Paper deep dive
Learning to Simulate Individuals from Macro Social Signals
Yining Zhao, Bushi Liu, Haofei Yu, Zhengyang Qi, Shanyong Wang, Chuyue Li, Yuxiang Liu, Jiaxuan You
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 91%
Last extracted: 10/7/2026, 8:29:53 AM
Summary
The paper introduces macro2mind, a framework that trains large language models to simulate individual behavior by learning behavioral reasoning from macro-level prediction market signals. Using Group Relative Policy Optimization (GRPO), the model employs a social behavioral decomposition to infer constituent groups, their belief updates, and interactions, guided by a hindsight-regret curriculum. This approach achieves state-of-the-art performance on market forecasting benchmarks (SWM-Bench, Polymarket) and transfers zero-shot to user-simulation benchmarks (Humanual, OvertonBench, PRISM, CAD), significantly outperforming baselines in generating synthetic training data for downstream simulators.
Entities (13)
Relation Signals (10)
Prediction Markets â providesdatafor â macro2mind
confidence 95% ¡ We propose to learn behavioral reasoning from prediction markets... We introduce macro2mind, which trains a language model with GRPO using market signals.
macro2mind â usesalgorithm â GRPO
confidence 95% ¡ We introduce macro2mind, which trains a language model with GRPO using market signals.
macro2mind â usestechnique â hindsight-regret curriculum
confidence 93% ¡ A hindsight-regret curriculum with difficulty-aware sampling focuses training on transitions
macro2mind â usestechnique â social behavioral decomposition
confidence 93% ¡ A social behavioral decomposition makes behavioral reasoning an explicit step of forecasting
macro2mind â achievessotaon â SWM-Bench
confidence 92% ¡ On SWM-Bench, macro2mind achieves state-of-the-art directional accuracy and correlation on Polymarket.
macro2mind â outperforms â Qwen3-4B
confidence 90% ¡ Relative gains of macro2mind over its Qwen3-4B backbone... macro2mind also raises a downstream simulator's accuracy... outperforming data generated by its backbone by 13.2 points.
macro2mind â transfersto â Humanual
confidence 88% ¡ Trained on market data, it transfers zero-shot to four user-simulation benchmarks (Humanual...)
â â
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Large language models are increasingly used to simulate how individuals respond to new situations, yet the behavioral reasoning behind these responses is either inherited from pretraining or learned from individual-level annotations, which offer limited behavioral diversity and little supervision of the reasoning itself. We propose to learn behavioral reasoning from prediction markets, whose price trajectories record how populations respond to real-world events at scale. We introduce macro2mind, which trains a language model with GRPO using market signals. A social behavioral decomposition makes behavioral reasoning an explicit step of forecasting: the model infers representative groups of market participants, predicts how each interprets the news and updates its beliefs, reasons about their interactions, and aggregates these responses into a price. A hindsight-regret curriculum with difficulty-aware sampling focuses training on transitions where hindsight-identified groups substantially improve the forecast while prioritizing examples that remain learnable for the current policy. The learned reasoning applies to user simulation without further training. On SWM-Bench, macro2mind achieves state-of-the-art directional accuracy and correlation on Polymarket. Trained on market data, it transfers zero-shot to four user-simulation benchmarks (Humanual, OvertonBench, PRISM, and CAD) and has competitive performance among zero-shot methods. Used as a data generator, macro2mind also raises a downstream simulator's accuracy on unseen users by 15.5 points, outperforming data generated by its backbone by 13.2 points.
Tags
Links
- Source: https://arxiv.org/abs/2610.07062v1
- Canonical: https://arxiv.org/abs/2610.07062v1
Trouble viewing inline? Open PDF directly â
Full Text
100,363 characters extracted from source content.
Expand or collapse full text
Learning to Simulate Individuals from Macro Social Signals Yining Zhao â thanks: Core Contributors. Email: yiningz9@illinois.edu Affiliation: University of Illinois Urbana-Champaign Bushi Liu11footnotemark: 1 Affiliation: University of Illinois Urbana-Champaign Haofei Yu Affiliation: University of Illinois Urbana-Champaign Zhengyang Qi Affiliation: Independent Researcher Shanyong Wang Chuyue Li Yuxiang Liu Jiaxuan You Affiliation: University of Illinois Urbana-Champaign Affiliation: Independent Researcher Abstract Large language models are increasingly used to simulate how individuals respond to new situations, yet the behavioral reasoning behind these responses is either inherited from pretraining or learned from individual-level annotations, which offer limited behavioral diversity and little supervision of the reasoning itself. We propose to learn behavioral reasoning from prediction markets, whose price trajectories record how populations respond to real-world events at scale. We introduce macro2mind, which trains a language model with GRPO using market signals. A social behavioral decomposition makes behavioral reasoning an explicit step of forecasting: the model infers representative groups of market participants, predicts how each interprets the news and updates its beliefs, reasons about their interactions, and aggregates these responses into a price. A hindsight-regret curriculum with difficulty-aware sampling focuses training on transitions where hindsight-identified groups substantially improve the forecast while prioritizing examples that remain learnable for the current policy. The learned reasoning applies to user simulation without further training. On SWM-Bench, macro2mind achieves state-of-the-art directional accuracy and correlation on Polymarket. Trained on market data, it transfers zero-shot to four user-simulation benchmarks (Humanual, OvertonBench, PRISM, and CAD) and has competitive performance among zero-shot methods. Used as a data generator, macro2mind also raises a downstream simulatorâs accuracy on unseen users by 15.5 points, outperforming data generated by its backbone by 13.2 points. 1 Introduction Large language models are increasingly used to simulate individual behavior (Park et al., 2023; Zhou et al., 2024; Aher et al., 2023; Ni et al., 2026), including how people respond to information (Argyle et al., 2023; Santurkar et al., 2023), make decisions (Kitadai et al., 2025; Horton et al., 2023), and interact with others (Yu et al., 2025). Such simulators support personalized agents (Jiang et al., 2024; Ji et al., 2026), social simulation (Yang et al., 2024; Gao et al., 2023), and the evaluation of interactive systems (Wang et al., 2025; Yao et al., 2024). Achieving high-fidelity simulation of a particular individual imposes a dual requirement. It demands not only static, descriptive information about that personâs background but also a dynamic capability to predict how their internal beliefs and subsequent actions evolve. We refer to this capability as behavioral reasoning; it is closely related to theory of mind, the ability to attribute beliefs and intentions to others (Premack and Woodruff, 1978). To improve response fidelity, training-based simulators directly fit models to observed user behaviors (Naous et al., 2026; Zhou et al., 2026); some frameworks further align latent user states with these observed outputs (Wu et al., 2026). However, while rich profiles and histories successfully specify whom to simulate, the core mechanics of how they reason are either passively inherited from pre-training priors or explicitly fitted to individual-level annotations. Consequently, the supervision for dynamic behavioral reasoning remains strictly confined to micro-level data. Challenges of user simulation. Generalizing to unseen users and situations is the central challenge for current simulators, fundamentally bottlenecked by two data limitations. (1) Limited behavioral diversity. Capturing how people dynamically update beliefs and adapt to peers requires temporally dense, socially situated data. However, such data is exceedingly scarce: traditional surveys yield only static snapshots, massive micro-level collection is costly and privacy-sensitive (Gabriel et al., 2024; Durmus et al., 2023), and synthetic personas (Ge et al., 2024) lack grounding in real behavior. (2) Costly reasoning supervision. Supervising how people respond requires internal cognitive traces that are prohibitively expensive to collect. Relying on affordable alternatives introduces critical flaws: simplistic tasks encourage models to exploit superficial heuristics, while LLM-generated rationales often inject spurious reasoning disconnected from genuine cognition. Both prevent the acquisition of robust behavioral reasoning needed for generalization. Figure 1: Learning to simulate individuals from macro signals. Left: Each prediction-market price aggregates the behavioral reasoning of its participants: how people interpret an event given their prior beliefs and goals, update those beliefs, and respond, including to one another. Prices record such reasoning at every move, across diverse types of markets. Individual-level data offers only a few answers per person and no record of the reasoning behind them. macro2mind learns behavioral reasoning from market prices and applies it zero-shot to simulate individuals. Right: Relative gains of macro2mind over its Qwen3-4B backbone on market forecasting and zero-shot user simulation. For synthetic data, students are trained on data generated by each model. Market prices as aggregated individual behavior. While individual behavioral data is scarce, prediction markets continuously record the underlying reasoning at scale (Wolfers and Zitzewitz, 2004). Economists recognize prices not as abstract labels, but as aggregations of dispersed beliefs (Hayek, 2013; Manski, 2006). Consequently, each price movement is the macroscopic trace of a micro-level process: heterogeneous individuals interpret news through personal priors (Kandel and Pearson, 1995), adapt to peers (Bikhchandani et al., 1992), and aggregate into a new priceâmirroring the macroâmicroâmacro chain of social theory (Coleman and others, 1990). For example, when inflation unexpectedly drops, rate-cut probabilities surge because trading desks instantly adjust, analysts react to the desks, and contrarians bet against the herd. Explaining this shift requires the exact behavioral reasoning necessary for user simulation, simply executed in reverse: inferring latent constituent behaviors from aggregate outcomes. Prediction markets thus solve both data bottlenecks (Figure 1). (1) Diverse and abundant dynamics. Markets naturally capture varied reasoning patterns from macroeconomic updates to partisan interpretations through continuous, high-stakes decisions, bypassing costly micro-level data collection. (2) Outcome-grounded supervision. The realized price provides objective outcome-level supervision for whether the aggregate implication of the inferred responses matches the observed market movement. This raises our central question: can learning to predict aggregate outcomes teach behavioral reasoning that transfers to simulating individuals? Our approach. Harnessing this abundance is not a free lunch: evaluating only a final numerical forecast (Gu et al., 2025) introduces two learning challenges. First, naive optimization risks rewarding shortcut learning over genuine behavioral reasoning (C1). Second, market transitions lack explicit labels to distinguish learnable social dynamics from routine noise (C2). We introduce macro2mind, which tackles both challenges through two complementary mechanisms trained via GRPO on macroscopic rewards. First, a social behavioral decomposition explicitly structures the reasoning trajectory (C1). The model must infer representative groups, predict their belief updates and interactions, and aggregate these micro-responses into a price. Second, a hindsight-regret curriculum with difficulty-aware sampling performs dynamic data selection (C2). Given the realized outcome in hindsight, the policy names the driving groups; regret is measured by how much this hint improves its forecast. Hindsight regret identifies behaviorally informative transitions, while the difficulty-aware term increases the sampling weight of examples with useful within-group variation. Ultimately, the decomposition dictates how the model reasons, while the curriculum dictates where that learning is most informative. Main findings. We train macro2mind on prediction-market transitions from SWM-Bench and evaluate both its macro forecasting accuracy and zero-shot transfer to four micro-level benchmarks (Humanual, OvertonBench, PRISM, and CAD). Our experiments answer the central question affirmatively through three key findings: (1) Superior market forecasting. macro2mind achieves state-of-the-art directional accuracy and correlations on Polymarket and highly competitive performance on Kalshi. (2) Robust zero-shot user simulation. Despite zero exposure to individual-level annotations, macro2mind consistently achieves competitive performance among zero-shot methods across the downstream benchmarks. Remarkably, it even outperforms several supervised simulators trained explicitly on the target domains. (3) Scalable synthetic supervision. Behavioral trajectories generated by our zero-shot model serve as high-quality training data. Fine-tuning downstream user simulators on these synthesized responses boosts their accuracy by 15.5 absolute points, proving macro2mind to be a stronger teacher compared to its standard backbone. 2 Related Works LLM-based User Simulation. LLMs have been used as stand-ins for human respondents in surveys and behavioral experiments (Argyle et al., 2023; Aher et al., 2023; Horton et al., 2023), although their default opinions often diverge from those of specific populations (Santurkar et al., 2023; Durmus et al., 2023). To simulate particular individuals, existing methods condition on profiles, interviews, or interaction histories through prompting (Park et al., 2023; Park et al., 2024b; Ryan et al., 2025), steer activations toward a target persona (Feng et al., 2026), or fine-tune on behavioral traces (Naous et al., 2026; Wu et al., 2026; Zhou et al., 2026); reward models of human preferences, including personalized ones, learn from pairwise feedback (Cai et al., 2026; Wang et al., 2026). Synthetic personas (Ge et al., 2024) and digital-twin datasets (Yang et al., 2025b) broaden coverage, but the former lack grounding in real behavior and the latter remain costly to collect. All of these approaches learn how people respond from individual-level data. We instead learn behavioral reasoning from aggregate outcomes and use user-specific data only at inference. Aggregate Social Signals and Social Simulation. Forecasting systems treat aggregate signals such as prediction-market prices as the end goal (Wolfers and Zitzewitz, 2004; Halawi et al., 2024; Karger et al., 2025; Chen et al., 2025). Social simulation runs bottom-up, generating collective dynamics from rule-based (Schelling, 1971; Brock and Hommes, 1998) or LLM-based agents (Gao et al., 2023; Yang et al., 2024; Zhang et al., 2025), with aggregate data used only to calibrate or validate the agents (Fagiolo et al., 2007). Ecological inference (King and Fox, 1999) and learning from label proportions (Quadrianto et al., 2008) do infer individuals from aggregates, but within a fixed population and label space. We instead use aggregate outcomes to train a reasoning process that transfers to new individuals and situations, and, since aggregate fit does not guarantee individual-level validity (Robinson, 2009; Zhang et al., 2026c), we evaluate it directly on individual behavior. Theory of Mind Modeling. Existing Theory of Mind (ToM) frameworks and benchmarks heavily rely on synthetic narratives or explicit human annotations (Gandhi et al., 2023; Kim et al., 2023; Wilf et al., 2024; Sclar et al., 2023; Zhou et al., 2025), sparking debate over whether LLMs perform robust causal reasoning or superficial pattern matching (Kosinski, 2024; Ullman, 2023; Sap et al., 2022). Our constituent-level decomposition directly operationalizes ToM by inferring how distinct groups update latent beliefs, but it induces this capability from macroscopic market outcomes. By bypassing micro-level annotations entirely, we learn grounded mental-state reasoning that transfers zero-shot to concrete individuals. 3 Problem Formulation We formalize two related prediction tasks. The macro task predicts how a population-level signal evolves in response to real-world events. The micro task predicts how a specific individual responds to a given situation. We structure both tasks under a unified framework of observation, policy, and evaluation. The observation captures the context and history available at a given step. The policy Ďθ _θ maps the observation to a reasoning trace followed by a prediction of the next outcome. Because both tasks share this inputâoutput interface, a single policy Ďθ _θ can be trained only on the macro task and evaluated zero-shot on the micro task, with individual human responses reserved for testing. This setting directly tests our core question: can behavioral reasoning learned from macroscopic outcomes transfer to simulating individuals? 3.1 Macro Task: Population-level Forecasting Observation. Each instance centers on a market question, a natural-language proposition such as the outcome of an election. Its price ytâ[0,1]y_tâ[0,1] reflects the collective decisions of market participants at step t. The model observes the question, the recent price history, and a set of text descriptions of real-world events, such as news items. We denote this macro observation by xtx_t. Policy. Given xtx_t, the policy generates a reasoning trace Ďt _t followed by a forecast of the next price, (Ďt,y^t+1)âźĎθ(â âŁxt)( _t, y_t+1) _θ(¡ x_t). Evaluation. The macro dataset collects transitions across many markets, macro=(xt,yt+1)D_macro=\(x_t,y_t+1)\. The task is evaluated by how accurately the forecast y^t+1 y_t+1 matches the realized next price yt+1y_t+1, using MAE, MASE, directional accuracy, and Pearson correlation. 3.2 Micro Task: Individual-level Simulation Observation. Each instance targets a specific user. At step t, the model observes the userâs persona and interaction history hth_t, such as a profile, stated views, or past messages, together with the current context, such as a news event, an ongoing conversation, or a candidate reply to be judged. We denote the full micro observation by xtx_t; all of it is available before the target response. Policy. Given xtx_t, the policy likewise generates a reasoning trace Ďt _t followed by a prediction of the userâs response, (Ďt,y^t+1)âźĎθ(â âŁxt)( _t, y_t+1) _θ(¡ x_t). Evaluation. The micro dataset collects individual interactions, micro=(xt,yt+1)D_micro=\(x_t,y_t+1)\, where yt+1y_t+1 is the userâs actual response. The task is evaluated by how well y^t+1 y_t+1 matches yt+1y_t+1, using task-specific scores such as rubric-based alignment for free-text replies, or error and rank correlation for ratings. 4 Learning to Simulate Individuals from Macro Signals Figure 2: The pipeline of macro2mind, illustrated with a running example. (a) Where to learn: The regret-and-difficulty curriculum performs data selection by focusing on high-regret transitions, such as the price jump after inflation data came in lower than expected, while ignoring the noise. (b) How to reason: Through social behavioral decomposition, the model infers how different groups react to the news (first order) and to one another (second order), and aggregates their belief updates into a forecast rewarded by the realized price. (c) Zero-shot transfer: This learned reasoning ability is transferred zero-shot to user simulations, effectively modeling how an individual (e.g., a small-business owner) reacts to the CPI news and peer comments. We realize our approach through a single autoregressive language-model policy Ďθ _θ. We first describe how the policy structures social behavioral reasoning during forecasting (Section 4.1), how it is optimized via GRPO using a hindsight-regret curriculum (Section 4.2), and how this learned reasoning transfers to individual simulation (Section 4.3). 4.1 Social Behavioral Reasoning Social behavioral decomposition. We instantiate this reasoning as a single autoregressive trajectory, avoiding the cost and credit-assignment difficulties of a full multi-agent simulation (Yang et al., 2024). Given an observation x, the policy generates a reasoning trace Ď=(Ďprop,ĎF,ĎA)Ď=(Ď^prop,Ď^F,Ď^A) before outputting the forecast y y. The resulting decomposition has natural counterparts in individual simulation. 1. Constituent initialization (ĎpropĎ^prop). The policy identifies representative constituents likely to drive the market shift, and constructs a prior state for each that describes its role, objectives, and beliefs. Micro counterpart: inferring which constituents drive a given change, and what they care about, requires linking behavioral patterns to the kinds of people who produce them. This is the same personaâbehavior association that individual simulation relies on. 2. First-order responses (ĎFĎ^F). The policy predicts how each constituent updates its beliefs and responds based on its own prior and the observation. Micro counterpart: predicting a userâs belief update and immediate response to the current context. 3. Second-order responses and aggregation (ĎAĎ^A). The policy models the most consequential interactions, assessing how one constituentâs reaction alters anotherâs beliefs and actions, and then aggregates the adjusted responses into the forecast y y. Micro counterpart: adapting to othersâ reactions, as in replying within a multi-turn dialogue. Behavioral reasoning as a shared transition function. Both tasks rest on a shared behavioral transition function, which maps an agentâs prior state and context to a belief update and an observable response, and can be viewed as a computational form of theory of mind. The two tasks differ in the observability of the prior state. In the micro task, it is given by the userâs history. In the macro task, it is unobserved, and the policy must first infer the prior states of key constituents before predicting their responses. Because the behavioral transition function is shared across constituents and tasks, the behavioral reasoning ability that the policy learns from macro rewards can be applied to individual simulation without further training (Section 4.3). Orders of social behavioral reasoning. Social reactions to an event unfold in orders. At the first order, each group reacts directly to the news; at the second order, groups react to one anotherâs reactions, and such cascades can in principle continue further. Individuals behave the same way, responding both to events and to what others say about them. Our template reasons up to the second order and focuses on the few interactions that matter most, which captures effects such as persuasion and conformity while maintaining a concise autoregressive trajectory. This way of reasoning bounds the intermediate reasoning process without altering the final macroscopic prediction target. 4.2 Learning Social Behavioral Reasoning from Macro Signals Algorithm 1 Overview of macro2mind training with the hindsight-regret curriculum and difficulty-aware sampling 1: Dataset macroD_macro; initial policy Ďθ _θ; reward function R; group size M; probe pool size P; difficulty filter d; weights Îťreg,Îťdiff _reg, _diff 2: Trained parameters θ 3: Initialize wâĄ(x)â1w(x)â 1 for all xâmacrox _macro 4: for each training step do 5: if periodic probe then $âł$Ërefresh curriculum statistics 6: Sample a pool P of P transitions uniformly from macroD_macro 7: for each (x,y)â(x,y) do 8: câźĎθhint(â âŁx,y)c _θ^hint(¡ x,y) $âł$Ëhindsight hint 9: (Ďi,y^i)âźĎθ(â âŁx)( _i, y_i) _θ(¡ x), (ĎiH,y^iH)âźĎθ(â âŁx,c)( _i^H, y_i^H) _θ(¡ x,c), i=1,âŚ,Mi=1,âŚ,M 10: ÎâĄ(x)â1Mââi=1MR~â(y^iH)â1Mââi=1MR~â(y^i) (x)â 1M _i=1^M R( y_i^H)- 1M _i=1^M R( y_i) 11: p(x)â1Mâi=1M[R~(y^i)âĽ0.5]p(x)â 1M _i=1^M 1\! [ R( y_i)⼠0.5 ] 12: wâĄ(x)âÎťregâmaxâĄ(ÎâĄ(x),0)+Îťdiffâdâ(pâĄ(x))w(x)â _reg ( (x),0)+ _diff\,d(p(x)) 13: âŹâźmacroâwâĄ(x)B _macro w(x) $âł$Ësample training batch 14: for each (x,y)ââŹ(x,y) do 15: (Ďi,y^i)âźĎθ(â âŁx)( _i, y_i) _θ(¡ x), i=1,âŚ,Mi=1,âŚ,M 16: p(x)â1Mâi=1M[R~(y^i)âĽ0.5]p(x)â 1M _i=1^M 1\! [ R( y_i)⼠0.5 ], wâĄ(x)âÎťregâmaxâĄ(Îâ(x)cached,0)+Îťdiffâdâ(pâĄ(x))w(x)â _reg ( (x)_cached,0)+ _diff\,d(p(x)) 17: θâGRPO-Updateâ(θ,âŹ,R)θâ GRPO-Update(θ,B,R) $âł$Ëgroup-normalized advantages 18: return θ Macro signals are inherently noisy and uneven in difficulty (C2): many transitions are dominated by random fluctuations, some are trivially solvable, and only some actually reward reasoning about the latent constituents driving the market. Because GRPO gains minimal signal from transitions the model always or never solves, we augment policy optimization with a hindsight-regret curriculum that combines behavioral informativeness with current learning difficulty. Hindsight regret prioritizes transitions where the model overlooks key constituents, while a difficulty-aware filter favors examples that have informative within-group variation for GRPO. Policy optimization. For a transition âmacro _macro, the policy samples a group of M outputs (Ďi,y^i)âźĎθ(â âŁx)( _i, y_i) _θ(¡ x), i=1,âŚ,Mi=1,âŚ,M, each consisting of a reasoning trace Ďi _i and a forecast y^i y_i. To bootstrap the structured reasoning format and stabilize subsequent reinforcement learning (Guo et al., 2025; Li et al., 2026b), we first perform a lightweight supervised fine-tuning (SFT) warmup on a small seed set of structured behavioral traces. For a given market observation x, the policy samples a group of M reasoning and forecasting trajectories. Each output is evaluated by a reward R that combines formatting, numerical accuracy, and directional accuracy. We optimize the policy via GRPO (Shao et al., 2024) using group-normalized advantages. To isolate forecasting capability, the curriculum below relies on a format-free reward function R~ R, which drops the format term and renormalizes the remaining terms, so that it measures forecast quality alone. Hindsight-regret curriculum. We formulate the curriculum as a dynamic data sampler. Periodically, we probe a random subset of transitions to estimate their learning value. For a given transition, the current policy is provided the true outcome y via a dedicated hint prompt and tasked with identifying the key constituents that drove the market. This generates a hint c containing constituent names and their qualitative price directions (up or down only). We inject c into the standard prompt as a simulated analystâs insight, and sample M hinted outputs (ĎiH,y^iH)âźĎθ(â âŁx,c)(Ď^H_i, y^H_i) _θ(¡ x,c) alongside M unhinted outputs (Ďi,y^i)âźĎθ(â âŁx)( _i, y_i) _θ(¡ x). We compute the hindsight regret Î and the baseline win rate p, estimating both from these samples: Î=âĄ[R~â(y^H)]ââĄ[R~â(y^)],p=PrâĄ(R~â(y^)âĽ0.5). =E [ R( y^H) ]-E [ R( y) ],\;p= ( R( y)⼠0.5 ). (1) Here, Î isolates the capability gap: what the policy can successfully explain when aware of the key constituents, yet fails to anticipate without them. Because c is generated by the policy itself, this gap shifts as the model improves. Sampling weights. We combine these two statistics into a composite sampling weight, adapting the designer reward formulation from SPADE (Liu et al., 2026a): w=ÎťregâmaxâĄ(Î,0)+Îťdiffâdâ(p),w= _reg ( ,0)+ _diff\,d(p), (2) where Îťreg _reg and Îťdiff _diff control the relative importance of each component. These terms serve complementary purposes: the regret term evaluates whether a transition is informative, while the difficulty term determines whether it is currently learnable. Specifically, the regret term prioritizes transitions whose outcomes hinge on critical constituents that the policy currently overlooks. We floor this value at zero because a negative regret implies that the self-generated hint misled the policy, rather than exposing a genuine capability gap. The difficulty term accounts for the optimization dynamics of GRPO. Because GRPO learns from the relative differences among M responses to the same transition, the advantage signals are strongest when a group contains a mix of successes and failures. This variance peaks at intermediate success rates and vanishes when the policy consistently succeeds or fails, regardless of the regret magnitude. Consequently, we define dâĄ(p)d(p) as a trapezoidal filter that equals 1 for pâ[plo,phi]pâ[p_lo,p_hi] and decays linearly to 0 as pâ0pâ 0 or pâ1pâ 1. Together, the two terms favor transitions that are both behaviorally informative and within the policyâs current learning capacity. Transitions not yet probed receive the mean weight of the probed ones, and the win rate p is also updated during regular GRPO steps. Each training batch âŹB is sampled from macroD_macro with probability proportional to w. Algorithm 1 gives an overview of the training procedure. 4.3 Applying Social Behavior Reasoning to Individual Simulation The core premise of our approach is that macroscopic market shifts and microscopic user actions are driven by the same underlying behavioral mechanics. Consequently, we transfer the reasoning capabilities learned from macro signals directly to individual simulation. Individual-level social reasoning. For individual simulation, the prior state is given rather than inferred. Depending on the task, it comes from the userâs interaction history, such as past posts or conversations, or from a persona, such as a profile and stated views. We therefore bypass constituent discovery and directly instantiate the prior state of the target individual with this information. Given this prior state and the current context, the policy applies the same learned behavioral transition to infer how the user interprets the context, updates their beliefs, and responds. Inference. We apply the trained policy to individual simulation zero-shot, without any parameter updates, and evaluate only the final response. During training, the policy reasons about coarse groups inferred from market context; here, it applies the same behavioral reasoning process to a specific person with an observed history or persona. Synthetic data. The same procedure can also generate training data to address the scarcity of individual behavioral data. Given a userâs persona and a few-shot history, optionally augmented with population-level behavioral statistics available in the downstream setting, we apply the trained policy to generate both novel questions and the corresponding responses. This synthetic data is then used to fine-tune and align a separate downstream simulator. 5 Main Experiments Table 1: Results on SWM-Bench. Bold / underline: best / second best results. Method Polymarket Kalshi MASEâ MAEâ DAâ Corrâ MASEâ MAEâ DAâ Corrâ Time-series only DLinear (Zeng et al., 2023) 1.111 0.048 0.580 0.348 1.245 0.075 0.492 â-0.035 TimeMixer (Wang et al., 2024a) 0.997 0.043 0.590 0.224 1.079 0.065 0.549 â-0.056 PatchTST (Nie et al., 2022) 0.994 0.043 0.592 0.306 1.174 0.071 0.536 â-0.035 Time-series + news (Prompting-based LLM) Qwen3-4B (Yang et al., 2025a) 2.297 0.099 0.584 0.088 1.848 0.112 0.559 0.098 Qwen3-4B-think (Yang et al., 2025a) 2.602 0.112 0.511 0.075 1.951 0.118 0.529 0.030 GPT-5.5 (OpenAI, 2026) 0.985 0.042 0.337 0.264 0.997 0.060 0.564 0.242 Time-series + news (Fine-tuned LLM) FNF (Wang et al., 2024b) 0.986 0.042 0.377 0.218 1.126 0.068 0.430 â-0.020 LLM4TS (Chang et al., 2025) 0.993 0.043 0.589 0.312 1.125 0.068 0.554 â-0.015 SWM (Yu et al., 2026) 1.141 0.049 0.535 0.088 1.013 0.061 0.589 0.167 GRPO-8B 1.044 0.045 0.475 0.012 1.001 0.060 0.448 0.031 macro2mind(4B) 0.999 0.043 0.674 0.362 1.088 0.066 0.569 0.016 5.1 Problem Setup Market prediction. We train and evaluate our model on SWM-Bench (Yu et al., 2026), which pairs prediction-market trajectories with contemporaneous news. The task is to predict the next observed market price, providing aggregate supervision without requiring any annotations of the underlying individual behaviors. User Simulation. To evaluate the zero-shot transfer of behavioral reasoning across both single-turn and multi-turn interactions, we select four benchmarks spanning three distinct individual-level tasks: free-text response generation (Humanual (Wu et al., 2026), Qwen3.8-Max as judge), representation judgment (OvertonBench (Poole-Dayan et al., 2026)), and conversational preference prediction (PRISM (Kirk et al., 2024)11 1 Following the original paper, we multiply the prompted 0â1 scores by 100 to report results on a 0â100 scale. and CAD (Zhang et al., 2026a)). Synthetic data. We evaluate on Twin-2K-500 (Toubia et al., 2025). Each generator writes new scenarios for 600 training users and labels them by answering as each participant, conditioned on the persona and the answer distribution of training users. A Qwen3-4B student is fine-tuned on each synthetic set and evaluated on real held-out questions, either for 100 unseen users or for the source users themselves. Details are in Appendix F. 5.2 Baseline settings. Market Prediction. We compare macro2mind against three categories of methods. (1) Time-series only baselines forecast solely from the numerical historical window and are trained on SWM-Bench: DLinear, PatchTST and TimeMixer. (2) Prompting-based LLMs process both the numerical window and all available news to generate forecasts via zero-shot prompting without parameter updates: Qwen3-4B, GPT-5.5. (3) Fine-tuned LLMs also use both modalities but are trained directly on SWM-Bench. SWM, FNF, LLM4TS and GRPO all use Qwen3-8B. macro2mind uses Qwen3-4B as backbone. User Simulation. We compare macro2mind against four categories of methods. (1) Same-backbone baselines: zero-shot Qwen3-4B and standard GRPO trained on SWM-Bench. (2) Prompting-based simulators: training-free inference methods, including Qwen3-4B, PERSONA-Flow, SynthesizeMe, Self-consistency, History-ICL and Self-report. (3) Trained user simulators: fine-tuned models including HumanLM-SFT, -SFT-think, and -GRPO. (4) Personalized reward models, repurposed as simulators: the generative P-GenRM, alongside the scalar models Skywork-RM, MRM, and WIMHF-RM. Scalar reward models generate no text and instead rerank eight Qwen3-4B samples. All methods use Qwen3-4B except Skywork-RM, and MRM. Details in Appendix. D. Table 2: Results on user simulation tasks. Parentheses denote training data (per bench.: one model trained per benchmark). Bold / underline: best / second best among zero-shot (out-of-domain) results. We multiply all Humanual scores by 100 for reporting. Humanual Overton PRISM CAD Method Stateâ Resp.â MAEâ ĎâĎ MAEâ Acc.â Acc.â Trained user simulators HumanLM-SFT (Humanual) (Wu et al., 2026) 8.2 3.2 0.734 0.566 22.5 59.6 35.0 HumanLM-SFT-think (Humanual) 6.3 1.8 0.754 0.606 23.3 60.1 36.5 HumanLM-GRPO (Humanual) 12.4 5.3 0.718 0.576 21.8 62.6 39.1 Personalized reward models P-GenRM (PRISM) (Zhang et al., 2026b) 3.4 0.7 0.786 0.455 21.8 66.2 44.0 Skywork-RM (pretrained) (Liu et al., 2026b) 8.3 2.4 1.024 0.090 21.5 64.7 41.8 MRM (per bench.) (Cai et al., 2026) 7.8 2.5 0.907 0.338 20.7 69.1 44.2 WIMHF-RM (per bench.) (Movva et al., 2026) 7.9 2.6 0.859 0.429 20.7 69.5 43.6 Prompting-based simulators PERSONA-Flow (Feng et al., 2026) 7.5 2.1 0.756 0.540 21.5 62.5 38.8 SynthesizeMe (Ryan et al., 2025) 7.7 2.2 0.763 0.529 21.3 63.2 40.4 Self-cons. (n=5n=5) (Wang et al., 2022) 7.9 2.1 0.758 0.538 21.3 62.6 38.6 History-ICL (Hwang et al., 2023) 8.8 2.9 0.741 0.552 20.5 63.6 42.9 Self-report (Park et al., 2024a) 6.9 1.8 0.750 0.544 21.5 62.7 38.5 Same-backbone controls Qwen3-4B (zero-shot) 7.7 2.4 0.756 0.539 21.5 62.4 38.9 GRPO-4B (SWMBench, zero-shot) 6.1 1.5 0.623 0.647 23.0 53.9 36.6 macro2mind (SWMBench, zero-shot) 9.5 3.4 0.635 0.653 21.2 63.3 42.2 5.3 Evaluation settings. For Market Prediction, we evaluate Mean Absolute Scaled Error (MASE), Mean Absolute Error (MAE), Directional Accuracy (DA) and Pearson Correlation (Corr). For User Simulation, we evaluate predictions against ground-truth behaviors. We report state and response alignment via LLM-as-a-judge on Humanual-Politics; MAE and correlations Ď on OvertonBench; MAE and preference accuracy on PRISM. Preference accuracy on CAD. Details in Appendix C.2. 5.4 Main Results macro2mind improves market forecasting, particularly in predicting directional movement. On SWM-Bench (Table 1), macro2mind achieves the highest directional accuracy (0.674) and correlation (0.362) on Polymarketâsurpassing both the strongest time-series baseline and GPT-5.5âwhile ranking second in directional accuracy on Kalshi (0.569). macro2mind demonstrates robust zero-shot transfer. Trained on prediction market signals, macro2mind has competitive performance among zero-shot methods (Table 2). It secures the best Humanual alignment (9.5/3.4) and OvertonBench Ď (0.653), remarkably surpassing two supervised HumanLM variants zero-shot. Furthermore, it firmly out-generalizes domain-specific simulators: macro2mind beats Humanual-trained HumanLM-GRPO across all five out-of-domain metrics (e.g., CAD: 42.2% vs. 39.1%) and avoids PRISM-trained P-GenRMâs catastrophic collapse on Humanual. On PRISM and CAD accuracy, macro2mind remains highly competitive (63.3% / 42.2%), trailing only History-ICL and dataset-specific reward models. macro2mind serves as a highly effective synthetic data generator for downstream user simulators. Fine-tuning a student model on data generated by macro2mind boosts its accuracy from 44.5% to 60.0% on unseen users (out-of-user) and from 50.9% to 61.3% on source users (in-user). In contrast, data generated by its Qwen3-4B backbone using the identical pipeline yields only marginal improvements (46.8% and 53.0%, respectively). Since the two generators differ exclusively in their exposure to macro market forecasting, this 13.2-point gap on unseen users demonstrates that macro2mind is a good teacher for micro-level simulation. 6 Discussion Figure 3: Component ablation results on Humanual. We use Qwen3-4B as base model. Figure 4: Training dynamics. SWM-Bench DA and Humanual Response score across training checkpoints. Figure 5: Human evaluation of reasoning plausibility. Pearson r=0.885r=0.885. RQ1: Which part of the training recipe drives transfer? Ablation results show that the full curriculum, combining hindsight regret with difficulty-aware sampling, is the key driver of zero-shot transfer. To isolate their contributions, we evaluate additive ablations on Humanual: Base+BR (zero-shot prompting), BR-SFT (fine-tuning on BR traces), BR-SFT+GRPO (GRPO without curriculum), Base+GRPO (GRPO with free form thinking) and macro2mind (full curriculum) under identical hyperparameters. Results (Figure 5) show the BR structure alone cannot explain the gains: Base+BR and BR-SFT yield similar State scores (7.6 vs. 7.5). Neither BR-SFT nor BR-SFT+GRPO improves over Base+BR: BR-SFT stays at 7.5 State and 1.4 Resp, and adding GRPO without the curriculum gives 6.3 State and 0.9 Resp. Thus, naively optimizing for market forecasting does not inherently transfer to individual simulation. Meaningful transfer emerges only when adding the full regret-and-difficulty curriculum: macro2mind reaches 9.5 State and 3.4 Resp, improving over Base+BR by 25% in State score. RQ2: Do macro forecasting and micro simulation ability improve together during training? Results indicate a concurrent improvement during training. As macro2mind learns to forecast macro signals, its zero-shot user-simulation capability exhibits corresponding gains. Evaluating four training checkpoints (steps 0, 20, 40, and 66) on both SWM-Bench and Humanual (Figure 5) reveals a striking trajectory. Over the first 40 steps, both abilities surge simultaneouslyâSWM-Bench directional accuracy increases from 50.2% to 64.7%, while Humanual response alignment more than doubles (1.2 â 2.6) despite zero exposure to its data. Beyond step 40, forecasting accuracy plateaus (64.9%), yet response alignment continues to climb to nearly three times its initial value (3.4). This continuous co-improvement demonstrates that zero-shot transfer is driven by the acquisition of macro behavioral reasoning. RQ3: Does the learned behavioral reasoning align with human judgments? Human evaluation indicates that the behavioral reasoning generated by macro2mind aligns well with human judgments. Five annotators evaluated 20 traces per model (0â100 scale), judging whether the inferred groups, news interpretations, and belief updates coherently explained the final forecast. As shown in Figure 5, macro2mind scores 72.2, surpassing both Qwen3-4B (47.1) and Qwen3.5-397B (51.4) under the exact same structural prompt. This suggests that structural prompting alone is insufficient to achieve human-aligned reasoning. In contrast, an R1-style model reasoning freely within <think> tags scores 18.3, often defaulting to disjointed numerical speculation rather than modeling actual human perspectives. These human ratings correlate strongly with our LLM judge (Pearson r=0.885r=0.885). 7 Conclusion We introduced macro2mind, which learns to simulate individual human behavior from aggregate prediction-market signals via a structured social behavioral decomposition and a dynamic hindsight-regret curriculum. macro2mind achieves state-of-the-art DA and Corr on Polymarket while demonstrating robust zero-shot transfer to user simulation. Furthermore, its generated data can serve as high-quality supervision, significantly boosting the accuracy of downstream models. In the future, exploring other forms of macroscopic supervision and deploying these simulators into dynamic, multi-agent ecosystems remain exciting directions. AI use statement AI systems, including LLMs, did not contribute to the development of the research ideas or the writing of this paper to a degree that would warrant authorship or contributor status. In this work, LLMs are treated strictly as objects of study and are the primary focus of our experimental evaluations. Ethics Statement This work introduces macro2mind, it is a methodological contribution rather than a deployed system, and its societal impact is largely mediated by how downstream practitioners apply it. All experiments are conducted on standard, publicly available benchmarks and open-weight or proprietary models used strictly for research evaluation; the work does not involve the collection of new personal data. Reproducibility statement We describe the full training and evaluation setup needed to reproduce our main results. Evaluation metrics are formally defined in Appendix §C. We will release our training code and configurations upon acceptance. References Aher et al. (2023) G. V. Aher, R. I. Arriaga, and A. T. Kalai Using large language models to simulate multiple humans and replicate human subject studies. In International conference on machine learning, p. 337â371. Cited by: §1, §2. Argyle et al. (2023) L. P. Argyle, E. C. Busby, N. Fulda, J. R. Gubler, C. Rytting, and D. Wingate Out of one, many: using language models to simulate human samples. Political Analysis 31 (3), p. 337â351. Cited by: §1, §2. Bikhchandani et al. (1992) S. Bikhchandani, D. Hirshleifer, and I. Welch A theory of fads, fashion, custom, and cultural change as informational cascades. Journal of political Economy 100 (5), p. 992â1026. Cited by: §1. Brock and Hommes (1998) W. A. Brock and C. H. Hommes Heterogeneous beliefs and routes to chaos in a simple asset pricing model. Journal of Economic dynamics and Control 22 (8-9), p. 1235â1274. Cited by: §2. Cai et al. (2026) H. Cai, Y. Li, T. Yu, F. Zhu, W. Wang, F. Feng, and W. Li One adapts to any: meta reward modeling for personalized llm alignment. arXiv preprint arXiv:2601.18731. Cited by: §2, Table 2. Chang et al. (2025) C. Chang, W. Wang, W. Peng, and T. Chen Llm4ts: aligning pre-trained llms as data-efficient time-series forecasters. ACM Transactions on Intelligent Systems and Technology 16 (3), p. 1â20. Cited by: Table 1. Chen et al. (2025) X. Chen, S. Wang, C. Qian, H. Wang, P. Han, and H. Ji Decisionflow: advancing large language model as principled decision maker. arXiv preprint arXiv:2505.21397. Cited by: §2. Coleman et al. (1990) J. S. Coleman et al. Foundations of social theory. Vol. 302, HeinOnline. Cited by: §1. Durmus et al. (2023) E. Durmus, K. Nguyen, T. I. Liao, N. Schiefer, A. Askell, A. Bakhtin, C. Chen, Z. Hatfield-Dodds, D. Hernandez, N. Joseph, et al. Towards measuring the representation of subjective global opinions in language models. arXiv preprint arXiv:2306.16388. Cited by: §1, §2. Fagiolo et al. (2007) G. Fagiolo, A. Moneta, and P. Windrum A critical guide to empirical validation of agent-based models in economics: methodologies, procedures, and open problems. Computational Economics 30 (3), p. 195â226. Cited by: §2. Feng et al. (2026) X. Feng, L. Zhao, W. Zhong, Y. Gu, L. Kong, X. Feng, B. Qin, et al. PERSONA: dynamic and compositional inference-time personality control via activation vector algebra. In International Conference on Learning Representations, Vol. 2026, p. 151839â151879. Cited by: §2, Table 2. Gabriel et al. (2024) I. Gabriel, A. Manzini, G. Keeling, L. A. Hendricks, V. Rieser, H. Iqbal, N. TomaĹĄev, I. Ktena, Z. Kenton, M. Rodriguez, et al. The ethics of advanced ai assistants. arXiv preprint arXiv:2404.16244. Cited by: §1. Gandhi et al. (2023) K. Gandhi, J. Fränken, T. Gerstenberg, and N. Goodman Understanding social reasoning in language models with language models. Advances in Neural Information Processing Systems 36, p. 13518â13529. Cited by: §2. Gao et al. (2023) C. Gao, X. Lan, Z. Lu, J. Mao, J. Piao, H. Wang, D. Jin, and Y. Li S3: social-network simulation system with large language model-empowered agents. arXiv preprint arXiv:2307.14984. Cited by: §1, §2. Ge et al. (2024) T. Ge, X. Chan, X. Wang, D. Yu, H. Mi, and D. Yu Scaling synthetic data creation with 1,000,000,000 personas. arXiv preprint arXiv:2406.20094. Cited by: §1, §2. Gu et al. (2025) S. Gu, C. Li, B. Jing, and K. Ren VerbalTS: generating time series from texts.. In ICML, Cited by: §1. Guo et al. (2025) D. Guo, D. Yang, H. Zhang, J. Song, P. Wang, Q. Zhu, R. Xu, R. Zhang, S. Ma, X. Bi, et al. Deepseek-r1: incentivizing reasoning capability in llms via reinforcement learning. arXiv preprint arXiv:2501.12948. Cited by: §4.2. Halawi et al. (2024) D. Halawi, F. Zhang, C. Yueh-Han, and J. Steinhardt Approaching human-level forecasting with language models. External Links: 2402.18563 Cited by: §2. Hayek (2013) F. A. Hayek The use of knowledge in society. In Modern understandings of liberty and property, p. 27â38. Cited by: §1. Horton et al. (2023) J. J. Horton, A. Filippas, and B. S. Manning Large language models as simulated economic agents: what can we learn from homo silicus?. Technical report National Bureau of Economic Research. Cited by: §1, §2. Hwang et al. (2023) E. Hwang, B. Majumder, and N. Tandon Aligning language models to user opinions. In Findings of the Association for Computational Linguistics: EMNLP 2023, p. 5906â5919. Cited by: Table 2. Ji et al. (2026) Z. Ji, L. Jin, S. Wang, J. Lu, C. Lu, Y. Wu, Y. Hu, L. Cui, and Y. Xu Momentum-coupled rubric adaptation for detailed image captioning. arXiv preprint arXiv:2609.36893. Cited by: §1. Jiang et al. (2024) H. Jiang, X. Zhang, X. Cao, C. Breazeal, D. Roy, and J. Kabbara PersonaLLM: investigating the ability of large language models to express personality traits. In Findings of the association for computational linguistics: NAACL 2024, p. 3605â3627. Cited by: §1. Kandel and Pearson (1995) E. Kandel and N. D. Pearson Differential interpretation of public signals and trade in speculative markets. Journal of political Economy 103 (4), p. 831â872. Cited by: §1. Karger et al. (2025) E. Karger, H. Bastani, C. Yueh-Han, Z. Jacobs, D. Halawi, F. Zhang, and P. E. Tetlock Forecastbench: a dynamic benchmark of ai forecasting capabilities. In International Conference on Learning Representations, Vol. 2025, p. 93943â93980. Cited by: §2. Kim et al. (2023) H. Kim, M. Sclar, X. Zhou, R. Bras, G. Kim, Y. Choi, and M. Sap FANToM: a benchmark for stress-testing machine theory of mind in interactions. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, p. 14397â14413. Cited by: §2. King and Fox (1999) G. King and J. Fox A solution to the ecological inference problem: reconstructing individual behavior from aggregate data. Canadian Journal of Sociology 24 (1), p. 150. Cited by: §2. Kirk et al. (2024) H. R. Kirk, A. Whitefield, P. RĂśttger, A. Bean, K. Margatina, J. Ciro, R. Mosquera, M. Bartolo, A. Williams, H. He, et al. The prism alignment dataset: what participatory, representative and individualised human feedback reveals about the subjective and multicultural alignment of large language models. Advances in Neural Information Processing Systems 37, p. 105236â105344. Cited by: §E.3.1, §5.1. Kitadai et al. (2025) A. Kitadai, S. D. Rico Lugo, Y. Tsurusaki, Y. Fukasawa, and N. Nishino Can ai with high reasoning ability replicate human-like decision making in economic experiments?. Group Decision and Negotiation 34 (6), p. 1303â1326. Cited by: §1. Kosinski (2024) M. Kosinski Evaluating large language models in theory of mind tasks. Proceedings of the National Academy of Sciences 121 (45), p. e2405460121. Cited by: §2. Li et al. (2026a) C. Li, Z. Tang, J. Wang, Y. Wu, K. Hashimoto, and L. Gao PyMETA: a benchmark dataset for hierarchical student code error classification with python-interpreter-based labels. arXiv preprint arXiv:2606.30610. Cited by: §C.2. Li et al. (2026b) C. Li, J. Yu, H. Wang, T. Xueyun, Z. Zhang, B. Li, S. Gu, K. Ren, J. Liu, and R. Huang AVA-encoder: towards agent-native video representation learning. arXiv preprint arXiv:2608.12313. Cited by: §4.2. Liu et al. (2026a) B. Liu, S. Yu, Y. Jiang, A. Qu, A. Zhao, Z. Liu, J. Kim, Z. Zhou, S. Kim, T. Ren, et al. SPADE: self-play in adaptive synthetic executable environments. arXiv preprint arXiv:2608.19197. Cited by: §4.2. Liu et al. (2026b) Y. Liu, L. Zeng, Y. Xiao, J. He, J. Liu, C. Wang, R. Yan, W. Shen, F. Zhang, J. Xu, et al. Skywork-reward-v2: scaling preference data curation via human-ai synergy. In International Conference on Learning Representations, Vol. 2026, p. 133805â133838. Cited by: Table 2. Manski (2006) C. F. Manski Interpreting the predictions of prediction markets. economics letters 91 (3), p. 425â429. Cited by: §1. Movva et al. (2026) R. Movva, S. Milli, S. Min, and E. Pierson Whatâs in my human feedback? learning interpretable descriptions of preference data. In International Conference on Learning Representations, Vol. 2026, p. 59420â59455. Cited by: Table 2. Naous et al. (2026) T. Naous, P. Laban, W. Xu, and J. Neville Flipping the dialogue: training and evaluating user language models. In International Conference on Learning Representations, Vol. 2026, p. 82079â82103. Cited by: §1, §2. Ni et al. (2026) B. Ni, Y. Wang, L. Wang, B. Kveton, F. Dernoncourt, Y. Xia, H. Chen, R. Luera, S. Basu, S. Mukherjee, et al. A survey on llm-based conversational user simulation. In Proceedings of the 19th Conference of the European Chapter of the Association for Computational Linguistics (Volume 1: Long Papers), p. 4266â4301. Cited by: §1. Nie et al. (2022) Y. Nie, N. H. Nguyen, P. Sinthong, and J. Kalagnanam A time series is worth 64 words: long-term forecasting with transformers. arXiv preprint arXiv:2211.14730. Cited by: Table 1. OpenAI (2026) OpenAI Introducing gpt-5.5. Note: https://openai.com/index/introducing-gpt-5-5/Accessed: 2026-05-29 Cited by: Table 1. Park et al. (2023) J. S. Park, J. OâBrien, C. J. Cai, M. R. Morris, P. Liang, and M. S. Bernstein Generative agents: interactive simulacra of human behavior. In Proceedings of the 36th annual acm symposium on user interface software and technology, p. 1â22. Cited by: §1, §2. Park et al. (2024a) J. S. Park, C. Q. Zou, J. Kamphorst, N. Egan, A. Shaw, B. M. Hill, C. Cai, M. R. Morris, P. Liang, R. Willer, et al. LLM agents grounded in self-reports enable general-purpose simulation of individuals. arXiv preprint arXiv:2411.10109. Cited by: Table 2. Park et al. (2024b) J. S. Park, C. Q. Zou, A. Shaw, B. M. Hill, C. Cai, M. R. Morris, R. Willer, P. Liang, and M. S. Bernstein Generative agent simulations of 1,000 people. arXiv preprint arXiv:2411.10109 52. Cited by: §2. Poole-Dayan et al. (2026) E. Poole-Dayan, J. Wu, T. Sorensen, J. Pei, and M. Bakker Benchmarking overton pluralism in llms. In International Conference on Learning Representations, Vol. 2026, p. 40773â40812. Cited by: §E.3.1, §5.1. Premack and Woodruff (1978) D. Premack and G. Woodruff Does the chimpanzee have a theory of mind?. Behavioral and brain sciences 1 (4), p. 515â526. Cited by: §1. Quadrianto et al. (2008) N. Quadrianto, A. J. Smola, T. S. Caetano, and Q. V. Le Estimating labels from label proportions. In Proceedings of the 25th International Conference on Machine learning, p. 776â783. Cited by: §2. Robinson (2009) W. S. Robinson Ecological correlations and the behavior of individuals. International journal of epidemiology 38 (2), p. 337â341. Cited by: §2. Ryan et al. (2025) M. J. Ryan, O. Shaikh, A. Bhagirath, D. Frees, W. Held, and D. Yang SynthesizeMe! inducing persona-guided prompts for personalized reward models in llms. In Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), p. 8045â8078. Cited by: §2, Table 2. Santurkar et al. (2023) S. Santurkar, E. Durmus, F. Ladhak, C. Lee, P. Liang, and T. Hashimoto Whose opinions do language models reflect?. In International conference on machine learning, p. 29971â30004. Cited by: §1, §2. Sap et al. (2022) M. Sap, R. Le Bras, D. Fried, and Y. Choi Neural theory-of-mind? on the limits of social intelligence in large lms. In Proceedings of the 2022 conference on empirical methods in natural language processing, p. 3762â3780. Cited by: §2. Schelling (1971) T. C. Schelling Dynamic models of segregation. Journal of mathematical sociology 1 (2), p. 143â186. Cited by: §2. Sclar et al. (2023) M. Sclar, S. Kumar, P. West, A. Suhr, Y. Choi, and Y. Tsvetkov Minding language modelsâ(lack of) theory of mind: a plug-and-play multi-character belief tracker. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), p. 13960â13980. Cited by: §2. Shao et al. (2024) Z. Shao, P. Wang, Q. Zhu, R. Xu, J. Song, X. Bi, H. Zhang, M. Zhang, Y. Li, Y. Wu, et al. Deepseekmath: pushing the limits of mathematical reasoning in open language models. arXiv preprint arXiv:2402.03300. Cited by: §4.2. Toubia et al. (2025) O. Toubia, G. Z. Gui, T. Peng, D. J. Merlau, A. Li, and H. Chen Database report: twin-2k-500: a data set for building digital twins of over 2,000 people based on their answers to over 500 questions. Marketing Science 44 (6), p. 1446â1455. Cited by: §5.1. Ullman (2023) T. Ullman Large language models fail on trivial alterations to theory-of-mind tasks. arXiv preprint arXiv:2302.08399. Cited by: §2. Wang et al. (2025) L. Wang, J. Zhang, H. Yang, Z. Chen, J. Tang, Z. Zhang, X. Chen, Y. Lin, H. Sun, R. Song, et al. User behavior simulation with large language model-based agents. ACM Transactions on Information Systems 43 (2), p. 1â37. Cited by: §1. Wang et al. (2026) S. Wang, S. Lin, Y. Zhao, X. Zhu, and Y. Zhang Meet dynamic individual preferences: resolving conflicting human value with paired fine-tuning. arXiv preprint arXiv:2604.12479. Cited by: §2. Wang et al. (2024a) S. Wang, H. Wu, X. Shi, T. Hu, H. Luo, L. Ma, J. Zhang, and J. Zhou Timemixer: decomposable multiscale mixing for time series forecasting. In International conference on learning representations, Vol. 2024, p. 38626â38652. Cited by: Table 1. Wang et al. (2024b) X. Wang, M. Feng, J. Qiu, J. Gu, and J. Zhao From news to forecast: integrating event analysis in llm-based time series forecasting with reflection. External Links: 2409.17515, Link Cited by: Table 1. Wang et al. (2022) X. Wang, J. Wei, D. Schuurmans, Q. Le, E. Chi, S. Narang, A. Chowdhery, and D. Zhou Self-consistency improves chain of thought reasoning in language models. arXiv preprint arXiv:2203.11171. Cited by: Table 2. Wilf et al. (2024) A. Wilf, S. Lee, P. P. Liang, and L. Morency Think twice: perspective-taking improves large language modelsâ theory-of-mind capabilities. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), p. 8292â8308. Cited by: §2. Wolfers and Zitzewitz (2004) J. Wolfers and E. Zitzewitz Prediction markets. Journal of economic perspectives 18 (2), p. 107â126. Cited by: §1, §2. Wu et al. (2026) S. Wu, E. Choi, A. Khatua, Z. Wang, J. He-Yueya, T. C. Weerasooriya, W. Wei, D. Yang, J. Leskovec, and J. Zou Humanlm: simulating users with state alignment beats response imitation. arXiv preprint arXiv:2603.03303. Cited by: §E.3.1, §1, §2, §5.1, Table 2. Yang et al. (2025a) A. Yang, A. Li, B. Yang, B. Zhang, B. Hui, B. Zheng, B. Yu, C. Gao, C. Huang, C. Lv, et al. Qwen3 technical report. arXiv preprint arXiv:2505.09388. Cited by: Table 1, Table 1. Yang et al. (2025b) Y. Yang, Y. Zhang, M. Wu, K. Zhang, Y. Zhang, H. Yu, Y. Hu, and B. Wang TwinMarket: a scalable behavioral and social simulation for financial markets. ArXiv abs/2502.01506. External Links: Link Cited by: §2. Yang et al. (2024) Z. Yang, Z. Zhang, Z. Zheng, Y. Jiang, Z. Gan, Z. Wang, Z. Ling, J. Chen, M. Ma, B. Dong, et al. Oasis: open agent social interaction simulations with one million agents. arXiv preprint arXiv:2411.11581. Cited by: §1, §2, §4.1. Yao et al. (2024) S. Yao, N. Shinn, P. Razavi, and K. Narasimhan tâaâutau-Bench: a benchmark for tool-agent-user interaction in real-world domains. arXiv preprint arXiv:2406.12045. Cited by: §1. Yu et al. (2025) H. Yu, Z. Qi, Y. Zhao, K. Nottingham, K. Xuan, B. P. Majumder, H. Zhu, P. P. Liang, and J. You Sotopia-rl: reward design for social intelligence. arXiv preprint arXiv:2508.03905. Cited by: §1. Yu et al. (2026) H. Yu, Y. Zhao, G. Lin, and J. You Building social world models with large language models. arXiv preprint arXiv:2606.11482. Cited by: §E.3.2, §5.1, Table 1. Zeng et al. (2023) A. Zeng, M. Chen, L. Zhang, and Q. Xu Are transformers effective for time series forecasting?. In Proceedings of the AAAI conference on artificial intelligence, Vol. 37, p. 11121â11128. Cited by: Table 1. Zhang et al. (2026a) L. Zhang, S. Milli, K. Jusko, J. Smith, B. Amos, W. Bouaziz, M. Revel, J. Kussman, Y. Sheynin, L. Titus, et al. Cultivating pluralism in algorithmic monoculture: the community alignment dataset. In International Conference on Learning Representations, Vol. 2026, p. 147076â147215. Cited by: §E.3.1, §5.1. Zhang et al. (2026b) P. Zhang, T. Lin, Y. Wu, J. Chen, Z. Wang, H. Yang, B. Zhao, F. Huang, Y. Li, and K. Zhang P-genrm: personalized generative reward model with test-time user-based scaling. In International Conference on Learning Representations, Vol. 2026, p. 70974â71002. Cited by: Table 2. Zhang et al. (2025) X. Zhang, J. Lin, X. Mou, S. Yang, X. Liu, L. Sun, H. Lyu, Y. Yang, W. Qi, Y. Chen, et al. Socioverse: a world model for social simulation powered by llm agents and a pool of 10 million real-world users. arXiv preprint arXiv:2504.10157. Cited by: §2. Zhang et al. (2026c) Z. Zhang, J. Yu, X. Song, B. Li, C. Li, C. Du, X. Fang, J. Liu, and R. Huang What to edit next: visually aligned image-editing follow-up suggestions in conversational systems. arXiv preprint arXiv:2608.07565. Cited by: §2. Zhou et al. (2025) X. Zhou, J. Liu, A. Yerukola, H. Kim, and M. Sap Social world models. arXiv preprint arXiv:2509.00559. Cited by: §2. Zhou et al. (2026) X. Zhou, W. Sun, W. Du, J. Liu, H. Sun, Q. Ma, T. Wu, Y. Yang, and M. Sap OdysSim: building foundation models for human behavior simulation. arXiv preprint arXiv:2606.14199. Cited by: §1, §2. Zhou et al. (2024) X. Zhou, H. Zhu, L. Mathur, R. Zhang, H. Yu, Z. Qi, L. Morency, Y. Bisk, D. Fried, G. Neubig, and M. Sap SOTOPIA: interactive evaluation for social intelligence in language agents. In International Conference on Learning Representations, B. Kim, Y. Yue, S. Chaudhuri, K. Fragkiadaki, M. Khan, and Y. Sun (Eds.), Vol. 2024, p. 40975â41019. External Links: Link Cited by: §1. Appendix A Limitation Our work has several limitations that suggest directions for future research. First, due to computational constraints, almost all experiments use Qwen3 as the shared backbone. Scaling macro2mind to larger models (e.g., 32Bâ70B) and validating its behavior across other model families (e.g., LLaMA-3) is an important next step. Second, due to the limitation of time and resources, we did not conduct large-scale human evaluation; we will conduct large-scale human evaluation in future research. Appendix B Reward Function of macro2mind Training For a generated output o on transition t, containing a forecast y^t+1 y_t+1 for the realized next price yt+1y_t+1, the training reward is a weighted sum of three terms: RâĄ(o)=ÎąfmtâRfmtâ(o)+ÎącorrâRcorrâ(o)+ÎądirâRdirâ(o).R(o)= _fmtR_fmt(o)+ _corrR_corr(o)+ _dirR_dir(o). (3) Format. Rfmtâ(o)=1R_fmt(o)=1 if the output contains every segment of the reasoning template and a parsable forecast y^t+1â[0,1] y_t+1â[0,1], and Rfmtâ(o)=0R_fmt(o)=0 otherwise. If no valid forecast can be parsed, we set Rcorrâ(o)=Rdirâ(o)=0R_corr(o)=R_dir(o)=0. Correctness. For a valid forecast, the correctness reward decays exponentially with the absolute forecast error: Rcorrâ(o)=expâĄ(â|y^t+1âyt+1|).R_corr(o)= \! (- | y_t+1-y_t+1 | ). (4) Since prices lie in [0,1][0,1], this term lies in [eâ1,1][e^-1,1] for valid forecasts. Direction. The direction reward indicates whether the predicted price change has the same sign as the realized change relative to the current price yty_t: Rdir(o)=[sign(y^t+1âyt)=sign(yt+1âyt)].R_dir(o)= 1\! [sign( y_t+1-y_t)=sign(y_t+1-y_t) ]. (5) Format-free reward. The curriculum (Section 4.2) evaluates forecast quality independently of output-format compliance. We therefore remove the format term and renormalize the remaining weights: R~â(o)âĄRâfmtâ(o)=ÎącorrâRcorrâ(o)+ÎądirâRdirâ(o)Îącorr+Îądirâ[0,1]. R(o)⥠R (o)= _corrR_corr(o)+ _dirR_dir(o) _corr+ _dirâ[0,1]. (6) Appendix C Evaluation Details C.1 SWM-Bench Let the test set contain N examples. For example i, let pip_i be the last observed price (at time t), yiy_i the realized next price (at t+1t+1), and y^i y_i the predicted price. Mean Absolute Error (MAE). The average magnitude of the prediction error: MAE=1Nââi=1N|y^iâyi|.MAE= 1N _i=1^N | y_i-y_i |. (7) Mean Absolute Scaled Error (MASE). MAE normalized by the error of the persistence (no-change) baseline y^i=pi y_i=p_i, which scales away cross-market differences in volatility; MASE<1MASE<1 means the model beats copying the last price: MASE=âi|Î^iâÎi|âi|Îi|.MASE= _i | _i- _i | _i | _i |. (8) Three-way Directional Accuracy (DA). We classify each change as up, down, or no-change via sgnδâ(x)â+1,â1,0sgn_δ(x)â\+1,-1,0\ with a small dead band δ=10â6δ=10^-6 (+1+1 if x>δx>δ, â1-1 if x<âδx<-δ, 00 otherwise), and measure agreement over the moved subset âł=i:|Îi|>δM=\i:| _i|>δ\: DA=1|âł|âiââł[sgnδ(Î^i)=sgnδ(Îi)].DA= 1|M| _i 1\! [sgn_δ( _i)=sgn_δ( _i) ]. (9) The three-way scheme penalizes predicting âno changeâ on a market that genuinely moved. Pearson Correlation (Corr). The Pearson correlation between predicted and realized belief changes over all N examples, capturing how well the model tracks the magnitude and sign of moves: Corr=âi(Î^iâÎ^ÂŻ)â(ÎiâÎÂŻ)âi(Î^iâÎ^ÂŻ)2ââi(ÎiâÎÂŻ)2,Corr= _i( _i- )( _i- ) _i( _i- )^2\, _i( _i- )^2, (10) where Î^ÂŻ and ÎÂŻ are the respective means. Scope. MAE, MASE, and Corr are computed over the evaluated subset (all or attributed); DA is computed over the moved subset âłM within that subset. MAE and MASE are lower-is-better; DA and Corr are higher-is-better. C.2 User Simulation Humanual-Politics. Given the userâs profile and an article, the simulator writes the comment the user would post. A judge Qwen3.8-Max compares it with the userâs real comment along seven aspects, following the rubric of Humanual-Politics: six aspects of the userâs latent state (stance, emotion, belief, value, goal, communication) and the response itself. For each aspect, the judge extracts one to three key points from the real comment, scores how well the generated comment reflects them (C), and subtracts a penalty for off-topic, contradictory, or hallucinated extra content (P); the aspect score is CâPC-P clipped to [0,1][0,1]. State alignment is the mean of the six state aspects and response alignment is the response aspect; they are never averaged together. Both are averaged over all test examples, with an example that has no parsable comment scoring zero. Our judge setup differs from the original in three ways: the judge model (qwen3.8-max instead of claude-4.5-haiku), all seven aspects scored in one request rather than one request per aspect, and each generation judged on its own rather than in a batch of rollouts for the same context. Absolute scores are therefore lower than the published ones and should be compared only within our table (Li et al., 2026a). We multiply all Humanual scores by 100 for reporting. OvertonBench. For a subjective question, a participant rated how well an LLM-generated answer represents their view on a scale from 1 (not at all) to 5 (fully represented); the simulator predicts that rating. We use the prompts of Overtonbench verbatim, with the fs+fr condition, their best-performing one: the prompt contains the participantâs own free-text answer to the question (fr) and their ratings of other LLM responses as examples (fs). We evaluate the 7,200 items of the political subset and report the mean absolute error (MAE) and Spearmanâs Ď between predicted and gold ratings, pooled over all items. An unparsable prediction is assigned the midpoint rating 3. PRISM. In each turn of a multi-turn conversation, a participant scored every candidate response from 1 to 100. Given the participantâs profile and the conversation history, the simulator predicts the score of each candidate. We evaluate all 27,172 scored turns in PRISM and report (i) MAE between predicted and gold scores on the original 1â100 scale, and (i) pairwise preference accuracy: over pairs of candidates in the same turn whose gold scores differ by more than 5 points, the fraction that the prediction orders the same way, with an exactly tied prediction counting as one half. An unparsable turn is assigned 50 for every candidate and loses every pair it takes part in. CAD. Given an annotatorâs profile and the conversation so far, the simulator predicts which of up to four candidate responses the annotator preferred in CAD. We evaluate a 2,000-comparison English subset and report accuracy; an unparsable prediction counts as wrong. Appendix D Details of User Simulation Baselines Prompting-based. These simulators keep the model parameters fixed: direct prompting of Qwen3-4B; PERSONA-Flow, which steers Qwen3-4B toward the userâs personality by composing activation vectors at inference time; History-ICL, which retrieves the userâs own past interactions with BM25 and places them in context, following LaMP; SynthesizeMe, which first induces a persona from the userâs history and then simulates the user conditioned on it; Self-report, which has the model first describe itself in the first person as the user and then answer; and Self-consistency, which samples five responses and aggregates them. Training-based. These are fine-tuned user simulators or personalized reward models. For HumanLM, we reproduce on Qwen3-4B the SFT, SFT-with-thinking, and standard GRPO recipes released with the paper, trained on Humanual (HumanLM-SFT, -SFT-think, -GRPO); P-GenRM is a generative personalized reward model; we reproduce its three-stage recipe on Qwen3-4B using the personalized split of PRISM: SFT on evaluation chains written by Qwen3.8-Max. Skywork-RM is the pretrained general-purpose reward model Skywork-Reward-V2-Llama-3.1-8B, used without any personalization training. MRM represents a userâs reward as a softmax-weighted mixture of two linear reward heads on frozen Skywork-Reward-V2-Llama-3.1-8B features; the heads and the initial mixture are meta-learned with the released code and hyper-parameters, and at test time the mixture weights are adapted to each user with one gradient step on that userâs preference pairs from their other conversations. WIMHF-RM is a LoRA reward model fine-tuned from Qwen3-4B. Applying scalar reward models. A scalar reward cannot produce a rating or a response by itself, so these reward models score CADâs options and take the highest, rank PRISM responses for pairwise accuracy, map the reward to the rating scale by isotonic regression for OvertonBench and PRISM MAE (for PRISM, fit with 5-fold cross-fitting grouped by user), and rerank eight Qwen3-4B samples on Humanual before the top one is judged. Scalar reward models do not transfer zero-shot, following their original per-dataset setting, MRM and WIMHF-RM are trained separately on each benchmark; these in-domain results are reported for reference rather than as a like-for-like comparison. Appendix E Artifact Details E.1 Model License Qwen3-4B License: Apache 2.0 Qwen3-8B License: Apache 2.0 Qwen3.5-397B License: Apache 2.0 Qwen3.8-Max License: Apache 2.0 E.2 Software Versions Our training infrastructure is built on top of verl22 2 https://github.com/volcengine/verl (v0.2.0.post2). E.3 Details of Datasets E.3.1 User Simulation Datasets Humanual (Wu et al., 2026). For free-text response generation, we use the Politics subset of Humanual. The model must generate a userâs free-text reply to a new context, given their profile and the preceding discussion thread. OvertonBench (Poole-Dayan et al., 2026). To test representation judgments, we employ OvertonBench, which tasks the model with predicting how a participant will rate a generated answer based on their stated views. PRISM (Kirk et al., 2024). For multi-turn preference prediction, the model predicts personalized ratings for candidate responses based on participant profiles and multi-turn dialogue history. Community Alignment (CAD) (Zhang et al., 2026a). Similarly, we use the English subset of CAD, which requires the model to predict an annotatorâs preferred response among four candidates, relying solely on their demographic profile and the preceding conversation. E.3.2 Market Prediction Dataset Details SWM-Bench (Yu et al., 2026). The dataset contains daily prediction-market trajectories sourced from Polymarket and Kalshi, paired with contemporaneous news retrieved from Google News. Each training example consists of a market proposition (a natural language question), its recent price history, and associated news events occurring at that step. This macroscopic observation is used to forecast the next market price. Appendix F Details of the Synthetic Data Experiment F.1 Dataset and Splits We use Twin-2K-500. Each participant provides a persona consisting of survey responses, together with answers to 53 held-out single-choice questions: ⢠40 pricing questions: âWould you buy <product> at $X?â for 40 branded products, with prices drawn from a grid of 0â2.0Ă each productâs reference price. ⢠13 behavioral-economics questions, one per experiment family: absolute_relative, allais, anchoring_african, anchoring_redwood, disease, less_is_more, myside, nonexp_QID196, nonexp_QID291, outcome_bias, proportion_dominance_1, proportion_dominance_2, and wta_wtp. The gold label is the participantâs actual answer. F.2 Student Input Format Training and evaluation use the same input format, rendered with the Qwen3 chat template with thinking disabled. Each input contains (i) the participantâs persona, given as 64 survey answers (32 on personality and values, 24 on economic and social attitudes, and 8 others); (i) several questions the participant has answered, as examples; (i) the target question and its options; and (iv) an instruction to output the option letter followed by â. â and a brief first-person answer. F.3 Evaluation Protocols Protocol Users In-context examples Out-of-user (main) 100 users 6 known answers of the user In-user 200 pilot training users same as in generation requests Table 3: Evaluation protocols for the synthetic data experiment. F.4 Data Generation Pipeline The pipeline is identical across generators. 1. Requests. We issue 6,000 requests: 10 for each of the 200 pilot users and 10 for each of 400 other training users. Each request pairs a participant (persona and example answers) with a target question family: about 40% pricing, with the rest spread over the 13 experiment families. Parameters are sampled from the familyâs real condition space. Pricing requests use invented everyday products, priced at a multiple (0.5â2.0Ă) of the categoryâs typical price, without revealing the typical price. 2. Scenario writing. The generator writes a scenario (question and options) and its own answer. Each scenario is automatically checked for format, compliance with the family parameters and price, and avoidance of held-out items: none of the 40 real pricing products may appear, and near-duplicates of any held-out question are rejected. 3. Rejection sampling. Failed requests are regenerated with a new seed (seed+7919â roundseed+7919¡round), for at most 3 rounds. 4. Population-anchored labeling. The generator answers each valid scenario k=6k=6 times as the participant. The prompt includes the answer distribution of training users under the corresponding real condition (for pricing, their purchase rate at the same price multiple). Anchors for pilot usersâ requests are computed over all 1,600 training users; anchors for the other requests exclude pilot users (1,400 users). The generatorâs original answer from the writing step is discarded; the label is the majority of the six answers, kept only if agreement is at least 0.6. Labels are generated text (option letter and first-person answer), not probabilities. 5. SFT export. The training target is â<letter>. <first-person answer>â with the student input format above; 5% of rows are held out for validation loss (seed 42). F.5 Student Training All students are trained from Qwen3-4B with identical settings: LoRA (rank 16, Îą=32Îą=32, dropout 0.05), learning rate 1Ă10â41Ă 10^-4, 1 epoch, global batch size 32 (micro batch 1), maximum length 16,384, and seed 42. We use AdamW (β=(0.9,0.999)β=(0.9,0.999), no weight decay) with a cosine schedule, 3% warmup, and gradient clipping at 1.0; the base model is kept in bf16 and LoRA parameters in fp32, with gradient checkpointing. The loss is the sum of response-token negative log-likelihoods over the global batch, divided by the number of response tokens. After training, the adapter is merged into Qwen3-4B and served with the same vLLM settings as the generators. Appendix G Detailed Prompts Conventions. Placeholders in are filled per example; italic text is our annotation and is not sent to the model; everything else is sent verbatim. Rows reported with thinking send the same prompt with thinking enabled; the other rows prefill an empty think block. G.0.1 macro2mind on SWM-Bench Behavioral reasoning prompt. macro2mind reads every market transition through this prompt, both in the training rollouts and for its SWM-Bench results in Table 1; the same reasoning rows are reused for individual simulation (Section G.0.2). macro2mind â behavioral reasoning prompt (training and SWM-Bench evaluation) System prompt You are a social-behavior simulator for prediction markets. Respond with exactly one JSON object and nothing else: no prose, no markdown fences, and no reasoning outside the JSON. User prompt You are given a forecasting task: question description window_data (one line per day; e.g. At date 2025-12-15, its target value is 0.0045) The associated news is: target_news (one line per news item, numbered [n1], [n2], âŚ; e.g. [n1] Published at: 2025-12-29T23:19:52Z; Source: The Dallas Morning News; Title: Jasmine Crockett touts progressive credentials in Dem Senate fight against James Talarico: Rep. Jasmine Crockett said Monday sheâs the more experienced, proven progressive option for Democrats as she competes with state Rep. James Talarico for theâŚ) Predict the market probability at date target_date by simulating the social groups that move this market. Write exactly 3 groups whose belief moves up after this news and exactly 3 groups whose belief moves down. Write every group as this analyst block: Group: the concrete group of people that behaves this way Goal: the goal, constraint or fear this group already had before this event Prior belief: what this group believed about the outcome before this news Interpretation: how this group reads the new evidence. Cite the news items you are reading by their id (n1, n2, ...), and say whether they are a surprise or a confirmation relative to that prior belief. Updated belief: what this group believes after reading it Response: the concrete behavior that follows from the change Each row of the block is one field of one JSON object: Group -> "group" Goal -> "goal" Prior belief -> "prior_belief" Interpretation -> "interpretation" Updated belief -> "updated_belief" Response -> "response" One filled-in block, from an unrelated merger market - copy the shape, never the content: Group: regulatory-sensitive institutional investors Goal: avoid unresolved regulatory risk Prior belief: moderately optimistic about deal completion Interpretation: the official antitrust objection in n3 is strong negative evidence, and it urprises a desk that expected a routine review Updated belief: substantially less confident that the deal will close Response: reduce exposure / become less willing to buy YES Return exactly one JSON object with this exact shape and nothing else: "up": [ "group": "...", "goal": "...", "prior_belief": "...", "interpretation": "...", "updated_belief": "...", "response": "...", "group": "...", "goal": "...", "prior_belief": "...", "interpretation": "...", "updated_belief": "...", "response": "...", "group": "...", "goal": "...", "prior_belief": "...", "interpretation": "...", "updated_belief": "...", "response": "..." ], "down": [ "group": "...", "goal": "...", "prior_belief": "...", "interpretation": "...", "updated_belief": "...", "response": "...", "group": "...", "goal": "...", "prior_belief": "...", "interpretation": "...", "updated_belief": "...", "response": "...", "group": "...", "goal": "...", "prior_belief": "...", "interpretation": "...", "updated_belief": "...", "response": "..." ], "interaction_analysis": "...", "aggregation": "...", "p_next": 0.0 Field rules: - "up" holds exactly 3 groups whose belief moves up after reading this news; "down" holds exactly 3 groups whose belief moves down. Which list a group is in is its direction, so every group must actually move, and the 6 groups must be different groups of people. - Every field is free text: write what its row of the analyst block asks for, in your own words. - "interaction_analysis": Plausible influence within or across the six groups, including transmission conditions and incremental behavioral changes; briefly state when none is justified. - "aggregation": Synthesize incremental implications, conditional interactions, unequal effect sizes, evidence quality, and current market information to justify the next price. - "p_next": your forecast (0 to 1) of the market probability at date target_date; the current market value is current_p. Assistant prefix (always prefilled; non-thinking): <think> </think> Hindsight-regret curriculum. During training, the current policy is shown the realized outcome and asked only to name the communities that drove it (hindsight prompt). The names are then placed into the behavioral reasoning prompt (privileged prompt), and the reward gain over the standard prompt is the hindsight regret that schedules training transitions. Both prompts are used only to score transitions; they never enter the training rollouts or any evaluation. macro2mind â hindsight prompt System prompt You are a social-behavior analyst for prediction markets. Respond with exactly one JSON object and nothing else: no prose, no markdown fences, and no reasoning outside the JSON. User prompt You are given a forecasting task: question description window_data (one line per day; e.g. At date 2025-12-15, its target value is 0.0045) The associated news is: target_news (one line per news item, numbered [n1], [n2], âŚ; e.g. [n1] Published at: 2025-12-29T23:19:52Z; Source: The Dallas Morning News; Title: Jasmine Crockett touts progressive credentials in Dem Senate fight against James Talarico: Rep. Jasmine Crockett said Monday sheâs the more experienced, proven progressive option for Democrats as she competes with state Rep. James Talarico for theâŚ) The realized market probability at date target_date was target_p (it was current_p before). Knowing that outcome, name the communities of market participants whose behavior explains it: 3 that pushed the probability up and 3 that pushed it down. Return exactly one JSON object with this exact shape and nothing else: "bullish": ["group name", "group name", "group name"], "bearish": ["group name", "group name", "group name"] Field rules: - "bullish" and "bearish" must each contain exactly 3 names of concrete groups of people (at most 12 words each); all names must be different. - Return names only: no reasons, no numbers, no other field. macro2mind â privileged prompt System prompt You are a social-behavior simulator for prediction markets. Respond with exactly one JSON object and nothing else: no prose, no markdown fences, and no reasoning outside the JSON. User prompt context_prompt (the user turn of the behavioral reasoning prompt up to, but excluding, its output-format instruction) Privileged context: an analyst who has already seen how this market resolved lists the communities that drove it. Communities that pushed the probability up: bullish Communities that pushed the probability down: bearish Use these communities when you decide which groups to simulate and how they move. schema_block (the output-format instruction of the behavioral reasoning prompt, from âPredict the market probability at date âŚâ to its end) G.0.2 Humanual Native prompt. The benchmark prompt used by Qwen3-4B, GRPO (with thinking), Self-consistency and P-GenRM in Table 2; the prompting methods below extend it. Humanual â native prompt System prompt You are simulating one specific person on a discussion site. You are given their persona and the article and comment thread they are about to reply to. Write the comment that person would post next, in their own voice, wording and length. Reply with the comment text and nothing else: no preamble, no quotation marks, no explanation, and never mention that you are a model. User prompt You are this person: persona Profile bio: bio Followers: followers_count; following following_count They are reading an article titled "title", subtitled "subtitle", published published_at. They are replying on reply_date. The article: article The thread so far: Commenter commenter: reply_content Write the comment this person posts next. Output only that comment. Rendered format of persona: The persona JSON, one line per key: "<Key Title>: <value>" (dict values as "sub: text; âŚ", lists as " - item"). Rendered format of reply_content: Each earlier comment of the thread, as "Commenter <commenter>: <content>", joined by blank lines; the "The thread so far:" section is omitted when there are none. A very long article is cut from the front and prefixed "[article truncated] âŚ". macro2mind. The user is posed as a single constituent of the behavioral reasoning prompt: the observed persona is the prior state and the thread is the context. macro2mind â Humanual prompt System prompt You are simulating one specific person on a discussion site. You are given their persona and the article and comment thread they are about to reply to. Respond with exactly one JSON object and nothing else: no prose, no markdown fences, and no reasoning outside the JSON. User prompt user_message Write this personâs reasoning as the analyst block below, then the comment they would actually post. Group: who this person is - the standing position or camp they speak from Goal: the goal, conviction or grievance they already held before reading this Prior belief: what this person thought about this topic before reading it Interpretation: how this person reads what they just read. Point at the specific part of the article or of the thread above they are reacting to, and say whether it surprises them or confirms what they already thought. Updated belief: what this person thinks after reading it Response: what they do about it Each row of the block is one field of one JSON object: Group -> "group" Goal -> "goal" Prior belief -> "prior_belief" Interpretation -> "interpretation" Updated belief -> "updated_belief" Response -> "response" Return exactly one JSON object with this exact shape and nothing else: "group": "...", "goal": "...", "prior_belief": "...", "interpretation": "...", "updated_belief": "...", "response": "...", "comment": "..." Field rules: - Every field is free text: write what its row of the analyst block asks for, in your own words. - "comment": the comment this person would post next, in their own voice, wording and length. This is the only field anyone else sees, so it must read as their comment and must not mention the reasoning above or that it was written by a model. Rendered format of user_message: the user turn of the native prompt without its last line (âWrite the comment this person posts next. âŚâ). Judge. Every generated comment is scored against the reference comment with this seven-aspect rubric (Qwen3.8-Max as judge); state alignment averages the first six aspects. Humanual â judge prompt User prompt You are a helpful and meticulous evaluator. Your task is to score how well the generated response matches the ground truth response with respect to each aspect below. Aspects (aspect_i: description_i): "stance": "HUMANâs agreement toward the explicitly named target, such as a claim or subject, in provided context. For example, \"strongly agrees with student loan forgiveness,\" or \"somewhat disagrees with a carbon tax\". In these cases, having only \"strongly agrees\" or \"somewhat disagrees\" is not enough, as they are missing targets. If there are multiple, include all of them separated by semicolons.", "emotion": "HUMANâs emotions with intensity toward an explicitly named target. For example, \"Moderate heartbreak for the wildfire victims; Mild irritation about governmentâs actions\". In this case, having only \"mild irritation,\" or \"moderate heartbreak\" are not sufficient, as the answer must express all three aspects: the emotion, the degree of emotion, and the target. If there are multiple, include all of them separated by semicolons.", "belief": "HUMANâs belief, namely a foundational assumption about how people, relationships, or the world fundamentally operate. Beliefs should reflect underlying mental models, not surface-level observations. Prefer beliefs that would explain multiple behaviors over beliefs that describe a single situation. Ask: \"What deeper assumption about human nature or the world would lead someone to say/do this?\" For example, \"people donât change unless theyâre forced to,\" \"loyalty is earned, not owed,\" \"conflict avoidance creates bigger problems later,\". Not beliefs: Practical advice, strategies, or statements about what should happen. Belief is not specific to a target or event, it should be a general statement about how HUMAN views the world.", "value": "HUMANâs value: what they think is important or should be prioritized. It is about \"what should matter\", not \"what is true\". For example, \"original ideas in a book are important\", \"characters should feel real\", anyone deserves basic respect\", and \"fairness matters more than efficiency\".", "goal": "HUMANâs goal: what they are trying to do with this comment. For example, \"persuade people that ...\", \"making fun of the poster on ...\", \"further seek help with ...\", \"offer support to ...\"", "communication": "HUMANâs communication approach: tone and how they structure their message. For examples, \"friendly, builds on a personal story then draws a lesson\", \"analytical, links claims with reasons and evidence step by step\", \"blunt, states conclusions with little explanation\"", "response": "HUMANâs response itself: what it says and does. Key points are about the response - the position it takes, what it offers, asks or objects to, and what it is reacting to - not the exact words it is made of. A different wording that makes the same points is a match." You will be given the context, the ground truth response, and ONE generated response. Provided Information: <|The Start of Context|> context <|The End of Context|> <|The Start of Ground Truth Response|> ground_truth <|The End of Ground Truth Response|> <|The Start of Generated Response|> generation <|The End of Generated Response|> Scoring Criteria (for each aspect_i): Assign a score in [0, 1] based on how much the aspects of generated response and the ground truth matches. For each aspect_i, follow this procedure: 1. Extract 1-3 key points: - Extract K key points from the ground truth response only about the aspect based on the description (e.g., if evaluating a "stance", pick key points related to the stance like "clearly disagrees with X"). - Each key point should be specific and distinct. 2. Score how well the generated response matches each key point: - For each key point i, compare it with the generated response and assign a match value m_i in range [0, 1]: - 1.0: The key point is precisely and perfectly reflected. - [0.7, 0.9]: Mostly reflected with small imperfections. - [0.4, 0.6]: Partially reflected or vague, but still leaning in the correct direction. - [0.1, 0.3]: Very weak reflection. - 0.0: Missed, contradicted, or reversed. 3. Compute coverage C = (m_1 + m_2 + ⌠+ m_K) / K, which measures how comprehensive the generated response reflects the ground truth response. 4. Compute penalty P for extra or conflicting content: - Examine additional content in the generated response beyond those key points: - Does it introduce unsupported evidence and assumptions? - Is it irrelevant to what ground truth response expresses? - Set a penalty P in [0, 1]: - 0.0: No problematic extra content; everything is perfectly matched. - [0.1, 0.3]: Slightly unnecessary or mildly speculative detail; meaning essentially unchanged. - [0.4, 0.6]: Moderate speculative or irrelevant content that somewhat shifts emphasis or adds unsupported ideas. - [0.7, 0.9]: Significant speculative, misleading, or conflicting content that clearly changes the meaning. - 1.0: Mostly off-topic, contradictory, or dominated by incorrect/hallucinated content. 5. If you are evaluating generated responses (skip if aspect_i is not response): - Length alone does NOT increase the score. Extra length is only ok if it is consistent and not redundant. - A generated response that is much longer than the ground truth response should be penalized via P. - The generated response may or may not reuse phrases from the context; however, if the generated response just directly copies previous context, without quoting them, treat that as off-task behavior and give a score of 0. 6. Compute the final score = max(0, min(1, C - P)) Additional considerations: - Follow the instruction carefully. - Be strict and reserve scores above 0.8 for clearly outstanding matches. Output format (JSON): "<aspect_1>": "thought": "<include (1) analysis of key points from ground truth along the aspect_1 dimension; (2) how well the generated response matches each key point; (3) compute the final score>", "score": <score>, "<aspect_2>": "thought": "<include (1) analysis of key points from ground truth along the aspect_2 dimension; (2) how well the generated response matches each key point; (3) compute the final score>", "score": <score>, "<aspect_3>": ... Format Notes: - Make sure to score the responses with respect to each aspect and consider only one aspect at a time. - All text in "thought" fields MUST be on a single line with no line breaks or newlines - Use standard JSON string format with double quotes. For any quotes needed inside strings, use single quotes (â) - Double check the JSON arrayâs format, especially for the comma and quotation marks - Ensure that ALL fields, especially "thought" and "score", are present for each item - You must provide exactly 1 score for each of the aspects: [âstanceâ, âemotionâ, âbeliefâ, âvalueâ, âgoalâ, âcommunicationâ, âresponseâ]. Your evaluation: Rendered format of context: **<role> <name>**: <content> (the article and every thread message of the dataset row, untruncated, blank-line separated) Rendered format of ground_truth: the gold comment Rendered format of generation: the simulated comment (for macro2mind, its "comment" field) G.0.3 User-simulation baselines We show each baseline on Humanual; the other benchmarks follow the same construction. PERSONA-Flow. Before each reply, the model predicts trait adjustments with the released prompt below (persona context: the Humanual persona; current input: the article and thread); the reply is then generated with the native prompt under the corresponding activation steering. PERSONA-Flow â trait-adjustment prompt User prompt You are a Personality Analyst AI. Your task is to determine the most appropriate **situational adjustments (deltas)** to a baseline AI personality based on the specific context and interaction needs. **Baseline Personality Profile:** The AI has these default traits: - **Agreeableness:** High (cooperative, trusting, helpful) - **Conscientiousness:** High (organized, reliable, disciplined) - **Extraversion:** Moderate (balanced social energy) - **Openness:** Moderate (balanced creativity and practicality) - **Neuroticism:** Low (generally calm and stable) **Context to Analyze:** - **Persona Context:** persona_context - **Current Input:** current_input **Your Task:** Determine which traits need adjustment (-2.0 to +2.0) based on what would be most effective for this specific interaction. Consider both directions equally and choose based on situational demands. **Trait Adjustment Guidelines:** **Extraversion:** - **Increase (+)** for: Group activities, public speaking, networking events, team leadership, social gatherings, collaborative brainstorming - **Decrease (-)** for: Individual focused work, quiet reflection, private study, solo creative tasks, introspective discussions **Agreeableness:** - **Increase (+)** for: Conflict resolution, team building, emotional support, consensus building, customer service - **Decrease (-)** for: Critical feedback, boundary setting, competitive situations, principled stands, resource constraints **Conscientiousness:** - **Increase (+)** for: Detailed planning, precision work, deadline management, quality control, formal procedures - **Decrease (-)** for: Spontaneous responses, creative brainstorming, flexible adaptation, crisis situations requiring speed **Neuroticism:** - **Increase (+)** for: Appropriate caution, emotional sensitivity, vulnerability, risk awareness, vigilant situations - **Decrease (-)** for: Calm leadership, confident decisions, reassuring others, crisis management, stability needs **Openness:** - **Increase (+)** for: Creative problem-solving, exploring new ideas, intellectual curiosity, innovation, novel situations - **Decrease (-)** for: Following procedures, traditional approaches, proven solutions, consistency needs, established protocols **Decision Principles:** 1. **Situational Fit:** Choose traits that best serve the interaction goals 2. **Context Sensitivity:** Consider what the human needs from this specific interaction 3. **Balanced Assessment:** Evaluate both positive and negative adjustments equally 4. **Natural Baseline:** Use 0.0 when baseline personality already fits the situation well **Output Format:** Provide only the numerical adjustment scores: Extraversion: [score] Agreeableness: [score] Conscientiousness: [score] Neuroticism: [score] Openness: [score] Qwen3-4B + History-ICL. The personâs own earlier comments are inserted into the native prompt. Qwen3-4B + History-ICL â Humanual User prompt [native Humanual user message up to, not including, the article header (or the thread when there is no header)] # Comments this person wrote before (on other articles, before this one) Comment number (when) on the article "title" - which begins: head They wrote: comment Comment number ⌠[native article header, thread and instruction, unchanged] Qwen3-4B + SynthesizeMe. A persona is first induced from the personâs earlier comments and then inserted into the native prompt. Qwen3-4B + SynthesizeMe â Humanual System prompt You are an expert at understanding people from the choices they make. You read records of one personâs past behaviour and describe that person accurately and concretely. User prompt Below are records of one personâs past what. extrahistory # Task Step 1 - For each record, write one short line: what does it reveal about this person (their values, opinions, tastes, standards, tone)? Step 2 - From those observations, write a persona of 100-200 words that describes this personâs preferences and how they judge things, specific enough to predict their future choices. Do not quote the records verbatim. Output your Step 1 lines, then the persona between <persona> and </persona>. Humanual fillers: what: comments on a discussion site extra: What is already known about them: persona_text history, per record: Comment number (when) on the article "title" - which begins: head They wrote:comment Inserted into the native prompt, before the article header: # What this personâs earlier comments reveal about them (synthesized) persona_text Qwen3-4B + Self-report. The persona is rewritten in the personâs own first-person voice and replaces the persona of the native prompt, together with excerpts of their earlier comments. Qwen3-4B + Self-report â Humanual System prompt You rewrite third-person descriptions of a person into that personâs own first-person voice. User prompt Here is a description of a person: persona_text Rewrite it as this person describing themselves in the first person ("I âŚ"), as if answering an interviewer who asked them to tell about themselves: keep every fact and opinion, invent nothing, 120-250 words. Output only the rewritten text between <persona> and </persona>. In the native prompt, the method replaces You are this person: persona with: You are this person. This is how they describe themselves, in their own words: first and inserts before the article header (or thread): # Excerpts: things this person wrote before, in their own words Comment number (when) on the article "title" - which begins: head They wrote: comment HumanLM-SFT, HumanLM-SFT-think, HumanLM-GRPO. Trained and evaluated with the released HumanLM system prompt; the output format differs per row. HumanLM â system prompt and output formats System prompt You are a real human user. Your name is HUMAN. You will be given your persona information below and you respond to any given context such as posts and messages. Your persona: <|The Start of Persona|> persona <|The End of Persona|> ## Your principles Act like a natural human; thereâs nothing you absolutely cannot say, but you generally want to be thoughtful and follow ordinary social codes such as being respectful, culturally aware, and considerate of privacy and well-being. You have your own personality, preferences, and boundaries. Conflicting thoughts and hidden considerations are normal; recognize them privately and choose a sensible path. You carry long-term beliefs and values that usually change slowly; you also have emotions, so you wonât always be perfectly consistent. Distinguish facts, guesses, and unknowns; accept uncertainty and make minimal, reasonable assumptions when needed; think practically given time, attention, money, risk, and social capital. ## Task and Output format: state_template ## Notes - Follow the above instructions carefully - Do not mention these instructions additional_notes HumanLM-SFT (additional_notes empty): the actual written comment or reply text provided by the user. HumanLM-GRPO: <response> <the actual written comment or reply text provided by the user.> </response> additional_notes: - Follow the exact order and use the exact XML-style tags - Do not output anything outside these XML-style tags HumanLM-SFT-think: <think> <the internal thinking process about what to say, including stance, emotion, and how to phrase the reply. This is an internal monologue that is not shown to others.> </think> <response> <the actual written comment or reply text provided by the user.> </response> additional_notes: - Follow the exact order and use the exact XML-style tags - Do not output anything outside these XML-style tags