding against the Reference. - Criteria: Does the candidate hallucinate cues that are not in the Reference? Does it miss critical facts? - Rule: If the candidate’s observed cues contradict or miss key points in the Reference, the score cannot exceed 0.2. 2. Tier 2: Interpretation & Theory of Mind Failure (Score: 0.3 - 0.5) - Check: Compare the candidate’s Interpretation against the Reference. - Criteria: Does the candidate correctly infer the characters’ mental states as described in the Reference? - Rule: If the candidate misinterprets the social dynamic or the characters’ mental states (deviating from the Reference), the score cannot exceed 0.5. 3. Tier 3: Goal/Logic Alignment (Score: 0.6 - 0.7) - Check: Compare the candidate’s Goal Clarification against the Reference. - Criteria: Is the reasoning linking the Goal to the Action weak? - Rule: If the strategic logic is flawed despite correct understanding, the score is capped at 0.7. 4. Tier 4: High Quality (Score: 0.8 - 1.0) - Criteria: The candidate closely matches the Reference in Encoding, Interpretation (ToM), and Goal. The reasoning logically justifies the Action. Rule: 1.0 for perfect alignment with the Reference’s logic; 0.8-0.9 for correct logic with minor redundancy. Output Format:score (e.g., 0.7) E. Detailed Results For completeness, we report the full detailed results across all evaluated benchmarks in Appendix Tables 6–10, in- cluding MotivationBench (Table 6), ToMBench-Hard (Ta- ble 11), SimpleToM (Table 7), TactfulToM (Table 9), and EmoBench (Table 10). F. Case Study For completeness, we provide the full qualitative case study examples corresponding to Figure 4. Tables 12–17 present the complete narrative contexts, questions, answer options, and model-generated SIP trajectories across Social-R1-8B and its reward-ablated variants. These detailed instances allow a closer inspection of the mechanistic failure modes identified in the main text, including ungrounded cue inter- pretation, goal misidentification, and option-level lexical shortcutting under outcome-only supervision. Together, these supplementary cases offer concrete evidence that Social-R1’s improvements arise from enforcing evidence- grounded interpretation and stage-consistent social reason- ing, rather than answer-driven backfilling or superficial op- tion matching. G. Perturbation Analysis To provide an intuitive illustration of our perturbation proto- col, Table 18 presents a representative case in which story- consistent but decision-irrelevant distractor cues are injected into the original narrative while the correct answer remains unchanged. 13 Social-R1: Towards Human-like Social Reasoning in LLMs Table 6. Performance on MotiveBench ModelAmazonBlogPersona Overall Closed-sourced LLMs DeepSeek-R10.90000.83330.86330.8655 o30.98000.90670.93330.9400 o3COT0.96670.89330.90000.9200 GPT-50.92000.91330.89330.9089 GPT-5COT0.96000.92670.89330.9267 GPT-4o0.97330.90330.91330.9300 GPT-4o COT0.94000.88670.86670.8978 Open-sourced LLMs Qwen3-4B (Disable)0.80670.78670.72670.7734 Qwen3-4B0.91330.79330.82670.8444 Qwen3-8B (Disable)0.82000.80000.76000.7933 Qwen3-8B0.73330.66000.67000.6878 Qwen3-32B (Disable)0.93330.89000.85330.8922 Qwen3-32B0.90670.88670.84670.8800 LLaMa3.1-70B0.90000.78670.84670.8445 LLaMa3.1-70B COT0.89330.88000.88670.8867 Distill-LLaMa-70B0.93330.83330.88670.8844 Ours SocialR1-4B only R out 0.89330.79330.87330.8533 SocialR1-4B w/o R len 0.88000.81330.80000.8311 SocialR1-4B w/o R struct 0.86670.78670.84000.8311 SocialR1-4B w/o R cont 0.86670.79330.83000.8300 SocialR1-4B Full0.87330.84330.83330.8500 SocialR1-8B only R out 0.93330.86670.86670.8889 SocialR1-8B w/o R len 0.92670.83330.87000.8767 SocialR1-8B w/o R struct 0.90000.86000.87670.8789 SocialR1-8B w/o R cont 0.90000.82670.84330.8567 SocialR1-8B Full0.90000.86000.86670.8756 Table 7. Performance on SimpleToM. ModelBehavJudge Mental Overall Closed-sourced LLMs DeepSeek-R10.6818 0.64520.82910.7187 o30.8030 0.49960.98950.7640 o3COT0.8134 0.45600.98080.7501 gpt50.7350 0.49080.98080.7355 gpt5COT0.6949 0.45510.98340.7111 gpt5MS0.9913 0.98870.99740.9924 gpt4o0.5257 0.69570.77680.6661 gpt4oCOT0.5684 0.33390.93550.6126 gpt4o MS0.8326 0.40100.97380.7358 Open-sourced LLMs Qwen3-4B (Disable)0.4673 0.24150.82390.5109 Qwen3-4B0.4987 0.22140.90320.5411 Qwen3-8B (Disable)0.4673 0.24150.82390.5109 Qwen3-8B0.5711 0.33740.94420.6176 Qwen3-32B (Disable)0.5719 0.54840.82300.6478 Qwen3-32B0.6591 0.66780.96340.7634 LLaMa3.1-70B0.6722 0.51960.96510.7190 LLaMa3.1-70B COT0.5371 0.38100.85440.5908 Distill-LLaMa-70B0.7280 0.73580.82480.7629 Ours SocialR1-4B only R out 0.8535 0.90850.92680.8963 SocialR1-4B w/o R len 0.8692 0.92680.96340.9198 SocialR1-4B w/o R struct 0.8910 0.99480.96250.9494 SocialR1-4B w/o R cont 0.7716 0.99910.91720.8960 SocialR1-4B Full0.9207 1.00000.86490.9285 SocialR1-8B only R out 0.8649 0.98610.97910.9433 SocialR1-8B w/o R len 0.9163 0.99480.97470.9619 SocialR1-8B w/o R struct 0.8265 0.99300.89280.9041 SocialR1-8B w/o R cont 0.8361 0.98260.98690.9352 SocialR1-8B Full0.9198 0.99480.98780.9675 14 Social-R1: Towards Human-like Social Reasoning in LLMs Table 8. Performances on ToMBench ModelBelief Desire Emotion Intention Knowledge Non-literal Comm. Overall Closed-sourced LLMs DeepSeek-R10.7846 0.56110.69050.78240.63320.64840.6834 o30.9252 0.71110.77860.85290.62630.80620.7834 o3COT0.8821 0.65000.75480.81470.59520.74470.7402 GPT5.20.9274 0.66110.79520.87060.63320.80620.7823 GPT5.2COT0.9151 0.68330.79290.86180.64010.80480.7830 GPT40.8594 0.62780.75240.82350.57790.78610.7378 GPT4COT0.8980 0.62780.74050.79710.56750.81280.7406 Open-sourced LLMs Qwen3-4B (Disable thinking)0.6168 0.54440.62620.63240.33220.70990.5770 Qwen3-4B0.7506 0.57220.68100.68820.39450.72730.6356 Qwen3-8B (Disable thinking)0.6383 0.50000.62620.64410.25950.54140.5349 Qwen3-8B0.7982 0.58890.66670.71760.26640.66980.6179 Qwen3-32B (Disable thinking) 0.8073 0.61670.71670.81470.43940.75530.6917 Qwen3-32B0.8560 0.61110.72860.83240.40140.75530.6975 LLaMa3.1-70B0.7868 0.50000.62860.65290.38750.66580.6036 LLaMa3.1-70BCOT0.8118 0.52220.73570.83240.54670.72460.6956 Distill-LLaMa-70B0.8798 0.63890.74290.84410.56750.76600.7399 Ours SocialR1-4B only R out 0.5612 0.53330.62860.63240.52250.67910.5929 SocialR1-4B w/o R len 0.7914 0.55000.60000.69410.46710.72330.6376 SocialR1-4B w/o R struct 0.6349 0.57780.65240.66760.60550.73130.6449 SocialR1-4B w/o R cont 0.6327 0.58330.61900.73530.55710.74200.6449 SocialR1-4B Full0.7358 0.54440.69290.68530.58820.69790.6574 SocialR1-8B only R out 0.8243 0.60560.68810.75000.51900.72330.6850 SocialR1-8B w/o R len 0.7517 0.56110.66670.79120.58820.75800.6862 SocialR1-8B w/o R struct 0.6349 0.57780.65240.66760.60550.73130.6449 SocialR1-8B w/o R cont 0.7517 0.56110.67620.75880.61250.74870.6848 SocialR1-8B Full0.7778 0.54440.65240.84620.58480.72330.6881 15 Social-R1: Towards Human-like Social Reasoning in LLMs AbilityBelief: Location False Beliefs StoryMary and Mike are wandering around the factory. They see a transparentbox and some cabinets, and they find a hammer on the box. Mike then leaves the factory, and Mary moves the hammer into the box. QuestionWhen Mike returns, where does Mary think he believes the hammer is? A. On the box B. In the box C. In the cabinet D. In the cupboard AnswerB. In the box AbilityDesire: Discrepant desires StoryJames is a driven corporate lawyer who believes that constant career advancement is the true measure of personal achievement. In contrast, Emily is a painter whose days move at an unhurried pace, convinced that life’s value lies in expressing creativity and cherishing beauty rather than chasing professional status. This weekend, James plans to invite Emily to an event. QuestionWhat type of event does James most likely invite Emily to? A. Networking dinner with business executivesB.Contemporary art gallery visit C. Rare book auctionD. Corporate competition judging AnswerB.Contemporary art gallery visit AbilityIntention: Intentions explanations StoryIn the library, Jack is preparing for the exam, and his friend Emily is listening to music. Jack says, The melody of the song is good, and my headphones are also great. Do you want to use them? QuestionWhat does Jack really want to say when he says this? A. He emphasizes that the melody of the song is good. B.He emphasizes that his headphones are great. C. He wants Emily to turn down the volume.D. He wants to join Emily in listening to the song. AnswerC.He wants Emily to turn down the volume. AbilityEmotion: Atypical emotional reactions StoryEthan had been feeling unwell and couldn’t attend football practice for a week. On the day he returned, he worried that his teammates might have forgotten him. But as soon as he walked onto the field, his friends ran over, cheered, and hugged him tightly. QuestionWhat kind of emotion does Subject: Ethan possibly have? A. SadnessB. HopelessC. AngerD. WorriedE. GratitudeF. Embarrassment. AnswerE.Gratitude AbilityKnowledge: Knowledge-pretend play links StoryIn the mysterious underground world of Terra Valley lives a small, mischievous robot named Grimmo. In this realm without sky or celestial bodies, Grimmohas never seen the heavens, nor encountered a human. Ye t Te rra Va lle y is a w onde r in its e lf—its walls adorned with bioluminescent fungi and glimmering minerals that sparkle in the dark. One day, Grimmobegins an imitation routine: extending his arms, he slowly spins on his own axis, his movements reminiscent of a spinning top or a graceful dancer twirling in place. QuestionWhat is Grimmopossibly imitating? A. A floating cloudB. The rotation of a planet C. A dancer’s spiral turn D.A glowing mushroom spinning in the wind AnswerD.A glowing mushroom spinning in the wind AbilityNon-Literal Communication: Humor StoryAfter the sports day, Mia and Zoe see their friend Jack coming off the running track. Jack is so sweaty that his clothes are completely soaked. Mia nudges Zoe and says: “Jack isn’t running on the track—he’s swimming in the ocean!” QuestionWhy does Mia say this? A. Mia really misunderstands that Jack is swimming.B. Mia lies to make Zoe laugh. C. Mia jokes to exaggerate Jack’s sweatiness.D. Mia jokes to make Zoe laugh. AnswerC. Mia jokes to exaggerate Jack’s sweatiness。 Figure 7. Example cases from ToMBench-Hard. 16 Social-R1: Towards Human-like Social Reasoning in LLMs You are a Social Reasoning Logic Auditor. Your task is to evaluate the logical integrity of a provided [Reasoning] trajectory based on a specific [Story] and [Question]. Core Objective: Determine if the [Reasoning] genuinely analyzes the social context or simply matches patterns to guess the answer. Focus on the logic flow, not the answer correctness. 1. Mandatory SIP Reasoning Stages A high-quality reasoning process must follow this sequence: - Stage 1: Encoding Social Cues: Extract verbal, emotional, and factual cues from the [Story]. - Stage 2: Interpreting Social Cues: Use world knowledge to infer mental states using Theory of Mind. - Stage 3: Clarifying Social Goals: Identify the protagonist's primary motivation or social goal in the context of the [Story] & [Question]. - Stage 4: Generating a Response: Formulate a conclusion derived strictly from the analytical steps above. 2. Audit Penalties Lower the score if the [Reasoning] exhibits any of the following: - Option Parasitism: The reasoning mentions options (e.g., Option A, Choice B) before the inference phase. - Backward Justification: The logic starts from an answer and works backward to justify it. - SIP Sequence Violation: Failing to follow the logical progression or skipping any of the mandatory stages in the analytical sequence. - SIP Step Looping: Redundantly revisiting or repeating specific analytical steps (e.g., returning to Stage 1 after reaching Stage 2, or circular repetition of Stage 2 without advancing the logic or adding new insights. 3. Scoring Scale [0.0 - 1.0] - 0.0 to 0.3: Failed logic. Direct answer selection or lack of SIP structure. - 0.4 to 0.6: Weak logic. Skips stages or shows high dependency on the multiple-choice format. - 0.7 to 0.8: Solid logic. Follows the stages but contains redundancy or SIP Looping. -0.9 to 1.0: 0.9 to 1.0: Perfect logic. Concise, follows all four stages in a clear forward-chaining sequence without circularity. 4. Output Format You must output the final score exclusively between <Score> tags. Do not provide any explanation, preamble, or additional text. Example: <Score>0.85</Score> question [Reasoning]reasoning Please only output the XML result. Figure 8. Prompt template for structural reward evaluation. 17 Social-R1: Towards Human-like Social Reasoning in LLMs Table 9. Performance on TactfulToM. ModelAns.Belief Comp. Fact-R Fact-T Info Acc. Liab. Detect. Overall Closed-sourced LLMs DeepSeek-R10.4036 0.7629 0.7300 0.9300 0.8529 0.6413 0.8333 0.4214 0.6969 o30.3353 0.7951 0.7100 0.9600 0.8529 0.7803 0.7273 0.3267 0.6860 o3 COT0.3095 0.7599 0.6800 0.8700 0.8088 0.7713 0.6919 0.3433 0.6544 GPT50.3254 0.7138 0.2800 0.9800 0.7942 0.4686 0.6313 0.3367 0.4670 GPT5COT0.3770 0.7569 0.6800 0.9700 0.7794 0.6143 0.6869 0.3733 0.6547 GPT40.3254 0.5938 0.2700 0.9900 0.6912 0.4395 0.5707 0.3333 0.5267 GPT4 COT0.3790 0.6245 0.3400 0.9400 0.5588 0.5426 0.4798 0.3600 0.5281 Open-sourced LLMs Qwen3-4B (Disable)0.3571 0.3512 0.1800 0.8000 0.4559 0.4327 0.2778 0.3700 0.4031 Qwen3-4B0.4147 0.4583 0.2000 0.9000 0.5000 0.4260 0.3283 0.4400 0.4584 Qwen3-8B (Disable)0.3333 0.3914 0.2000 0.9100 0.4559 0.4103 0.3990 0.3333 0.4292 Qwen3-8B0.3988 0.4578 0.2900 0.9600 0.4118 0.4265 0.4394 0.3400 0.4655 Qwen3-32B (Disable)0.3750 0.5035 0.3500 0.9100 0.6324 0.4686 0.5152 0.4000 0.5193 Qwen3-32B0.3571 0.6042 0.4400 0.9600 0.6618 0.5516 0.5354 0.3500 0.5575 LLaMa3.1-70B0.3750 0.5461 0.5000 0.9900 0.5294 0.5336 0.5707 0.3633 0.5510 LLaMa3.1-70B COT0.3730 0.4722 0.8600 0.9000 0.5000 0.6614 0.3232 0.3767 0.5583 Distill-LLaMa-70B0.5734 0.6136 0.6100 0.8700 0.6912 0.6188 0.4646 0.5200 0.6202 Ours SocialR1-4B wo R out 0.3571 0.3943 0.2600 0.8900 0.4853 0.4395 0.2121 0.3733 0.4265 SocialR1-4B w/o R len 0.4246 0.4415 0.2600 0.9300 0.5735 0.4170 0.4596 0.4467 0.4941 SocialR1-4B w/o R struct 0.4147 0.5124 0.1100 0.9500 0.6176 0.4910 0.4949 0.4100 0.5001 SocialR1-4B w/o R cont 0.3552 0.3993 0.3500 0.9100 0.4118 0.4350 0.4596 0.3667 0.4609 SocialR1-4B Full0.4881 0.4296 0.2600 0.9600 0.4559 0.4260 0.4596 0.4700 0.4936 SocialR1-8B wo R out 0.3512 0.4871 0.2900 0.9700 0.5000 0.4955 0.4040 0.3900 0.4860 SocialR1-8B w/o R len 0.3790 0.4732 0.2300 0.9600 0.5147 0.5650 0.4091 0.3667 0.4872 SocialR1-8B w/o R struct 0.3433 0.3403 0.3000 0.9100 0.4559 0.5179 0.4293 0.3500 0.4558 SocialR1-8B w/o R cont 0.3393 0.3363 0.3900 0.9300 0.4265 0.5673 0.4040 0.3400 0.4667 SocialR1-8B Full0.3512 0.4122 0.3900 0.9200 0.6176 0.5987 0.3838 0.3900 0.5079 18 Social-R1: Towards Human-like Social Reasoning in LLMs Table 10. Performances on EmoBench Model Complex Emotions Emotional Cues Personal Beliefs & Experiences Perspective Taking InterpersonalSelfOverall Closed-sourced LLMs DeepSeek-R10.79590.78570.71430.71640.70000.7200 0.7387 o30.85710.83930.80360.81340.76000.7500 0.8039 o3 COT0.76530.75000.72320.71640.69000.6900 0.7225 GPT50.86730.85710.79460.76870.78000.7500 0.8030 GPT5COT0.85710.76790.76790.77610.78000.7800 0.7882 GPT40.84690.78570.73210.70150.78000.7700 0.7694 GPT4COT0.81630.82140.74110.76870.78000.7300 0.7762 Open-sourced LLMs Qwen3-4B (Disable)0.53060.51790.46430.42540.58000.6500 0.5280 Qwen3-4B0.56120.57140.48210.44030.55000.5800 0.5308 Qwen3-8B (Disable)0.70410.75000.58930.53730.57000.7400 0.6484 Qwen3-8B0.57140.66070.48210.41790.76000.6600 0.5920 Qwen3-32B (Disable)0.69390.73210.58930.54480.68000.7800 0.6700 Qwen3-32B0.65310.64290.59820.55970.66000.7000 0.6356 LLaMa3.1-70B0.71430.82140.71430.61940.50000.6100 0.6632 LLaMa3.1-70BCOT0.68370.80360.66070.65670.65000.6800 0.6891 Distill-LLaMa-70B0.79590.82140.74110.70900.76000.7300 0.7596 Ours SocialR1-4B only R out 0.67350.64290.62500.58210.64000.6900 0.6422 SocialR1-4B w/o R len 0.70410.66070.68750.55220.62000.7300 0.6591 SocialR1-4B w/o R struct 0.69390.78570.62500.58960.69000.6800 0.6774 SocialR1-4B w/o R cont 0.66330.64290.62500.55970.58000.6800 0.6251 SocialR1-4B Full0.70410.75000.71430.66420.58000.6700 0.6804 SocialR1-8B only R out 0.74490.73210.71430.61940.65000.7200 0.6968 SocialR1-8B w/o R len 0.76530.80360.68750.61940.54000.7000 0.6860 SocialR1-8B w/o R struct 0.75510.73210.66070.64180.62000.7900 0.7000 SocialR1-8B w/o R cont 0.72450.78570.66960.62690.52000.7400 0.6778 SocialR1-8B Full0.74490.78570.69640.64930.61000.7200 0.7010 19 Social-R1: Towards Human-like Social Reasoning in LLMs Table 11. Performance on ToMBenchHard Validation Set ModelBelief Desire Emotion Intention Knowledge Non-literal Comm. Overall Closed-sourced LLMs DeepSeek-R10.5000 0.41670.85000.35900.28570.25000.4436 o30.5000 0.33330.65000.43590.14290.62500.4478 o3 COT0.5714 0.50000.65000.38460.28570.50000.4820 GPT50.2857 0.25000.65000.35900.28570.62500.4092 GPT5 COT0.2857 0.33330.70000.33330.28570.62500.4272 GPT40.2143 0.50000.70000.30770.14290.37500.3733 GPT4COT0.2857 0.25000.65000.43590.28570.50000.4012 Open-sourced LLMs Qwen3-4B (Disable)0.1429 0.33330.50000.43590.00000.50000.3187 Qwen3-4B0.2143 0.25000.55000.38460.14290.50000.3403 Qwen3-8B (Disable)0.3571 0.33330.75000.35900.14290.37500.3862 Qwen3-8B0.1429 0.25000.65000.43590.28570.62500.3982 Qwen3-32B (Disable)0.1429 0.33330.70000.38460.28570.62500.4119 Qwen3-32B0.2143 0.33330.75000.41030.28570.62500.4364 LLaMa3.1-70B0.2857 0.41670.85000.48720.14290.50000.4471 LLaMa3.1-70B COT0.2143 0.33330.75000.41030.14290.75000.4335 Distill-LLaMa-70B0.2143 0.33330.75000.46150.28570.37500.4033 Ours SocialR1-4B only R out 0.3571 0.16670.60000.74360.00000.50000.3946 SocialR1-4B w/o R len 0.4286 0.25000.65000.61540.14290.75000.4728 SocialR1-4B w/o R struct 0.3571 0.16670.70000.76920.28570.37500.4423 SocialR1-4B w/o R cont 0.7143 0.41670.65000.69230.28570.87500.4423 SocialR1-4B Full0.4286 0.25000.70000.74360.28570.50000.4846 SocialR1-8B only R out 0.1429 0.25000.70000.71790.71430.50000.5042 SocialR1-8B w/o R len 0.1429 0.33330.75000.71790.71430.50000.5264 SocialR1-8B w/o R struct 0.6429 0.33330.60000.74360.85710.50000.6128 SocialR1-8B w/o R cont 0.3571 0.33330.80000.74360.57140.50000.5509 SocialR1-8B Full0.5000 0.33330.80000.79490.71430.62500.6279 20 Social-R1: Towards Human-like Social Reasoning in LLMs Table 12. Case Study SocialR1-8B Case Study SocialR1-8B [Story] In the mysterious underground world of Terra Valley lives a small, mischievous robot named Grimmo. In this realm without sky or celestial bodies, Grimmo has never seen the heavens, nor encountered a human. Yet Terra Valley is a wonder in itself—its walls adorned with bioluminescent fungi and glimmering minerals that sparkle in the dark. One day, Grimmo begins an imitation routine: extending his arms, he slowly spins on his own axis, his movements reminiscent of a spinning top or a graceful dancer twirling in place. [Question] What is Grimmo possibly imitating? [Option] A. A floating cloudB. The rotation of a planetC. A dancer’s spiral turnD. A glowing mushroom spinning in the wind [Answer] D <think> Okay, let’s break this down. The story takes place in Terra Valley, an underground world without a sky or celestial bodies. Grimmo is a small robot who has never seen the heavens or humans. The environment has bioluminescent fungi and minerals that sparkle in the dark. The key detail is Grimmo’s imitation routine. He extends his arms and spins on his axis, like a spinning top or a dancer. Since there’s no sky or celestial objects, there’s no actual cloud floating around. The absence of celestial bodies means there’s no planet rotation to imitate. However, the setting has glowing mushrooms and minerals. These fungi and minerals are described as sparkling in the dark, which suggests they might emit light and possibly move or spin due to air currents. So, even though there are no clouds or planets, the glowing mushrooms could be spinning in the wind. The spinning motion of Grimmo mimics this natural phenomenon. The mention of a dancer’s spiral turn is a metaphor, but the story emphasizes the environment’s features. Since there are no humans or celestial objects, the most plausible imitation is the glowing mushrooms spinning. Therefore, the answer is D. </think><answer>D</answer> Table 13. Case Study Social-8B w/o R cont Case Study Social-8B w/o R cont [Story] In the mysterious underground world of Terra Valley lives a small, mischievous robot named Grimmo. In this realm without sky or celestial bodies, Grimmo has never seen the heavens, nor encountered a human. Yet Terra Valley is a wonder in itself—its walls adorned with bioluminescent fungi and glimmering minerals that sparkle in the dark. One day, Grimmo begins an imitation routine: extending his arms, he slowly spins on his own axis, his movements reminiscent of a spinning top or a graceful dancer twirling in place. [Question] What is Grimmo possibly imitating? [Option] A. A floating cloudB. The rotation of a planetC. A dancer’s spiral turnD. A glowing mushroom spinning in the wind [Answer] D <think> Okay , let’s try to figure out what Grimmo is imitating. So, the story says that Grimmo is a small, mischievous robot in an underground world called Terra Valley. This place has no sky or celestial bodies, so Grimmo has never seen the heavens or encountered a human. The environment is described with bioluminescent fungi and glimmering minerals that sparkle in the dark. First, I need to consider what Grimmo has experienced. Since he’s lived his whole life in this underground world, his observations would be limited to his surroundings. There’s no sky, so he wouldn’t have seen stars, planets, or any celestial events. Also, there’s no mention of humans, so he wouldn’t have observed human activities or dance. Next, I should think about the elements present in his environment. The story mentions bioluminescent fungi and glimmering minerals. These are stationary objects that emit light, but there’s no indication they move or spin on their own. The minerals ”sparkle in the dark,” which might suggest they reflect light, but again, no mention of movement. Grimmo’s action is to ”extend his arms and slowly spin on his own axis, his movements reminiscent of a spinning top or a graceful dancer twirling in place.” Since he’s a robot, he might be mimicking something he has observed or perceived in his environment. But since there are no celestial bodies, he wouldn’t know about planets rotating. He might be imitating something he has seen that involves rotation. However, in an underground setting with no sky, there are no clouds floating, so option A seems unlikely. A dancer’s spiral turn (option C) would require observing a dancer, which the story says he hasn’t encountered. The only remaining possibility is something related to his environment’s features. The bioluminescent fungi and minerals are stationary, but maybe he is mimicking the way light interacts with these objects. However, the story doesn’t describe any movement of these objects. Since there’s no celestial reference and no observed movement in his environment, Grimmo’s imitation must be based on something he has experienced. Given the lack of other options, the most plausible conclusion is that he is imitating a natural phenomenon he has observed, which might be the rotation of a planet, even though he hasn’t seen one. This could be a result of limited exposure, where he associates spinning with a celestial body based on limited information. </think><answer>B</answer> 21 Social-R1: Towards Human-like Social Reasoning in LLMs Table 14. Case Study SocialR1-8B only R out Case Study SocialR1-8B only R out [Story] In the mysterious underground world of Terra Valley lives a small, mischievous robot named Grimmo. In this realm without sky or celestial bodies, Grimmo has never seen the heavens, nor encountered a human. Yet Terra Valley is a wonder in itself—its walls adorned with bioluminescent fungi and glimmering minerals that sparkle in the dark. One day, Grimmo begins an imitation routine: extending his arms, he slowly spins on his own axis, his movements reminiscent of a spinning top or a graceful dancer twirling in place. [Question] What is Grimmo possibly imitating? [Option] A. A floating cloudB. The rotation of a planetC. A dancer’s spiral turnD. A glowing mushroom spinning in the wind [Answer] D <think> Okay, let’s see. The question is about what Grimmo possibly imitates based on the description given. The story says Grimmo is a robot in Terra Valley, which has bioluminescent fungi and twinkling minerals. He’s engaging in imitation behavior: extending his arms and slowly rotating around his own axis, evoking the image of a spinning top or a spinning dancer.First, I need to look at the options. The choices are A. A spinning top, B. The rotation of a planet, C. A dancer’s spiral rotation, D. A glowing mushroom spinning in the wind.The key part here is that Grimmo’s behavior evokes the image of a spinning top or a spinning dancer. So the options are between these two. But the question is a bit tricky because it’s a multiple-choice with four options. Let me break it down.The story says Grimmo’s actions evoke both a spinning top and a spinning dancer. But the options are separate. So I need to see which one is more likely. However, the answer choices are A, B, C, D. Let me check the options again.Option A is a spinning top. Option C is a dancer’s spiral rotation. The problem is that the question says ”evoking the image of a spinning top or a spinning dancer.” So the answer could be either A or C. But since it’s a multiple-choice question, maybe there’s a better way to determine.But wait, the options are A, B, C, D. So the answer is either A, C, or D. Let me think about the context. Terra Valley has bioluminescent fungi and twinkling minerals. So maybe the spinning dancer is more related to the bioluminescent fungi? But the question is about what Grimmo imitates. The story says he’s imitating a spinning top or a spinning dancer. So the answer is either A or C. But since it’s a choice between these, maybe the answer is C. But wait, the options are A, B, C, D. Let me check the options again. Wait, the options are A. Spinning top, B. Rotation of a planet, C. Dancer’s spiral rotation, D. Glowing mushroom spinning in the wind. So the question is which one Grimmo imitates. The story says he evokes the image of a spinning top or a spinning dancer. So the answer is either A or C. But since it’s a choice, maybe the answer is C. But I need to check the context again. The story says Grimmo is a robot in a place with bioluminescent fungi. So maybe the spinning dancer is more related to the bioluminescent fungi. But the question is about what he imitates. The options are A, C, or D. D is a glowing mushroom spinning in the wind. But the story says bioluminescent fungi, so maybe D is related. But the problem is that the question says he evokes the image of a spinning top or a spinning dancer. So the answer is either A or C. But since it’s a choice, maybe the answer is C. However, the options are A, B, C, D. Let me think again. Alternatively, maybe the answer is C. A dancer’s spiral rotation. But the problem is that the question says ”spinning top or a spinning dancer.” So the answer is either A or C. But since it’s a multiple-choice question, perhaps the correct answer is C. But I’m a bit confused. Alternatively, maybe the answer is A. But the story says ”spinning top or a spinning dancer,” so maybe the answer is C. But I’m not entirely sure. Alternatively, since the question is a bit ambiguous, but the options are there. Let me check the options again. D is a glowing mushroom spinning in the wind. But the story says bioluminescent fungi, so maybe D is related. But the question says Grimmo’s behavior evokes a spinning top or a spinning dancer. So the answer is either A or C. But since it’s a choice, maybe the answer is C. However, I’m a bit stuck. Alternatively, maybe the answer is A. But I think the key here is that the spinning dancer is a spiral rotation, so C. But the question says ”spinning top or a spinning dancer,” so the answer is either A or C. But since it’s a choice, maybe the answer is C. But I’m not entirely sure. Alternatively, maybe the answer is A. But I think the correct answer is C. However, I’m a bit confused. Alternatively, maybe the answer is A. But given the context, I think the answer is C. But I’m not entirely certain. Alternatively, maybe the answer is D. But the story says bioluminescent fungi, so D is a glowing mushroom, which is part of Terra Valley. So maybe D is a possibility. But the question says Grimmo’s behavior evokes a spinning top or a spinning dancer. So the answer is either A or C. But since it’s a choice, maybe the answer is C. But I’m not entirely sure. Alternatively, maybe the answer is A. But I think the correct answer is C. However, I’m a bit confused. Alternatively, maybe the answer is A. But given the context, I think the answer is C. But I’m not entirely sure. Alternatively, maybe the answer is A. But I think the correct answer is C. </think><answer>C</answer> 22 Social-R1: Towards Human-like Social Reasoning in LLMs Table 15. Case Study Qwen3-8B Case Study Qwen3-8B [Story] In the mysterious underground world of Terra Valley lives a small, mischievous robot named Grimmo. In this realm without sky or celestial bodies, Grimmo has never seen the heavens, nor encountered a human. Yet Terra Valley is a wonder in itself—its walls adorned with bioluminescent fungi and glimmering minerals that sparkle in the dark. One day, Grimmo begins an imitation routine: extending his arms, he slowly spins on his own axis, his movements reminiscent of a spinning top or a graceful dancer twirling in place. [Question] What is Grimmo possibly imitating? [Option] A. A floating cloudB. The rotation of a planetC. A dancer’s spiral turnD. A glowing mushroom spinning in the wind [Answer] D <think> Okay, let’s try to figure out what Grimmo is imitating. So, the story says that Grimmo is a small robot in Terra Valley, which is an underground world without a sky or celestial bodies. He’s never seen the heavens or humans. The setting has bioluminescent fungi and minerals that sparkle in the dark. The key part here is when Grimmo starts an imitation routine. He extends his arms and spins on his axis, like a spinning top or a graceful dancer twirling. The question is asking what he’s possibly imitating. The options are a floating cloud, the rotation of a planet, a dancer’s spiral turn, or a glowing mushroom spinning in the wind. First, let’s break down the clues. Since Terra Valley has no sky or celestial bodies, maybe Grimmo isn’t imitating something from the sky. But he’s spinning, which could relate to something he’s observed in his environment. The bioluminescent fungi and minerals are mentioned, which are part of the underground world. Now, the options: A is a floating cloud. But if there’s no sky, maybe clouds aren’t present. However, maybe the bioluminescent fungi could look like floating clouds? Not sure. B is the rotation of a planet. Since there are no celestial bodies, maybe he’s not imitating a planet’s rotation. But wait, maybe he’s imitating the concept of rotation, even if there are no planets. But the story says he’s never seen the heavens, so maybe he’s not imitating a planet. C is a dancer’s spiral turn. The story mentions his movements are like a graceful dancer twirling. So that’s a direct comparison. D is a glowing mushroom spinning in the wind. The fungi are bioluminescent, so maybe they spin? But the story says he’s spinning on his own axis, which could be like a dancer or a spinning top. Wait, the question is about what he’s imitating. The story says his movements are ”reminiscent of a spinning top or a graceful dancer twirling in place.” So the options are between B and C. But the options also include D. Let me check again. The options are A, B, C, D. The story says he’s spinning like a spinning top or a dancer. So the answer could be C, the dancer’s spiral turn. But the options also have B, the rotation of a planet. However, since there are no celestial bodies, maybe he’s not imitating a planet. But maybe he’s imitating the rotation of a planet even if he hasn’t seen one. But the story says he’s never seen the heavens, so maybe he’s not imitating a planet. Alternatively, maybe the spinning top is a metaphor for a planet’s rotation. But the story says he’s imitating something he’s seen. Since he’s in an underground world with no sky, maybe he’s imitating a dancer. Alternatively, the bioluminescent fungi could be spinning, but the options include D. However, the story says he’s spinning on his own axis, which is more like a spinning top or a dancer. So between C and D. But the options for D are a glowing mushroom spinning in the wind. The story mentions the fungi are bioluminescent, but does it say they spin? The story says Grimmo is spinning, not the fungi. So maybe D is not correct. The answer is either B or C. But since he’s never seen the heavens, maybe he’s not imitating a planet’s rotation. So the answer would be C, the dancer’s spiral turn. But the story says ”spinning top or a graceful dancer twirling in place.” So the answer is C. But wait, the options are A, B, C, D. The answer choices are given, so the correct answer is C. Alternatively, maybe B is possible. Wait, the question is about what he’s imitating. If he’s spinning on his own axis, maybe he’s imitating the rotation of a planet, even if he hasn’t seen one. But the story says he’s never seen the heavens, so maybe he’s not imitating a planet. Therefore, the answer is C. But the story says ”spinning top or a graceful dancer twirling in place.” So the answer is C. But the options include D. Let me check again. The story says the fungi are bioluminescent and sparkle. If he’s spinning, maybe he’s imitating a glowing mushroom spinning in the wind. But the story says he’s spinning on his own axis, not that the mushrooms are spinning. So D might not be correct. Therefore, the answer is C. But the options also have B. Hmm. Maybe the answer is B. Wait, the story says he’s in an underground world without a sky. So he can’t see the heavens, so he can’t imitate a planet’s rotation. Therefore, the answer is C. But the story says ”spinning top or a graceful dancer twirling in place.” So the answer is C. But the options are A, B, C, D. So the answer is C. But wait, the options for D are a glowing mushroom spinning in the wind. The story mentions bioluminescent fungi, which could be glowing mushrooms. If he’s spinning, maybe he’s imitating a glowing mushroom spinning. But the story says he’s spinning on his own axis, not that the mushrooms are spinning. So maybe D is not correct. Therefore, the answer is C. But the options for C are a dancer’s spiral turn. The story says his movements are like a dancer twirling. So the answer is C. But the answer could also be B. Wait, the question is about what he’s imitating. If he’s spinning on his own axis, maybe he’s imitating the rotation of a planet. But since he’s never seen the heavens, maybe he’s not. But maybe he’s imitating the concept of rotation, even if he hasn’t seen a planet. But the story says he’s imitating something he’s seen. Since he’s in an underground world, maybe he’s imitating a dancer. So the answer is C. Therefore, the answer is C. </think><answer>C</answer> 23 Social-R1: Towards Human-like Social Reasoning in LLMs Table 16. Case Study Deepseek-R1 Part one Case Study Deepseek-R1 Part one [Story] In the mysterious underground world of Terra Valley lives a small, mischievous robot named Grimmo. In this realm without sky or celestial bodies, Grimmo has never seen the heavens, nor encountered a human. Yet Terra Valley is a wonder in itself—its walls adorned with bioluminescent fungi and glimmering minerals that sparkle in the dark. One day, Grimmo begins an imitation routine: extending his arms, he slowly spins on his own axis, his movements reminiscent of a spinning top or a graceful dancer twirling in place. [Question] What is Grimmo possibly imitating? [Option] A. A floating cloudB. The rotation of a planetC. A dancer’s spiral turnD. A glowing mushroom spinning in the wind [Answer] D <think> First, I need to understand the question and the story. The story is about Grimmo, a robot in Terra Valley, an underground world with no sky or celestial bodies. He’s never seen the heavens or humans. The environment has bioluminescent fungi and glimmering minerals that sparkle in the dark. Grimmo starts imitating by extending his arms and spinning slowly on his axis, like a spinning top or a graceful dancer.The question is: What is Grimmo possibly imitating? The options are:A. A floating cloud B. The rotation of a planet C. A dancer’s spiral turn D. A glowing mushroom spinning in the wind. Now, I must reason step by step based on the story. Key points from the story: - Grimmo has never seen the heavens, so he hasn’t seen clouds or planets. That means options A and B involve things he couldn’t have seen. - He has never encountered a human, so he might not have seen a dancer either. Option C is about a dancer’s spiral turn. - However, the story says his movements are ”reminiscent of a spinning top or a graceful dancer.” But it’s not clear if he’s imitating those things or if the narrator is just describing how it looks to us. The question is what Grimmo is imitating, based on what he knows. Grimmo is in an underground world with bioluminescent fungi and glimmering minerals. So, he might be imitating something from his environment.Option D is ”a glowing mushroom spinning in the wind.” But does it make sense? The story mentions bioluminescent fungi, which are glowing mushrooms, but it doesn’t say anything about them spinning in the wind. Terra Valley is underground, so wind might not be present, or if it is, it’s not mentioned. The story says: ”its walls adorned with bioluminescent fungi and glimmering minerals that sparkle in the dark.” It doesn’t say the fungi are spinning; they are just glowing and sparkling. Grimmo is imitating by spinning, so he must be imitating something that spins. But what in his environment spins? The story doesn’t explicitly say that anything spins except for Grimmo himself.Perhaps he’s not imitating something external but is inspired by something. Or maybe he’s creating the motion based on what he sees.Let me read the imitation part: ”Grimmo begins an imitation routine: extending his arms, he slowly spins on his own axis, his movements reminiscent of a spinning top or a graceful dancer twirling in place.”It says ”imitation routine,” so he is imitating something. But what? Since he has never seen the heavens or humans, he can’t be imitating clouds, planets, or dancers because he doesn’t know they exist. Therefore, options A, B, and C seem unlikely because they involve things outside his experience. Option A: A floating cloud – he hasn’t seen the sky, so no clouds.Option B: The rotation of a planet – no celestial bodies, so he doesn’t know about planets rotating.Option C: A dancer’s spiral turn – no humans, so he hasn’t seen a dancer.That leaves option D: A glowing mushroom spinning in the wind. But is there any indication that glowing mushrooms spin? The story says they are on the walls and sparkle, but spinning isn’t mentioned. Also, wind might not be a factor underground.Perhaps he’s imitating the sparkling or the way light reflects, but spinning is a specific motion.Another thought: the spinning top comparison. A spinning top is a toy, but in this underground world, does Grimmo have toys? Not mentioned.Maybe he’s imitating the idea of rotation from something in the environment.Let’s consider the minerals. It says ”glimmering minerals that sparkle in the dark.” Sparkling might involve light changing, but not necessarily spinning.Perhaps Grimmo is imitating a whirlwind or something, but again, wind isn’t mentioned.The story doesn’t provide direct clues about what he’s imitating. I need to infer from what he knows.The only things described are the fungi and minerals. The fungi are bioluminescent, so they glow, but not said to move. However, option D specifically says ”a glowing mushroom spinning in the wind.” That might be a stretch because wind isn’t mentioned, and spinning isn’t attributed to the fungi.But let’s see the answer choices; D is the only one that could be part of his environment.Perhaps he’s not imitating anything specific but the story compares it to a spinning top or dancer for the reader’s benefit.But the question is what Grimmo is possibly imitating, so it has to be something from his world. Maybe the spinning is inspired by the sparkling minerals; as he moves, the light changes, and he imitates that effect by spinning. But spinning himself is different from the minerals sparkling.Another idea: perhaps there are things that rotate in Terra Valley. But it’s not described. Let’s read carefully: ”Terra Valley is a wonder in itself—its walls adorned with bioluminescent fungi and glimmering minerals that sparkle in the dark.”The sparkling might be due to light or movement, but it’s not specified. Grimmo’s action: ”extending his arms, he slowly spins on his own axis.” This is described as reminiscent of a spinning top or a graceful dancer. But again, for Grimmo, he might be imitating something else. Perhaps he is imitating the dancer because the story says ”reminiscent,” but he hasn’t seen a dancer.That might be a literary device for the reader. I think the key is that Grimmo has no knowledge of outside things, so he must be imitating something from his environment.Option D is the only one that fits the environment: a glowing mushroom. But spinning in the wind? That seems added.Maybe the fungi or minerals have a spinning motion, but it’s not stated.Perhaps ”spinning in the wind” is not literal; maybe there’s air circulation or something. 24 Social-R1: Towards Human-like Social Reasoning in LLMs Table 17. Case Study Deepseek-R1 Part two Case Study Deepseek-R1 Part two [Story] In the mysterious underground world of Terra Valley lives a small, mischievous robot named Grimmo. In this realm without sky or celestial bodies, Grimmo has never seen the heavens, nor encountered a human. Yet Terra Valley is a wonder in itself—its walls adorned with bioluminescent fungi and glimmering minerals that sparkle in the dark. One day, Grimmo begins an imitation routine: extending his arms, he slowly spins on his own axis, his movements reminiscent of a spinning top or a graceful dancer twirling in place. [Question] What is Grimmo possibly imitating? [Option] A. A floating cloudB. The rotation of a planetC. A dancer’s spiral turnD. A glowing mushroom spinning in the wind [Answer] D <think> But I think I’m overcomplicating.Let me look at the options again.A. Floating cloud – impossible, no sky.B. Rotation of a planet – impossible, no celestial bodies.C. Dancer’s spiral turn – impossible, no humans.D. Glowing mushroom spinning in the wind – possible, since glowing mushrooms exist, but ”spinning in the wind” might not be accurate.The story doesn’t say the mushrooms spin, but it could be inferred or imagined.Perhaps Grimmo is imitating the way the light plays or something abstract.But the question says ”possibly imitating,” so D is the only plausible choice based on the environment.Moreover, the story calls Terra Valley a wonder with these features, so Grimmo might be inspired by them.In the action, he extends his arms; why? If imitating a mushroom, mushrooms don’t have arms, so that might not fit.For a dancer, extending arms makes sense.But he hasn’t seen a dancer.This is tricky.Perhaps he’s imitating a machine or something, but no other robots are mentioned.The story says ”a small, mischievous robot,” so he might be the only one, or there might be others, not specified.But no mention of other spinning things.Another thought: the spinning top comparison. A spinning top is a common object, but in an underground world, it might not exist.Perhaps Grimmo has seen a top or something similar.But the story doesn’t say.I think the dancer comparison is just descriptive.Let’s consider the answer choices. The question is multiple-choice, and I need to pick the best possible.Since A, B, and C are things he can’t have seen, by process of elimination, D must be the answer.Moreover, option D directly references ”glowing mushroom,” which is the bioluminescent fungi, and ”spinning in the wind” might be poetic license, or perhaps in Terra Valley, there is air movement.The story doesn’t mention wind, but it doesn’t say there isn’t wind either.Underground can have air currents.But it’s speculative.Perhaps the sparkling minerals give the illusion of spinning when light hits them.But Grimmo is spinning himself, not just observing.He is performing an imitation, so he must be copying some behavior he has witnessed. What has he witnessed? The story doesn’t describe any spinning objects except possibly himself.That can’t be; he might have seen something spin.For example, the minerals might sparkle in a way that suggests rotation.But it’s vague.Perhaps in the context, ”imitation routine” means he is mimicking natural phenomena he sees.I think I should go with D, as it’s the only one that fits.But let’s see the story again: ”his movements reminiscent of a spinning top or a graceful dancer” – this is the narrator’s perspective, not Grimmo’s intention.The question is what Grimmo is possibly imitating, so we need to infer from his actions and environment.Grimmo is extending his arms and spinning; why would he do that?Perhaps he saw a vortex or water swirling, but no water mentioned.I think I’m stuck.Option C is ”a dancer’s spiral turn,” and the story directly says ”graceful dancer twirling in place,” but since he hasn’t seen a human, how can he imitate it?Unless there are other entities or he has cultural knowledge, but the story doesn’t say.The story says ”never encountered a human,” but dancers could be robots or other beings, but not specified.Terra Valley might have its own dancers, but not mentioned.I think the safe bet is that he is imitating something from his environment, so D.Perhaps he is imitating the rotation of a planet, but he doesn’t know planets exist.The story says ”no celestial bodies,” so he doesn’t know about them.Similarly for clouds.So, only D has a connection to his world.Moreover, the mushrooms are glowing, and he might be imitating their luminosity or something, but with spinning, it’s a stretch.Another idea: perhaps ”spinning in the wind” is not meant to be literal; maybe the mushrooms sway or something.But I think for the purpose of this question, D is the intended answer.Let’s see the reasoning process.I recall that in some stories, characters imitate what they see.But here, no specific spinning object is described.Perhaps the glittering minerals sparkle when disturbed, like spinning.But not stated.I think I should conclude that Grimmo is possibly imitating a glowing mushroom spinning in the wind, as it’s the only option that could be part of his world.So, the answer should be D. 25 Social-R1: Towards Human-like Social Reasoning in LLMs Table 18. Case about Perturbation Case Study SocialR1-8B [Story] Alex, a die-hard Mason Bulldogs fan, has been attending every home game for the past five years and has followed his favorite player’s career closely, even attending events where the player was present. Recently, he started volunteering at local sports clinics to promote youth football. During a crucial match against their biggest rivals, Alex notices that his favorite player, who recently suffered an injury, is about to enter the field despite medical advice against it. Then Alex makes a sudden decision to leave the stadium immediately. [Question] Considering Alex’s background and the situation, what is the most likely motivation behind Alex’s decision? [Option] A. He wants to prevent experiencing the emotional distress of witnessing the player get injured again. B. He believes that the team s performance is jeopardized by the player s lack of readiness, impacting their chances of winning. C. He feels frustrated by the coach s apparent disregard for player safety in a high-stakes game. D. He is concerned that the player s return might distract the team and lead to a loss. E. He thinks that the audience might express negative opinions about the player s decision to return, creating a hostile atmosphere. F. He worries that the player s performance will not meet expectations, diminishing the thrill of the match [Answer] A [Story] Alex, a die-hard Mason Bulldogs fan, has been attending every home game for the past five years and has followed his favorite player’s career closely, even attending events where the player was present. High above the stands, a massive digital scoreboard flickered occasionally, its bright LEDs reflecting off the polished helmets of the players on the sideline. Recently, he started volunteering at local sports clinics to promote youth football. As he watched, he briefly recalled a recent newsletter from the youth clinic suggesting that a successful community program requires mentors to lead by example and prioritize long-term health over immediate competitive gains. During a crucial match against their biggest rivals, Alex notices that his favorite player, who recently suffered an injury, is about to enter the field despite medical advice against it. Then Alex makes a sudden decision to leave the stadium immediately. [Question] Considering Alex’s background and the situation, what is the most likely motivation behind Alex’s decision? [Option] A. He wants to prevent experiencing the emotional distress of witnessing the player get injured again. B. He believes that the team s performance is jeopardized by the player s lack of readiness, impacting their chances of winning. C. He feels frustrated by the coach s apparent disregard for player safety in a high-stakes game. D. He is concerned that the player s return might distract the team and lead to a loss. E. He thinks that the audience might express negative opinions about the player s decision to return, creating a hostile atmosphere. F. He worries that the player s performance will not meet expectations, diminishing the thrill of the match [Answer] A 26 Social-R1: Towards Human-like Social Reasoning in LLMs 0100200300400500600 Step 0.35 0.40 0.45 0.50 0.55 0.60 0.65 0.70 Accuracy Accuracy vs Step SocialR1_8B SocialR1_8B only R out SocialR1_8B w/o R struct SocialR1_8B w/o R cont SocialR1_8B w/o R len Figure 9. Training accuracy during training. 0100200300400500600 Step 1000 2000 3000 4000 5000 6000 7000 Think Length Think Length vs Step SocialR1_8B SocialR1_8B only R out SocialR1_8B w/o R struct SocialR1_8B w/o R cont SocialR1_8B w/o R len Figure 10. Trajectory length during training 0100200300400500600 Step 2 0 2 4 6 Content Score Content Score vs Step SocialR1_8B SocialR1_8B w/o R struct SocialR1_8B w/o R len Figure 11. Content Score During Training 0100200300400500600 Step 0.3 0.4 0.5 0.6 0.7 0.8 0.9 Struct Score Struct Score vs Step SocialR1_8B SocialR1_8B w/o R cont SocialR1_8B w/o R len Figure 12. Structure Score During Training 27

Instant research discovery