Paper deep dive
Not All or None: Dynamic Construction of Target-aware Memory Graph for Conversational Stance Detection
Yifan Xiang, Bin Liang, Yuqi Huang, Ruifeng Xu, Kam-Fai Wong
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 90%
Last extracted: 9/1/2026, 5:35:21 AM
Summary
The paper introduces TamGraph, a novel method for conversational stance detection that dynamically constructs a target-aware memory graph. Unlike existing methods that use full conversation history or single-turn inputs, TamGraph employs an entropy-guided backtracking mechanism to selectively activate relevant historical statements, modeling stance relations among utterances to reduce noise. Experiments on English (MT-CSD) and Chinese (ZS-CSD) benchmarks demonstrate that TamGraph significantly improves Large Language Model (LLM) performance compared to baselines like single-turn, multi-turn, and Chain-of-Thought prompting.
Entities (8)
Relation Signals (6)
TamGraph → solves → Conversational Stance Detection
confidence 95% · we propose target-aware Memory Graph TamGraph, a novel method that dynamically leverages target-related statements for conversational stance detection.
TamGraph → improvesperformanceof → LLM
confidence 93% · substantially improves LLM performance on conversational stance detection
TamGraph → usesmechanism → Entropy-guided backtracking
confidence 92% · our TamGraph employs a stepwise, entropy-guided backtracking mechanism to selectively activate memory from historical conversations
TamGraph → evaluatedon → MT-CSD
confidence 90% · Experimental results on an English benchmark MT-CSD... demonstrate that our TamGraph substantially improves LLM performance
TamGraph → evaluatedon → ZS-CSD
confidence 90% · Experimental results on... a Chinese benchmark ZS-CSD demonstrate that our TamGraph substantially improves LLM performance
TamGraph → outperforms → GPT-4o-mini
confidence 80% · applying TamGraph on the models Qwen2.5-14B-Instruct... achieves performance comparable to the strong closed-source baseline GPT-4o-mini
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Stance detection is crucial for understanding the underlying attitude of an expression towards a target. Conversational stance detection is a more challenging stance detection task in real-world social media scenarios, as it involves detecting the user's stance by leveraging the target-related historical statements across conversational sessions. In this paper, we propose target-aware Memory Graph TamGraph, a novel method that dynamically leverages target-related statements for conversational stance detection. Instead of considering all preceding historical conversations or using no prior conversation information for stance detection, our TamGraph employs a stepwise, entropy-guided backtracking mechanism to selectively activate memory from historical conversations and dynamically constructs a target-aware graph to model the stance relations among utterances. This allows the exploitation of target-related information from the conversation history for stance detection while preventing the introduction of noise. Experimental results on both English and Chinese benchmarks demonstrate that our TamGraph substantially improves LLM performance on conversational stance detection.
Tags
Links
- Source: https://arxiv.org/abs/2608.29066v1
- Canonical: https://arxiv.org/abs/2608.29066v1
Trouble viewing inline? Open PDF directly →
Full Text
65,264 characters extracted from source content.
Expand or collapse full text
Not All or None: Dynamic Construction of Target-aware Memory Graph for Conversational Stance Detection Yifan Xiang 1,2 , Bin Liang 1,2 * , Yuqi Huang 3 , Ruifeng Xu 4 , Kam-Fai Wong 1,2 1 The Chinese University of Hong Kong, Hong Kong, China 2 MoE Key Laboratory of High Confidence Software Technologies, China 3 The Hong Kong Polytechnic University, Hong Kong, China 4 Harbin Institute of Technology, Shenzhen, China yfxiang, kfwong@se.cuhk.edu.hk, bin.liang@cuhk.edu.hk Abstract Stance detection is crucial for understanding the underlying attitude of an expression to- wards a target. Conversational stance detec- tion is a more challenging stance detection task in real-world social media scenarios, as it involves detecting the user’s stance by lever- aging the target-related historical statements across conversational sessions.In this pa- per, we propose target-aware Memory Graph TamGraph, a novel method that dynamically leverages target-related statements for conver- sational stance detection. Instead of consider- ing all preceding historical conversations or using no prior conversation information for stance detection, our TamGraph employs a stepwise, entropy-guided backtracking mech- anism to selectively activate memory from historical conversations and dynamically con- structs a target-aware graph to model the stance relations among utterances. This allows the exploitation of target-related information from the conversation history for stance detection while preventing the introduction of noise. Ex- perimental results on both English and Chinese benchmarks demonstrate that our TamGraph substantially improves LLM performance on conversational stance detection. 1 Introduction Stance detection aims to determine people’s opin- ionated standpoint or attitude (e.g., Favor, Against, or Neutral) in a post towards a specific target, entity, topic, or claim (Augenstein et al., 2016). While conversational stance detection requires determin- ing the stance of a conversation utterance as favor, against, or neutral toward a given target based on the conversation information (Niu et al., 2024). Existing research on LLM-based stance detec- tion (Zhang et al., 2022, 2023; Gatto et al., 2023) demonstrates the strong capability of LLMs across * Corresponding author. various prompting strategies. Some works design reasoning chains to elicit incremental and inter- pretable inference (Ma et al., 2024; Weinzierl and Harabagiu, 2024), while others demonstrate the feasibility of data augmentation with fine-tuning (Zhao et al., 2024; Ding et al., 2024; Wagner et al., 2024). Additionally, bias-oriented analyses exam- ine LLM behaviours across diverse settings, offer- ing mitigation directions (Li et al., 2025; Nguyen and Kim, 2025). However, these approaches typi- cally focus on the stance of a single utterance, thus failing to reflect real-world scenarios (e.g., social media threads), where utterances are embedded in conversations. Recent advances (Niu et al., 2024; Ding et al., 2025; Niu et al., 2025) introduce conversational stance detection benchmarks and propose promis- ing methods that exploit conversational information for stance detection. However, a key limitation of these approaches lies in their rough reliance on full conversation history. Such a strategy fails to ac- count for the fact that absorbing excessive sessions often includes substantial target-irrelevant or mis- leading content that not only introduces noise but also hinders stance detection. To address these limitations, we propose a novel method for conversational stance detection that dynamically constructs a Target-aware Memory Graph during the conversation, called TamGraph. Built upon LLMs, our TamGraph employs a step- wise and entropy-guided backtracking mechanism to selectively activate target-related information from conversation history as memories for the dy- namic construction of a target-aware graph, en- abling the modeling of stance relations across utter- ances. This process gradually incorporates mem- ories that can boost the confidence of stance pre- diction into the graph and stops when evidence suffices, essentially avoiding rough and indiscrim- inate injection of full conversations. To this end, this produces a target-focused, pruned context that arXiv:2608.29066v1 [cs.CL] 29 Aug 2026 retain? Exactly. Renewables are safer. We should just pour all that money into solar and wind instead of building new reactors. #Target: Building more nuclear power plants I’m not a big fan either, but the alternative is burning even more coal in fact, thus your measures may not be great ideas. I hate the idea of building more nuclear power plants. The waste lasts forever and one accident can ruin a whole region. I get that. The disasters we’ve already had are terrifying, and governments always seem to downplay the risks. Conversation contexts target utterance turn 1st turn 2nd turn n-1 turn n LLM single-turn input LLM LLM Not fan? Irrelevant? Alternative? What are the measures??? Hate? Disasters? Terrifying? How negative the dialogue is!!! Stance: Against Stance: None proposition 4 proposition 3 proposition 2 proposition 1 proposition 5 Update TamGraph Alternative causes more pollution. Your measures, not building new reactors, is not great. Stance: Favor Impactedbytheoverallstance, failed to learn the relations No enough information, failed to understand the meaning...... Dynamic TamGraph Construction entropy based control update? Figure 1: The comparison between our method and other prompting baselines. filters irrelevant noise and misleading information, enabling more reliable stance detection in complex, multi-turn interactions. More concretely, TamGraph performs target- aware backtracking over the conversation history. Starting from the last utterance, it traverses back- wards along the reply chain, sequentially retrieving only those prior turns whose propositions are target- related and activating them as memory. At each step, it induces the relations (e.g., support, against, questions) between the newly activated proposi- tions from the retrieved turn and the propositions from the last retained turn. Then, the dynamic target-aware memory graph is updated correspond- ingly, where propositions serve as nodes and the induced inter-turn relations as edges. The updated graph is serialized into a structured prompt and provided to the LLM for stance detection. Cru- cially, the entropy of the LLM’s intermediate infer- ence provides a principled signal for both selective memory updates and early stopping: (i) a newly retrieved turn is retained, and the memory graph is updated only when it increases the predictive confi- dence, and (i) the backtracking process terminates once the entropy reaches a predefined threshold or no further conversation turns are available. This procedure yields a progressively refined memory graph that preserves supportive evidence and dis- cards unhelpful history. Experimental results on an English benchmark MT-CSD and a Chinese bench- mark ZS-CSD demonstrate that our TamGraph substantially improves LLM performance on con- versational stance detection. Our main contributions are as follows: •We are the first to introduce a stepwise, back- tracking method for conversational stance detec- tion, motivated by the observation that incorpo- rating the entire conversation history can intro- duce target-irrelevant or misleading information and thus degrade the prediction accuracy. • WeproposeTamGraph(Target-aware Memory Graph), which selectively activates target-related memories from the preceding turns in the conversation and encodes them into a target-aware memory graph with their relations, thereby providing informative and target-centric contexts while mitigating noise from irrelevant or misleading turns. •Experimental results on English benchmark MT- CSD and Chinese benchmark ZS-CSD show that TamGraph consistently improves the per- formance over multiple prompting baselines and ablation variants across a wide range of LLM backbones, demonstrating the effective- ness of the stepwise, backtracking, and entropy- guided strategy.Moreover, applying Tam- Graph on the models Qwen2.5-14B-Instruct and Qwen3-4B-Instruct-2507 achieves performance comparable to the strong closed-source baseline GPT-4o-mini in both multi-turn and single-turn input settings. 2 Related Works 2.1 Stance Detection with LLMs Recently, a stream of works applies LLMs to stance detection. TR-ZSSD (Weinzierl and Harabagiu, 2024), and CoS (Ma et al., 2024) propose prompt- ing schemes to guide LLMs to perform incremen- tal reasoning, yielding interpretable and accurate predictions. Meanwhile, fine-tuning offers a prac- tical mechanism for endowing LLMs with task- specific knowledge; accordingly, data augmenta- tion with fine-tuning has demonstrated the feasi- bility in stance detection (Zhao et al., 2024; Ding et al., 2024; Wagner et al., 2024). However, these efforts primarily focus on single-utterance stance detection without multi-turn context, whereas we extend this research area to conversational settings. 2.2 Conversational Stance Detection: Methods and Datasets Beyond single-utterance settings, recent works have explored stance detection in multi-turn con- texts and benchmarks. GLAN (Niu et al., 2024) introduces a large-scale English conversational stance detection dataset, MT-CSD, and GLAN, a framework that models reply dependencies and local interactions via attention, CNN, and GCN components. ZS-CSD (Ding et al., 2025) extends this line of research to Chinese, releasing another dataset and proposing SITPCL, which combines a speaker-interaction graph over both intra- and inter- speaker relations with target-aware prototypical contrastive learning. Yet, these studies generally do not consider LLM-based frameworks, which are the primary focus of our work. 3 Methods Given a conversationD = (U,E,τ ), whereU = u 1 ,...,u n is the sequence of utterancesu i ,E are reply links between utterances, andτis the tar- get, the task is to predict the stance of the final utter- anceu n ∈ U. We propose Target-aware Memory Graph (TamGraph), which performs target-aware backward retrieval over the conversation history to generate memory contexts for LLM-based stance detection step by step. It incrementally activates τ-related propositions from earlier turns as mem- ory, and constructs a dynamic target-aware mem- ory graph whose edges encode the relations (e.g., support, against, questions) among these propo- sitions. Subsequently, the stance is detected by prompting an LLM with the progressively recon- structed memory graph. At each step, we adopt an entropy threshold to determine (i) whether the newly retrieved memory increases predictive confi- dence and should be retained, and (i) whether the current memory graph provides sufficient evidence for an early-stopping decision. 3.1 TamGraph Initialization We construct an LLM-based classifierf θ over a structured promptΠ(·)that produces a single-token labely ∈ Y = favor,against,none. At the stept = 0, we initialize the TamGraphG 0 = (V 0 ,E 0 )withV 0 = u n andE 0 =∅, and build the promptC 0 fromG 0 . By promptingf θ withΠ(C 0 ), we obtain the logits z 0 , the predicted label y 0 : z 0 ,y 0 = f θ (Π(C 0 )),(1) and the probabilitiesp 0 = softmax(z 0 ). For each step t, we define the entropy H(p t ) as: H(p t ) =− X y∈Y p t (y) logp t (y),(2) and select an entropy thresholdθ. IfH(p 0 ) ≤ θ, we output y 0 as the final decision. 3.2 TamGraph Updates and Inference Procedure Otherwise, ifH(p 0 ) > θ, we retainu 0 , set H(p r ) = H(p 0 )and regard the prediction as un- certain and iteratively expand the TamGraph. At stept ≥ 1, we have the last retained utteranceu r (e.g., at stept = 1,u r = u n ) and the next preced- ing utteranceu t onE. We activate memory inu t via proposition extraction to obtain target-related propositions S(u t ): S(u t ) =s t,1 , ..., s t,m , s t,i ⊆ u n .(3) and alsoS(u r ). Next, we perform relation induc- tion between S(u t ) and S(u r ), yielding: R t =(s i ,s j ,ρ)| s i ∈ S(u t ), s j ∈ S(u r ), (4) whereρ ∈ support, against, questions,...de- notes the stance ofs j towardss i . We then construct an updated TamGraph ̃ G t = ( ̃ V t , ̃ E t )by adding the node u t and edges R t : ̃ V t = V t−1 ∪ u t ,(5) ̃ E t = E t−1 ∪s i ρ ←− s j | (s i ,s j ,ρ)∈ R t . (6) We build the promptC t from ̃ G t , and recompute z t ,y t = f θ (Π(C t )),p t = softmax(z t ), andH(p t ). IfH(p t )≤ H(p r ), we retain this memory update, settingG t = ̃ G t ,u r = u t andH(p r ) = H(p t ); otherwise, we discardu t , keepG t = G t−1 , and proceed to the next preceding utteranceu t+1 onE. The recursion terminates onceH(p t ) ≤ θor no preceding utterances remain, and the final decision is the predictionyproduced at the last retained TamGraph update. The detailed procedure is presented in the Algo- rithm 1 and all prompts are detailed in the Table 17, Table 18, Table 19, and Table 20 in the section A in the Appendix. Algorithm 1: Procedure of TamGraph Con- struction Input: D = (U,E,τ ), entropy threshold θ Output: y ∈favor, against, none G←u n ; C = build_prompt(G); (z,y)← f θ (Π(C)); p← softmax(z); H ← Entropy(p); if H ≤ θ then return y; u r ← u 0 ; u t ← the replied-to utterance of u r in E; while u t ̸=∅ and H > θ do S(u t )← proposition_extraction(q,τ ); S(u r )← proposition_extraction(u r ,τ ); R← relation_induction(S(u t ),S(u r ),τ ); ̃ G← G∪ u t ∪ R; C = build_prompt( ̃ G); ( ̃z, ̃y)← f θ (Π(C)); ̃p← softmax( ̃z); ̃ H ← Entropy( ̃p); if ̃ H ≤ H then G← ̃ G; u r ← u t ; p← ̃p; H ← ̃ H ; y ← ̃y; u t ← the next preceding utterance on E; return y; 4 Experimental Setup Benchmarks. We conduct experiments on two conversational stance detection benchmarks. For English, we use MT-CSD (Niu et al., 2024), which contains 15876 instances with Bitcoin, Tesla, SpaceX, Biden, and Trump as targets. For Chinese, we use ZS-CSD (Ding et al., 2025), which consists of 17063 conversations with 113 noun-phrase-type targets and 167 claim-type targets. Their labels are either favor, against, or none. We conduct experi- ments on the test set of both benchmarks. Models.We evaluate five LLMs of various sizes: Qwen2.5-Instruct (3B, 7B, 14B) (Team, 2024), Qwen3-4B-Instruct-2507 (Yang et al., 2025), Llama-3.1-8B-Instruct (Grattafiori et al., 2024). All models are evaluated on MT-CSD. On ZS-CSD, we exclude Llama-3.1-8B-Instruct due to its limited native support for Chinese. Besides, we analyze GPT-4o-mini (Achiam et al., 2023) as a strong closed-source baseline. Baselines. We consider three input settings: (i) single-turn input, where LLMs only receive the target utterance, (i) multi-turn input, where LLMs also receive the entire conversation, and (i) Chain- of-Thought (CoT) prompting (Wei et al., 2022), where LLMs are given the entire conversation and the target utterance and instructed to perform step- by-step reasoning before outputting the stance label. Additionally, we include GPT-4o-mini (Achiam et al., 2023) as a strong closed-source baseline, GLAN and SITPCL as non-LLM baselines. For MT-CSD, we design prompts in English, whereas for ZS-CSD, we use Chinese translations consistent with the native language of the benchmark. The prompts are detailed in the Table 13, Table 14, and Table 15 in the section A in the Appendix. Implement Details. Since stance detection re- quires stable predictions, and to ensure repro- ducibility, we set the temperature to 0. Model per- formance is evaluated using accuracy and macro F1. In our method, we set the entropy threshold at 0.6 in the main experiments. We report results for each target in MT-CSD and the overall results on both benchmarks. 5 Results 5.1 Main Results We regard each combination of a target and a model as one group for MT-CSD, and each model as a group for ZS-CSD, which is not pre-partitioned by target. Within each group, we compare our method against the three baselines. Table 1 showcases the results. Our method achieves the highest overall weighted accuracy and macro F1 on all targets and both bench- marks. Specifically, averaged over all models, it yields massive improvements of 4.93%, 5.68%, and 2.75% on weighted accuracy over the single-turn input, multi-turn input, and CoT prompting base- lines, respectively, and gains of 6.68%, 5.56%, and 1.82% on macro F1. In detail, across 29 groups, our method ranks first in 19 groups and second in 5 groups. Moreover, as shown in Figure 2, our methods on Qwen2.5-14B and Qwen3-4B achieve performance comparable to, or even better than, GPT-4o-mini under single-turn and multi-turn in- puts, which indicates that our method unlocks strong capabilities in small-size LLMs. These ad- vances demonstrate the effectiveness of our method in LLM-based conversational stance detection in multi-language settings. In particular, our method consistently achieves a substantial lead on “Tesla” and “SpaceX” in MT- CSD, where multi-turn input often degrades per- formance compared with single-turn input. It also performs well in noise-heavy settings: for “Biden” in MT-CSD and ZS-CSD, where GPT-4o-mini rel- evance detection identifies 60.89% and 56.51% of utterances as target-irrelevant, respectively (prompt detailed in the Table 16 in the section A in the Ap- pendix), our method achieves the highest accuracy Model MT-CSD (acc. / F1 in %)ZS-CSD (acc. / F1 in %) BitcoinTeslaSpaceXBidenTrumpOverall Qwen2.5-3B (single)47.03/42.96 47.22/43.78 46.11/46.77 31.39/30.02 48.84/34.43 44.66/42.8833.55/25.53 Qwen2.5-3B (multi)48.69/43.80 37.41/35.26 41.74/40.98 34.31/34.6938.15/28.49 40.19/39.1331.11/20.77 Qwen2.5-3B (CoT)49.44/49.29 49.63 /44.1549.22/48.3242.58/43.29 46.70/37.4947.62/45.5935.41/33.91 Qwen2.5-3B (Ours)46.09/45.8862.22/54.6264.80/61.0045.50/33.1456.51/43.1654.66/50.7439.51/34.68 Qwen2.5-7B (single)44.98/42.32 70.91/58.2464.17/57.37 45.74/31.76 66.67/51.15 58.58/51.3239.43/28.53 Qwen2.5-7B (multi)50.00/49.2466.11/57.33 63.24/58.5045.74/36.6869.16/46.71 59.26/54.3039.63/31.45 Qwen2.5-7B (CoT)49.81/49.29 70.00/58.09 62.31/57.86 46.72/35.59 69.16/52.2960.14/55.6340.25/35.09 Qwen2.5-7B (Ours)43.68/41.0672.22/70.7566.98/66.7948.18/37.9270.05/70.1460.35/57.6840.91/36.89 Qwen2.5-14B (single)46.65/46.09 66.67/59.6863.86/61.2542.82/33.86 59.18/42.10 55.84/53.0146.32/43.96 Qwen2.5-14B (multi)54.09/51.1950.00/45.67 53.89/53.62 41.85/41.6049.38/39.10 49.89/49.4245.09/44.00 Qwen2.5-14B (CoT)52.23 /51.93 63.52/56.16 59.19/56.70 48.91/42.85 59.00/43.7756.77/55.3345.67/45.50 Qwen2.5-14B (Ours)47.03/46.6473.52/64.4671.03/67.0046.72/33.9267.74/51.8661.16/57.1046.94/46.59 Qwen3-4B-2507 (single) 43.87/42.97 69.63/59.5269.16/62.31 46.47/28.93 67.91/51.09 59.30/53.2739.20/34.72 Qwen3-4B-2507 (multi) 44.61/44.69 57.96/53.17 58.57/55.30 54.74/48.96 59.54/46.53 54.83/54.0339.67/36.60 Qwen3-4B-2507 (CoT)46.84/46.66 60.74/56.10 63.86/61.8748.66/42.7957.93/43.59 55.25/54.1543.61/38.89 Qwen3-4B-2507 (Ours)45.91/45.8070.56/60.7267.29/60.4545.99/30.9669.52/50.8160.02/54.9944.23/41.41 Llama-3.1-8B (single)36.99/34.21 39.81/37.71 44.86/41.71 44.04/33.57 40.64/36.15 40.78/40.31— Llama-3.1-8B (multi)49.81/49.51 58.70/52.2358.88/56.78 50.85/43.8041.35/37.13 51.24/50.31— Llama-3.1-8B (CoT)52.97/52.53 54.07/51.09 53.27/53.34 49.64 /47.35 52.41/40.6952.55/53.12— Llama-3.1-8B (Ours)51.49/50.4660.93/60.7655.14/55.8741.90/40.4159.18/55.6154.28/53.53— GLAN46.65/56.95 62.22/52.38 70.09/55.98 41.61/38.15 63.10/48.91 56.73/50.47— Table 1: Comparison of single-turn, multi-turn, CoT, and our method on MT-CSD and ZS-CSD. On ZS-CSD, we exclude Llama-3.1-8B-Instruct due to its limited native support for Chinese. For each model block and each column, the best (second-best) value for accuracy and macro F1 is shown in bold (underline). in 6 out of 9 groups and ranks second in 1 group. These observations illustrate that our method is effective when conversations contain substantial irrelevant or potentially harmful context. Build- ing on these insights, we demonstrate that the core of our method, selectively and dynamically acti- vating target-related memory and constructing a target-aware memory graph, is principled and ef- fective in distilling informative expressions from conversation history and brings reliable gains. We also compare our method with non-LLM baselines, including GLAN for MT-CSD and SIT- PCL for ZS-CSD. Our method outperforms GLAN in most target-specific settings and achieves bet- ter overall performance on MT-CSD. On ZS-CSD, our best result on Qwen2.5-14B reaches 46.59% in macro-F1, highly outperforming the 43.81% re- ported by SITPCL. These results further demon- strate the effectiveness of our method beyond LLM- based prompting baselines. 5.2Comparison with retrieval-based baselines We consider a retrieval-based baseline that selects evidence from the full conversation history using BM25. Specifically, we evaluate three query modes for retrieval: using only the target, using only the 0 20 40 60 80 BitcoinTeslaSpaceXBidenTrumpZS-CSD Accuracy (Ours vs. GPT-4o-mini) Ous (Qwen2.5 14B)Ous (Qwen3 4B) GPT-4o-mini (single)GPT-4o-mini (multi) 0 20 40 60 80 BitcoinTeslaSpaceXBidenTrumpZS-CSD Macro F1 (Ours vs. GPT-4o-mini) Ous (Qwen2.5 14B)Ous (Qwen3 4B) GPT-4o-mini (single)GPT-4o-mini (multi) Figure 2: The comparison between the performance of our method on Qwen2.5-14B-Instruct and Qwen3-4B- Instruct-2507 and GPT-4o-mini. The detailed results are presented in Table 12 in the Appendix. target utterance, and using the combination of tar- get and utterance as the query. For each query mode, we retrieve the top-2 evidence turns as the input context for stance detection as our baselines. Model MT-CSD (acc. / F1 in %) BitcoinTeslaSpaceXBidenTrumpOverall Qwen2.5-3B (BM25, q=target)45.35/40.26 44.26/38.63 47.98/49.16 36.01/32.37 49.55/35.58 44.83/41.51 Qwen2.5-3B (BM25, q=utterance)44.98/40.4249.81/42.9452.96/52.9139.90/34.00 51.34/37.0647.79/44.17 Qwen2.5-3B (BM25, q=u+t)43.87/39.08 48.15/40.27 51.40/51.80 37.71/33.70 49.38/35.56 46.10/42.65 Qwen2.5-3B (Ours)46.09/45.8862.22/54.6264.80/61.0045.50/33.1456.51/43.1654.66/50.74 Qwen2.5-7B (BM25, q=target)46.65/45.95 63.89/56.26 63.86/60.8549.15/41.28 66.67/47.93 58.08/54.36 Qwen2.5-7B (BM25, q=utterance)46.28/45.3369.81/60.3861.99/58.89 48.18/36.92 68.09/53.53 59.26/54.80 Qwen2.5-7B (BM25, q=u+t)46.28/45.32 68.15/59.96 62.93/59.86 47.45/37.01 67.74/53.7858.79/54.59 Qwen2.5-7B (Ours)43.68/41.0672.22/70.7566.98/66.7948.18/37.9270.05/70.1460.35/57.68 Qwen2.5-14B (BM25, q=target)50.37/50.15 59.44/54.31 61.99/60.46 45.74/40.1660.07/44.86 55.50/54.04 Qwen2.5-14B (BM25, q=utterance)49.44/49.00 65.00/57.52 64.17/62.61 46.96/33.63 61.68/51.2057.44/55.42 Qwen2.5-14B (BM25, q=u+t)50.00 /49.6665.00/58.0763.24/62.08 48.66/41.58 61.14/48.57 57.61/55.78 Qwen2.5-14B (Ours)47.03/46.6473.52/64.4671.03/67.0046.72/33.9267.74/51.8661.16/57.10 Qwen3-4B-2507 (BM25, q=target)47.40/47.15 52.59/47.93 60.44/58.2047.20/41.97 62.57/45.71 53.90/52.86 Qwen3-4B-2507 (BM25, q=utterance) 47.21/47.0359.07/52.6757.32/55.69 47.93/39.31 62.39/47.1055.00/53.31 Qwen3-4B-2507 (BM25, q=u+t)47.40/47.0356.30/50.50 57.63/56.15 48.66/41.6361.68/46.57 54.41/52.93 Qwen3-4B-2507 (Ours)45.91/45.8070.56/60.7267.29/60.4545.99/30.9669.52/50.8160.02/54.99 Llama-3.1-8B (BM25, q=target)48.14/46.2447.96/42.91 50.47/48.55 49.88/45.8355.26/39.73 50.40/50.14 Llama-3.1-8B (BM25, q=utterance)42.94/41.26 53.15/47.8251.09/49.8749.64/45.15 57.04/42.1550.86/50.40 Llama-3.1-8B (BM25, q=u+t)44.42/42.40 50.93/45.89 49.84/48.51 51.09/47.85 54.55/39.82 50.19/49.76 Llama-3.1-8B (Ours)51.49/50.4660.93/60.7655.14/55.8741.90/40.4159.18/55.6154.28/53.53 Table 2: Detailed results of BM25 retrieval baselines with different query modes on MT-CSD for five models, compared with our method. For each model block, we report three query variants and our method. Table 2 reports the results across the five models on MT-CSD. Overall, while BM25 provides a strong and simple global-retrieval baseline, our method achieves consistently better performance across dif- ferent models and retrieval settings. This suggests that the gains of our method are not solely due to selecting several topically related turns, but also depend on entropy-guided evidence selection and structured evidence integration. 5.3 Ablation Studies We conduct ablation studies on MT-CSD to analyze the effect of entropy threshold control. We vary the entropy threshold in0.0, 0.2, 0.4, 0.6, 0.8, where a higher threshold encourages earlier stopping and uses fewer conversation turns. The threshold0.0 corresponds to disabling entropy-based early stop- ping. We further evaluate two variants: (i) w/o entropy threshold, which disables entropy-based memory selection and early stopping, and instead incorporates all conversation turns and relations to construct the TamGraph; and (i) random turn selection, which randomly samples conversation turns and induces their relations for TamGraph con- struction. We report accuracy, macro F1, and to- ken usage during inference to evaluate both perfor- mance and computational cost. Figure 3 presents the effect of the various entropy thresholds on performance. Across all five models, our method consistently yields performance gains under a wide range of settings, and the accuracy and macro F1 remain within a relatively narrow band as the threshold varies, indicating robustness to thresholds. Meanwhile, the impact of different thresholds is not uniform across models, reflect- ing their varying sensitivity to the early-stopping mechanism. The detailed results for each target are presented in the section D in the Appendix. Figure 4 illustrates the effect of the entropy threshold on token usage. The results demonstrate the effectiveness of the early-stopping mechanism in reducing computational cost. Compared to the variant without entropy-based early-stopping (en- tropy threshold= 0), the average number of token usage during inference drops noticeably. As the threshold increases, token usage decreases even more while preserving competitive performance. Tables 6 and Table 7 report the results of the “w/o entropy threshold” and “random turn selec- tion”, respectively. Compared with our method, removing the entropy threshold results in an over- all 3.60% accuracy drop and 1.46% macro F1 drop, while random turn selection leads to an overall 3.09% accuracy drop and 1.32% macro F1 drop. Figure 5 further compares our method with the w/o entropy threshold variant across the five mod- 54.58 51.0 0 0.2 0.4 0.6 0.8 0 0.2 0.4 0.6 0.8 0 0.2 0.4 0.6 0.8 0 0.2 0.4 0.6 0.8 59.93 60.06 59.89 60.02 59.60 0 0.2 0.4 0.6 0.8 55.00 54.87 56.65 Qwen2.5-3B Qwen2.5-7B Qwen2.5-14B Qwen3-4B LLaMA3.1-8B 60.82 0 0.2 0.4 0.6 0.80 0.2 0.4 0.6 0.80 0.2 0.4 0.6 0.80 0.2 0.4 0.6 0.8 0 0.2 0.4 0.6 0.8 Accuracy (%) Macro F1 (%) 54.94 60.65 53.35 60.52 52.85 60.35 57.68 60.10 51.59 61.0 60.8 60.6 60.4 60.2 60.0 60.0 58.0 56.0 54.0 52.0 50.0 54.74 50.96 50.74 54.66 53.69 49.97 53.82 50.23 53.77 50.55 54.8 54.6 54.4 54.2 54.0 53.8 50.8 50.6 50.4 50.2 50.0 56.33 54.89 54.72 54.99 54.43 61.2 61.0 60.8 60.6 60.4 60.2 57.6 57.4 57.2 57.0 56.8 56.6 60.4 60.2 60.0 59.8 59.6 59.4 56.0 55.8 55.6 55.2 54.8 54.4 52.35 54.28 53.53 53.47 54.13 53.96 56.0 55.6 55.2 54.8 54.6 54.2 54.4 54.0 53.6 53.2 52.8 52.4 60.40 56.66 60.69 56.69 61.16 57.10 61.20 56.92 59.60 57.25 Figure 3: Ablation studies on the entropy threshold vs. performance across five models. The horizontal axis denotes the entropy threshold, and the vertical axis shows the overall weighted accuracy and macro F1. 700 Qwen2.5-3B Qwen2.5-7B Qwen2.5-14B Qwen3-4B LLaMA3.1-8B 0 0.2 0.4 0.6 0.8 TokenUsage 324.73 382.66 471.73 580.31 975.82 600 500 400 300 200 800 900 1000 1100 0 0.2 0.4 0.6 0.8 213.02 226.38 244.91 734.17 0 0.2 0.4 0.6 0.8 215.88 226.69 240.72 265.10 1171.69 0 0.2 0.4 0.6 0.8 1321.50 0 0.2 0.4 0.6 0.8 356.00 401.64 465.82 1117.17 272.67 161.38 169.65 176.00 187.98 544.87 Figure 4: Ablation studies on the entropy threshold vs. token usage across five models. The horizontal axis denotes the entropy threshold, and the vertical axis shows the average number of token usage per sample. els. These degradations highlight the effective- ness of our entropy-based TamGraph construction and early-stopping strategy in selecting informative context. In particular, rather than incorporating all conversation turns, TamGraph supports stepwise and selective memory update that can filter noise and terminate once sufficient evidence has been accumulated, yielding more reliable results than using the full conversation history. 5.4 Discussion 5.4.1 Performance Gap Analysis We observe clear performance variations across targets and models. On MT-CSD, our method achieves particularly strong gains on “Tesla” and “Trump”, where it obtains the best accuracy across all models, while the improvement on “Bitcoin” is relatively limited. From the model perspective, our method brings substantial improvements on Qwen2.5-3B and Qwen2.5-7B across multiple tar- gets and benchmarks, whereas the gains on Llama- 3.1-8B are more moderate. Building on LLMs, our method unavoidably depends on the models’ un- derlying capability. Nevertheless, the strong gains on relatively small models and the consistently pos- itive margins across most settings jointly demon- strate that our method effectively enhances weaker backbones while maintaining robust performance. 5.4.2Entropy–Accuracy Correlation Analysis Although LLM-generated intermediate outputs may contain hallucinations or extraction errors, our method achieves consistent improvements over baselines, suggesting that these outputs remain ef- fective in aggregate for organizing target-related evidence. To further validate entropy-guided back- tracking, we analyze results on MT-CSD across five models, totaling 11,855 test samples. We find that cases where an initially correct prediction be- comes incorrect after backtracking are rare, occur- ring in only 317 cases (2.67%). In addition, cases where entropy decreases but the final prediction re- mains incorrect account for only 908 cases (7.66%). These results demonstrate that entropy serves as an 0 10 20 30 40 50 60 70 Qwen2.5-3BQwen2.5-7BQwen2.5-14BQwe3-4BLlama3.1-8B Accuracy (Ours vs. w/o entropy threshold) Ousw/o entropy threshold 0 10 20 30 40 50 60 70 Qwen2.5-3BQwen2.5-7BQwen2.5-14BQwe3-4BLlama3.1-8B Macro F1 (Ours vs. w/o entropy threshold) Ousw/o entropy threshold Figure 5: The comparison between the performance of our method and the “w/o entropy threshold”. Turn 1:US has over 750 complaints that Teslas brake for no reason. Turn 2:... 750 people have filed a formal complaint to NHTSA... I've experienced it many times in my 2018 Model 3...I do not use Autopilot in traffic anymore... Turn 3:Everyone ... has seen this problem, and some sections of road the issue is easily reproduceable between cars as well. We also stopped using autopilot in traffic. It's totally not worth it... Turn 4:... I no longer drive behind Teslas because I swear... I've seen it happen multiple times. Turn 5:... I think the logic panics and thinks they're semi trailers very suddenly on top of the car. Happens less than it did a couple years ago (2019 model 3) ... Turn 6 (Target utterance):Glad to know that public roads are a Tesla testing facility that everyone gets to participate in. Target: Tesla Gold stance: against COT baseline Ours Turn 6 Turn 5: refers_to the previous negative experience, suggesting that tesla vehicles are being tested on public roads... Turn 5 Turn 4: elaborates on the previous observation, specifying that ... occurs most reliably while driving under bridges and large overhead signs, ... noting that it has happened less frequently in recent times. pred: against let's analyze the utterance step by step: 1. the final utterance is: "glad to know that public roads are a tesla testing ..." 2. this statement expresses a positivesentiment towards tesla by suggesting that the public roads serve as a testing ground for tesla vehicles. 3. the speaker is expressing satisfaction or approval at the idea of public roads being used for testing purposes, which implies supportfor tesla's practices. based on the assessment criteria provided: -the utterance supports or agrees with the target entity (tesla). therefore, the stance is favor. pred: favor Figure 6: Qualitative example of our method and CoT baseline. effective and reliable signal for selecting and orga- nizing target-related evidence and the robustness of LLM-generated intermediate outputs. 5.5 Case Study As illustrated in Figure 6, the qualitative example on Qwen2.5-7B shows how TamGraph supports conversational stance detection. Our method iden- tifies target-related propositions, induces their rela- tions, and constructs a target-aware memory graph, enabling the model to capture the negative and sar- castic meaning of the target utterance. In contrast, the CoT baseline misinterprets the utterance as posi- tive and supportive toward “public roads being used for testing purposes,” leading to an incorrect pre- diction. This suggests that baselines have limited capability to exploit conversation structure, while TamGraph better organizes conversational evidence and helps the model handle implicit stance cues such as sarcasm. Additional qualitative examples are provided in the section E in the Appendix. 6 Conclusion In this paper, we first conduct preliminary exper- iments across multiple conversational stance de- tection settings on various LLMs and reveal that indiscriminately incorporating the entire conversa- tion history impedes the performance of LLMs due to the introduction of noise and target-irrelevant information. To address this limitation, we pro- pose TamGraph (Target-aware Memory Graph), a novel stepwise, backtracking method that selec- tively activates target-related propositions as mem- ory from preceding turns in the conversation and dynamically constructs and updates a target-aware memory graph. Utilizing the entropy-guided selec- tive memory update and early-stopping mechanism, TamGraph retains only confidence-improving evi- dence and filters misleading turns, providing infor- mative and target-centric contexts for LLMs and mitigating noise from irrelevant conversation his- tory. Experimental results on an English bench- mark, MT-CSD, and a Chinese benchmark, ZS- CSD, illustrate that TamGraph substantially im- proves LLM performance on conversational stance detection in multi-language settings and also out- performs benchmark-specific non-LLM baselines. Further ablation studies demonstrate the effective- ness of our entropy-guided selective memory up- date and early-stopping mechanism. Limitations There are several limitations in this work. First, TamGraph is built upon LLMs; therefore, it un- avoidably depends on the underlying capabilities of the backbone models for memory activation and entropy-based control. Errors in these intermedi- ate steps may propagate to the memory graph and affect the final result. Second, the stepwise pro- cedure can result in substantial overhead for long conversations, as it may require multiple iterations, leading to increased latency and token consump- tion. Third, due to computational constraints, we evaluate TamGraph on models of comparatively small sizes. Although the method consistently im- proves performance and even reaches performance comparable to GPT-4o-mini on several models, fur- ther validation on larger and more diverse frontier models can be conducted to characterize its scaling behaviours. Despite these limitations, we hope our research offers useful insights for future investiga- tion on conversational stance detection. Ethical considerations This research focuses on conversational stance detection using publicly available datasets. All datasets used in this paper are released under per- missive licenses, are used consistently with their intended purposes, and do not contain personally identifiable information. However, due to the na- ture of the task, the conversations may include con- troversial or sensitive opinions about real-world en- tities (e.g., public figures, organizations, or events). No human subjects were involved in data collec- tion in this paper. While our method is designed to detect the stance expressed in a given utterance, it may produce incorrect predictions, and any model outputs or reported results reflect the content of the datasets rather than the views of the authors. Acknowledgements This work is partially supported by Hong Kong RGC GRF No. 14206324, CUHK direct grant No. 4055291, National Natural Science Foundation of China 62576120, and CIPS-SMP-Zhipu Large Model Fund. References Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, and 1 others. 2023. Gpt-4 techni- cal report. arXiv preprint arXiv:2303.08774. Isabelle Augenstein, Tim Rocktäschel, Andreas Vla- chos, and Kalina Bontcheva. 2016. Stance detection with bidirectional conditional encoding. In Proceed- ings of the 2016 Conference on Empirical Methods in Natural Language Processing, pages 876–885. Daijun Ding, Rong Chen, Liwen Jing, Bowen Zhang, Xu Huang, Li Dong, Xiaowen Zhao, and Ge Song. 2024. Cross-target stance detection by exploiting target analytical perspectives.In ICASSP 2024- 2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 10651–10655. IEEE. Yuzhe Ding, Kang He, Bobo Li, Li Zheng, Haijun He, Fei Li, Chong Teng, and Donghong Ji. 2025. Zero- shot conversational stance detection: Dataset and approaches. In Findings of the Association for Com- putational Linguistics: ACL 2025, pages 3221–3235, Vienna, Austria. Association for Computational Lin- guistics. Joseph Gatto, Omar Sharif, and Sarah M. Preum. 2023. Chain-of-thought embeddings for stance detection on social media. In Findings of the Association for Com- putational Linguistics: EMNLP 2023, pages 4154– 4161, Singapore. Association for Computational Lin- guistics. Aaron Grattafiori, Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al- Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Alex Vaughan, and 1 others. 2024. The llama 3 herd of models. arXiv preprint arXiv:2407.21783. Ang Li, Jingqian Zhao, Bin Liang, Lin Gui, Hui Wang, Xi Zeng, Xingwei Liang, Kam-Fai Wong, and Ruifeng Xu. 2025. Mitigating biases of large lan- guage models in stance detection with counterfactual augmented calibration. In Proceedings of the 2025 Conference of the Nations of the Americas Chap- ter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Pa- pers), pages 7075–7092, Albuquerque, New Mexico. Association for Computational Linguistics. Junxia Ma, Changjiang Wang, Hanwen Xing, Dong- ming Zhao, and Yazhou Zhang. 2024. Chain of stance: Stance detection with large language mod- els. In CCF International Conference on Natural Language Processing and Chinese Computing, pages 82–94. Springer. Quang Minh Nguyen and Taegyoon Kim. 2025. Is ex- ternal information useful for stance detection with LLMs? In Findings of the Association for Computa- tional Linguistics: ACL 2025, pages 14798–14807, Vienna, Austria. Association for Computational Lin- guistics. Fuqiang Niu, Min Yang, Ang Li, Baoquan Zhang, Xi- aojiang Peng, and Bowen Zhang. 2024. A chal- lenge dataset and effective models for conversational stance detection. In Proceedings of the 2024 Joint International Conference on Computational Linguis- tics, Language Resources and Evaluation (LREC- COLING 2024), pages 122–132, Torino, Italia. ELRA and ICCL. Fuqiang Niu, Yi Yang, Xianghua Fu, Genan Dai, and Bowen Zhang. 2025. C-mtcsd: A chinese multi- turn conversational stance detection dataset. In Com- panion Proceedings of the ACM on Web Conference 2025, pages 769–772. Qwen Team. 2024. Qwen2.5: A party of foundation models. Stefan Sylvius Wagner, Maike Behrendt, Marc Ziegele, and Stefan Harmeling. 2024. Sqbc: Active learning using llm-generated synthetic data for stance detec- tion in online political discussions. arXiv preprint arXiv:2404.08078. Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, and 1 others. 2022. Chain-of-thought prompting elic- its reasoning in large language models. Advances in neural information processing systems, 35:24824– 24837. Maxwell Weinzierl and Sanda Harabagiu. 2024. Tree- of-counterfactual prompting for zero-shot stance de- tection. In Proceedings of the 62nd Annual Meet- ing of the Association for Computational Linguistics (Volume 1: Long Papers), pages 861–880, Bangkok, Thailand. Association for Computational Linguistics. An Yang, Anfeng Li, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Gao, Chengen Huang, Chenxu Lv, and 1 others. 2025.Qwen3 technical report. arXiv preprint arXiv:2505.09388. Bowen Zhang, Daijun Ding, Liwen Jing, Genan Dai, and Nan Yin. 2022. How would stance detection techniques evolve after the launch of chatgpt? arXiv preprint arXiv:2212.14548. Bowen Zhang, Xianghua Fu, Daijun Ding, Hu Huang, Genan Dai, Nan Yin, Yangyang Li, and Liwen Jing. 2023. Investigating chain-of-thought with chatgpt for stance detection on social media. arXiv preprint arXiv:2304.03087. Chenye Zhao, Yingjie Li, Cornelia Caragea, and Yue Zhang. 2024. ZeroStance: Leveraging ChatGPT for open-domain stance detection via dataset generation. In Findings of the Association for Computational Lin- guistics: ACL 2024, pages 13390–13405, Bangkok, Thailand. Association for Computational Linguistics. A Prompts In this section, we present the prompts used in our experiments. Table 13, Table 14, and Table 15 show the prompt templates for the single-turn input, the multi-turn input, and the CoT baseline, respectively. DatasetLabelProportion Precision RecallF1 MT-CSD favor19.65%52.3733.69 41.00 MT-CSD against30.54%53.0460.58 56.56 MT-CSD none49.81%62.8466.20 64.47 ZS-CSDfavor34.87%56.6011.54 19.17 ZS-CSDagainst29.06%35.6843.81 39.33 ZS-CSDnone36.07%45.7072.48 56.05 Table 3: Pooled per-label performance of TamGraph across the evaluated model backbones. Table 17 and Table 18 present the prompts used for proposition extraction and relation induction, respectively, as described in the section 3. Table 19 shows the prompt that converts the memory graph into natural language as the input for LLMs, and Table 20 provides the evaluation prompt used in our TamGraph. B Per-label Analysis To provide a more fine-grained evaluation under label imbalance, we further report per-label preci- sion, recall, and F1, together with the proportion of instances associated with each gold stance label, in Table 3. For MT-CSD, the results cover five backbone models across five targets; for ZS-CSD, they cover the four Chinese-capable backbones re- ported in Table 1. On MT-CSD, the against and none labels achieve comparatively strong F1 scores, whereas the lower recall of favor indicates that favor-oriented instances remain more difficult to identify. On ZS-CSD, none achieves the highest F1, while the low recall of favor reveals a similar label-specific challenge. C Target-irrelevant History Analysis. To more systematically quantify the amount of noisy conversational context, we condcut the GPT- 4o-mini relevance analysis described in Table 16 to all targets in MT-CSD and the complete ZS-CSD benchmark. We report both the proportion of target- irrelevant utterances among all utterances and the proportion among history turns after excluding the final utterance to be evaluated. As shown in Table 5, a substantial proportion of conversational history is unrelated to the tar- get, particularly for the Biden subset and ZS-CSD. We further compare full multi-turn and single- turn prompting over the 29 model-target or model- dataset settings reported in Table 1. Full multi- turn input performs worse in 14 of 29 accuracy settings and 12 of 29 macro-F1 settings. These results demonstrate that additional conversational context is not uniformly beneficial and provide di- rect empirical motivation for selectively activating target-related history. D Additional Ablation study In this section, we present the detailed results of our ablation study. Table 8, Table 9, Table 10, and Table 11 showcase the accuracy and macro F1 for entropy thresholds of 0.0, 0.2, 0.4, and 0.8 respectively. The case of 0.6 is already presented in the Table 1 in the section 5. E Additional Case Study In this section, we provide additional qualitative ex- amples comparing our method with the CoT base- line. Figure 7 illustrates a case where the CoT base- line produces garbled output and thus fails to yield a valid prediction. Figure 8 presents another exam- ple in which the CoT baseline incorrectly predicts a favor stance, due to being impacted by the over- all conversational stance. Besides, the TamGraph provided in the cases also illustrates the entropy- guided memory (conversation turn) selection. We additionally report a failure case where, for a user asking whether Tesla offers a simpler and less error-prone adaptive cruise-control option, all baselines and TamGraph predict against instead of none. TamGraph correctly connects the question to the preceding complaint about Autopilot, but the remaining error reflects the difficulty of distinguish- ing an information-seeking question with negative presuppositions from an explicit negative stance. F Detailed Results of GPT-4o-mini In this section, we report the detailed results of the GPT-4o-mini on the MT-CSD and the ZS-CSD, corresponding to the Figure 2 in the section 5.1. Table 12 showcases the accuracy and the macro F1 for each target and benchmark. G Inference Efficiency Analysis Although backtracking may introduce additional inference calls, its cost is effectively controlled by the entropy-based early-stop strategy. The back- tracking depth is highly concentrated: 83.35% of test cases terminate after only one additional back- tracking step, and 97.28% terminate within two ad- ditional backtracking steps. Cases requiring three or more additional backtracking steps account for Dataset# ModelsCoTTamGraph Speedup MT-CSD5 ∼33.0 ∼27.01.22× ZS-CSD4 ∼29.0 ∼24.01.21× Overall– ∼31.0 ∼25.51.22× Table 4: Dataset-level end-to-end latency in seconds per example. The overall row averages the two dataset-level results. only 2.72%. This distribution indicates that back- tracking typically stops early once sufficient evi- dence has been collected, enabling our method to integrate useful contextual evidence with limited additional inference overhead. Consistent with the step-level analysis, the token- level results in Figure 4, which show that our method substantially reduces token usage com- pared with the baselines, these findings further demonstrate the cost efficiency of our framework. We further measure end-to-end runtime latency against the CoT baseline under the same model backbones, datasets, and hardware. The mea- surement includes all intermediate and final LLM calls made by TamGraph, including proposition extraction, relation induction, memory construc- tion, and final stance prediction. As shown in Ta- ble 4, TamGraph is faster than CoT on both bench- marks and achieves an overall speedup of approx- imately 1.22×. Together with the token-usage re- sults in Figure 4 and the concentrated backtracking- depth distribution, this result shows that the sav- ings produced by entropy-guided early stopping contribute to lower practical inference cost despite TamGraph’s intermediate reasoning stages. Dataset / Target# InstancesAvg. TurnsIrrelevant UtterancesIrrelevant History Turns MT-CSD / Biden4114.7960.75%54.86% MT-CSD / Bitcoin5384.6527.64%25.69% MT-CSD / SpaceX3214.5217.03%15.15% MT-CSD / Tesla5404.8622.18%18.80% MT-CSD / Trump5614.9124.69%22.44% MT-CSD Overall2,3714.7730.04%27.02% ZS-CSD Overall2,5843.7056.51%54.64% Table 5: Dataset-level analysis of target-irrelevant conversational content. History turns exclude the final utterance whose stance is evaluated. Model MT-CSD (acc. / F1 in %) BitcoinTeslaSpaceXBidenTrumpOverall Qwen2.5-3B (w/o entropy threshold)45.35/45.48 52.59/48.09 58.26/56.14 45.26/39.97 49.55/39.29 49.73/47.97 Qwen2.5-3B (Ours)46.09/45.8862.22/54.6264.80/61.0045.50/33.1456.51/43.1654.66/50.74 Qwen2.5-7B (w/o entropy threshold)48.33/48.52 60.93/55.26 63.86/60.13 47.93/40.45 64.35/50.94 57.02/55.23 Qwen2.5-7B (Ours)43.68/41.0672.22/70.7566.98/66.7948.18/37.9270.05/70.1460.35/57.68 Qwen2.5-14B (w/o entropy threshold)58.18/57.16 58.33/54.69 61.37/59.53 44.28/43.58 55.26/42.24 55.55/55.32 Qwen2.5-14B (Ours)47.03/46.6473.52/64.4671.03/67.0046.72/33.9267.74/51.8661.16/57.10 Qwen3-4B-2507 (w/o entropy threshold) 45.54/45.45 55.00/50.94 57.01/54.21 47.93/42.30 60.78/48.60 53.27/52.57 Qwen3-4B-2507 (Ours)45.91/45.8070.56/60.7267.29/60.4545.99/30.9669.52/50.8160.02/54.99 Llama-3.1-8B (w/o entropy threshold)50.37/49.18 61.48/54.16 58.26/56.82 52.80/46.83 61.68/46.83 57.06/56.09 Llama-3.1-8B (Ours)51.49/50.4660.93/60.7655.14/55.8741.90/40.4159.18/55.6154.28/53.53 Table 6: Detailed results of ablation study on the “w/o entropy threshold” on MT-CSD for five models and the comparison with our method. For each group, the best value for accuracy and macro F1 is shown in bold. Model MT-CSD (acc. / F1 in %) BitcoinTeslaSpaceXBidenTrumpOverall Qwen2.5-3B (random turn selection)46.84/46.82 55.56/50.64 57.01/54.90 43.55/36.29 49.73/39.10 50.32/48.15 Qwen2.5-3B (Ours)46.09/45.8862.22/54.6264.80/61.0045.50/33.1456.51/43.1654.66/50.74 Qwen2.5-7B (random turn selection)45.17/45.41 65.00/58.54 64.80/60.70 47.45/38.44 63.46/49.35 57.06/54.53 Qwen2.5-7B (Ours)43.68/41.0672.22/70.7566.98/66.7948.18/37.9270.05/70.1460.35/57.68 Qwen2.5-14B (random turn selection)52.79/52.26 60.00/56.62 60.44/58.33 46.23/42.39 57.58/44.08 55.46/55.02 Qwen2.5-14B (Ours)47.03/46.6473.52/64.4671.03/67.0046.72/33.9267.74/51.8661.16/57.10 Qwen3-4B-2507 (random turn selection) 47.21/47.30 59.81/56.04 57.94/54.73 48.42/40.54 62.92/49.89 55.46/54.14 Qwen3-4B-2507 (Ours)45.91/45.8070.56/60.7267.29/60.4545.99/30.9669.52/50.8160.02/54.99 Llama-3.1-8B (random turn selection)50.19/49.30 63.70/57.96 55.76/53.88 50.12/43.74 61.68/46.51 56.73/55.60 Llama-3.1-8B (Ours)51.49/50.4660.93/60.7655.14/55.8741.90/40.4159.18/55.6154.28/53.53 Table 7: Detailed results of ablation study on the “random turn selection” on MT-CSD for five models and the comparison with our method. For each group, the best value for accuracy and macro F1 is shown in bold. Model MT-CSD (acc. / F1 in %) BitcoinTeslaSpaceXBidenTrumpOverall Qwen2.5-3B44.80/44.73 60.37/54.30 63.86/61.16 46.72/34.94 55.44/42.93 53.77/50.55 Qwen2.5-7B44.61/42.98 72.78/62.50 69.16/62.97 47.20/33.60 70.05/55.33 60.82/54.94 Qwen2.5-14B50.56/50.14 71.67/66.47 64.49/62.28 48.18/39.29 62.21/46.86 59.60/57.25 Qwen3-4B-2507 46.65/46.73 68.89/61.26 65.11/60.62 48.91/34.88 69.16/52.76 59.93/56.33 Llama-3.1-8B50.93/49.26 62.96/58.93 53.58/52.78 45.01/38.87 59.36/46.85 55.00/53.96 Table 8: Detailed results of ablation study on the entropy threshold (entropy = 0.0) on MT-CSD for five models. Model MT-CSD (acc. / F1 in %) BitcoinTeslaSpaceXBidenTrumpOverall Qwen2.5-3B44.42/44.30 60.74/53.97 62.62/59.28 46.47/34.18 56.51/43.51 53.82/50.23 Qwen2.5-7B43.31/41.05 72.41/60.80 67.91/60.09 48.18/32.39 70.94/52.37 60.65/53.35 Qwen2.5-14B46.10/45.87 73.52/64.53 68.22/64.64 45.74/33.53 67.74/53.73 60.40/56.66 Qwen3-4B-2507 45.91/45.77 70.56/59.83 66.98/60.30 46.23/31.10 69.70/50.93 60.06/54.89 Llama-3.1-8B50.74/49.09 63.15/59.49 53.58/52.61 43.80/37.65 59.71/47.32 54.87/53.71 Table 9: Detailed results of ablation study on the entropy threshold (entropy = 0.2) on MT-CSD for five models. Model MT-CSD (acc. / F1 in %) BitcoinTeslaSpaceXBidenTrumpOverall Qwen2.5-3B43.87/43.74 61.11/54.23 63.24/59.23 46.23/33.93 55.97/42.95 53.69/49.97 Qwen2.5-7B43.31/40.79 71.67/59.64 67.91/59.72 47.93/31.71 71.30/52.73 60.52/52.85 Qwen2.5-14B46.65/46.35 73.33/63.90 68.85/65.17 46.47/33.75 67.74/51.42 60.69/56.69 Qwen3-4B-2507 45.54/45.46 70.37/59.82 66.98/60.12 46.47/31.24 69.34/50.67 59.89/54.72 Llama-3.1-8B51.12/49.55 62.41/58.28 53.27/52.36 43.07/37.02 59.54/47.66 54.58/53.47 Table 10: Detailed results of ablation study on the entropy threshold (entropy = 0.4) on MT-CSD for five models. Model MT-CSD (acc. / F1 in %) BitcoinTeslaSpaceXBidenTrumpOverall Qwen2.5-3B45.17/45.05 62.78/55.08 64.49/60.19 44.53/33.08 58.11/44.64 54.74/50.96 Qwen2.5-7B43.31/40.33 70.93/58.52 66.36/57.45 47.69/30.73 71.30/52.68 60.10/51.59 Qwen2.5-14B47.58/47.02 73.15/63.44 71.34/66.95 46.96/33.67 67.38/51.63 61.20/56.92 Qwen3-4B-2507 46.28/46.04 68.89/58.71 67.60/60.52 46.23/31.58 68.63/50.19 59.60/54.43 Llama-3.1-8B51.30/49.53 60.00/55.65 54.83/53.92 41.12/35.81 58.29/44.75 53.65/52.35 Table 11: Detailed results of ablation study on the entropy threshold (entropy = 0.8) on MT-CSD for five models. Model MT-CSD (acc. / F1 in %)ZS-CSD (acc. / F1 in %) BitcoinTeslaSpaceXBidenTrump GPT-4o-mini (single) 49.07/48.33 71.48/67.73 61.99/59.38 44.28/34.61 66.49/53.7243.65/38.94 GPT-4o-mini (multi) 53.53/53.48 69.26/62.61 66.67/64.86 48.18/39.96 68.81/51.9645.99/42.70 Table 12: The detailed results of GPT-4o-mini under single-turn and multi-turn inputs on MT-CSD and ZS-CSD. Turn 2 Turn 0: this sentence questions the consistency of the quality build of the tesla model, specifically the plaid, which contradicts the previous statement about its quality. Turn 1:$140k Tesla quality. Turn 2:Some assembly is required! All jokes aside, that must sting for the owner (buyers remorse) Turn 3:My friend has the same model...plaid. Can confirm the same quality build. But he showed me the dash bends and moves just as bad! For that amount I’d rather pick up the lower end Taycan! Turn 4 (Target utterance): is literally why I bought a Mercedes EQS! I have zero of these issues! Target: Tesla Gold stance: against COT baseline Ours pred: against \ [answer] not forget to add the [assessment criteria] in your response. [assessment criteria] - if the utterance supports or agrees with the target entity → \"favor\". - if the utterance opposes, criticizes, or questions the target entity → \"against\". - if the utterance is neutral or irrelevant to the target entity → \"none\". follow the format to the letter. am ready to receive the answer. ...... Garbled text Turn 3 Turn 2: this sentence elaborates on the previous sentence, providing evidence for the speaker's decision to purchase a mercedes eqs, specifically highlighting the issues with the tesla model and contrasting them with their own positive experience with the mercedes." Figure 7: Additional qualitative example of our method and CoT baseline. The CoT baseline provides garbled text in some cases. Turn 1:"$140k Tesla quality Turn 2:There are body shops that specialize in de-sucking tesla's build quality Turn 3:Imagine spending 140k and this being a necessity tho Turn 4:It won't last long. Every manufacturer has electric cars now so theres real competition, and Lucid makes a much faster car now with actual luxury car quality. The only ones buying teslas are tech dorks with more money than brains and people who put a deposit down before Ol' Musky Super Genius went full mask off rightwing nutbar. Turn 5:It’s so crazy I used to look up to the guy and now omg..... it’s depressing AF. Turn 6 (Target utterance):Good on you for changing your mind on him though, a lot of people seemingly can't or won't do that." Target: Tesla Gold stance: against COT baseline Ours Turn 4 Turn 3: ...questions the necessity of spending 140k on a tesla, suggesting ... and that the brand's appeal may be limited to a specific niche. Turn 3 Turn 1: this sentence questions the necessity of spending $140k on a tesla, implying that it may not be a necessary expense. pred: against ... the utterance “......" expresses a positive sentiment toward the person who changed their opinion about tesla, implying that this person is commendable for doing so. the utterance does not directly express a stance toward tesla itself but rather toward the person's change of heart. however, considering the context of the conversation, which heavily criticizes tesla and its founder, the utterance can be seen as indirectly supporting tesla by acknowledging someone who has changed their opinion about it. therefore, the stance toward tesla, based on the provided criteria, is "favor". ... pred: favor Turn 6 Turn 5: this sentence elaborates on the previous statement about people's willingness to change their opinions... Figure 8: Additional qualitative example of our method and CoT baseline. Prompt of single-turn input stance detection evaluation You are an expert stance detection evaluator. Your task is to assess the stance of the given SENTENCE toward a given TARGET entity. Follow the [Assessment Criteria] strictly when making your judgment. [SENTENCE to evaluate] utterance_text [TARGET entity] target_entity [Assessment Criteria] - If the utterance supports or agrees with the TARGET entity→ output "favor". - If the utterance opposes, criticizes, or questions the TARGET entity→ output "against". - If the utterance is neutral or irrelevant to the TARGET entity→ output "none". Your output must be exactly one label from favor, against, none, with no extra words or punctuation. Your output is: Table 13: Prompt of single-turn input stance detection evaluation. Prompt of multi-turn input stance detection evaluation You are an expert stance detection evaluator. Your task is to assess the stance expressed in the final utterance of a conversation toward a given TARGET entity. Follow the [Assessment Criteria] strictly when making your judgment. [Conversation] conversation_texts [Utterance to evaluate] (This is always the last utterance in the conversation) utterance_text [TARGET entity] target_entity [Assessment Criteria] - If the utterance supports or agrees with the TARGET entity→ output "favor". - If the utterance opposes, criticizes, or questions the TARGET entity→ output "against". - If the utterance is neutral or irrelevant to the TARGET entity→ output "none". Your output must be exactly one label from favor, against, none, with no extra words or punctuation. Your output is: Table 14: Prompt of multi-turn input stance detection evaluation. Prompt of CoT baseline stance detection evaluation You are an expert stance detection evaluator. Your task is to assess the stance expressed in the final utterance of a conversation toward a given TARGET entity. The stance can be "favor", "against", or "none". Follow the [Assessment Criteria] strictly when making your judgment. [Conversation] conversation_texts [Utterance to evaluate] (This is always the last utterance in the conversation) utterance_text [TARGET entity] target_entity [Assessment Criteria] - If the utterance supports or agrees with the TARGET entity→ "favor". - If the utterance opposes, criticizes, or questions the TARGET entity→ "against". - If the utterance is neutral or irrelevant to the TARGET entity→ "none". Let’s think step by step, and put your answer in the . Your output is: Table 15: Prompt of CoT baseline stance detection evaluation. Prompt of Relevance Detection You are a relevance detection assistant. Your task is to determine whether the given UTTERANCE is relevant to the TARGET entity. You are also provided with the whole CONVERSATION for context. [CONVERSATION] conversation_text [UTTERANCE TO EVALUATE] utterance_text [TARGET ENTITY] target_entity [Decision Rule] - If the utterance talks about, comments on, or is clearly related to the TARGET entity, answer “yes”. - If the utterance is not about the TARGET entity, or is clearly irrelevant, answer “no”. Your answer must be exactly one word: “yes” or “no”, with no extra text. Your answer: Table 16: Prompt of Relevance Detection. Prompt of Target-related Proposition Extraction You are an information extraction system. From the given SENTENCE and the TARGET, extract minimal atomic statements (propositions) that are explicitly about the TARGET from the SENTENCE. You should output JSON only. [Input] - SENTENCE: <PUT_SENTENCE_HERE> - TARGET: <PUT_TARGET_HERE> [Instructions] - Extract ONLY statements that concern the TARGET. - An "atomic statement" is the shortest target-related statement that stands alone in meaning. Split coordination (e.g., "A and B") into separate statements. - Check if pronouns or vague in the SENTENCE refer to the TARGET. - Replace any pronouns or vague references in the sentence (e.g., "it", "this car", "that model") with the exact TARGET phrase so each statement is explicitly target-related. - Focus on content types such as facts, experiences, evaluations, comparisons, causes/effects, time, and conditions. Ignore greetings, fillers, and chit-chat. - Attributes are an OPEN set of concise tags you choose. - If there are no target-related statements, return "statements": []. [Output JSON schema] "statements": [ "text": "<statement1>", "attributes": "<tag1>" , "text": "<statement2>", "attributes": "<tag2>" , ...... ] You should output ONLY valid JSON. Your output is: Table 17: Prompt of Target-related Proposition Extraction. Prompt of Relation Induction You are an information extraction system. Given statements from the CURRENT SENTENCE and the PREVIOUS SENTENTCE that the current sentence replies to, extract relations between statement pairs. You should output JSON only. [Input] TARGET: <PUT_TARGET_HERE> CURRENT_SENTENCE: <PUT_CURRENT_SENTENCE_HERE> PREVIOUS_SENTENCE: <PUT_PREVIOUS_SENTENCE_HERE> statements from CURRENT SENTENCE: "statements": [ "text": "<statement 1>", "attribute": "<tag 1>", "text": "<statement 2>", "attribute": "<tag 2>" ... ] statements from PREVIOUS SENTENCE: "statements": [ "text": "<statement 1>", "attribute": "<statement 1>", "text": "<statement 2>", "attribute": "<statement 2>" ... ] [Instructions] Link ONLY pairs that clearly talk about the same aspect of the TARGET. Each pair gets EXACTLY ONE relation label (no multi-label). Choose a concise label; examples: ["supports","against","doubts","compares","questions","quotes","refers_to","evidence_for"]. You CAN use another short labels if more precise. The direction of the relation is always from the statement of CURRENT SENTENCE to the statement of PREVIOUS SENTENCE. Do NOT invent links. If no clear link, output no edge for that pair. Provide a very short rationale. If there are no valid links at all, return "edges": []. [Output JSON schema] "edges": [ "source_text": "<one statement from statements from CURRENT SENTENCE>", "target_text": "<one statement from statements from PREVIOUS SENTENCE>", "relation": "<a relation>", "rationale": "<very brief cue or phrase from the sentences>" , ... , ...... ] You should output ONLY valid JSON. Your output is: Table 18: Prompt of Relation Induction. Prompt of Natural Language Relation Rephraser You are a faithful rephraser. Using ONLY the inputs below, write ONE short English paragraph that summarizes how the CURRENT sentence’s statements relate to the PREVIOUS sentence’s statements about the TARGET. Do NOT add facts. Output ONLY the paragraph. [TARGET] target_entity [PREVIOUS sentence] prev_sentence [CURRENT sentence] current_sentence [STATEMENTS of the previous sentence] // verbatim list already extracted statements_prev_json [STATEMENTS of the current sentence] // verbatim list already extracted statements_curr_json [RELATIONS from the current sentence to the previous sentence] // verbatim edges already extracted relations_json [Style] - Refer to the relation labels provided (e.g., supports/against/elaborates/questions/compares/refers_to/evidence_for/causal). - You should start with "This sentence ...". Your output is: Table 19: Prompt of Natural Language Relation Rephraser. Prompt of Stance Detection of TamGraph You are an expert stance detection evaluator. Your task is to assess the stance expressed in the final utterance of a conversation toward a given TARGET entity. You should consider the information in the [Per-turn Internal Relations], and follow the [Assessment Criteria] strictly when making your judgment. [Conversation] conversation_texts [Per-turn Internal Relations] # Natural-language descriptions of how each utterance relates to its immediately previous utterance (i→ i-1). # One line per adjacent pair: Turn i→ Turn i-1, faithfully describes how Turn i relates to Turn i-1 per_turn_internal_relations [Utterance to evaluate] (This is always the last utterance in the conversation) utterance_text [TARGET entity] target_entity [Assessment Criteria] - If the utterance supports or agrees with the TARGET entity→ output "favor". - If the utterance opposes, criticizes, or questions the TARGET entity→ output "against". - If the utterance is neutral or irrelevant to the TARGET entity→ output "none". Your output must be strictly one word, exactly one label from favor, against, none, with no extra words or punctuation. Your output is: Table 20: Prompt of Stance Detection of TamGraph.