Pith. sign in

REVIEW 4 major objections 6 minor 79 references

Enhancing Social Intelligence in LLMs with Hierarchical Reasoning and Utterance-Level Goal Rewarding

T0 review · 4 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read A two-stage think-strategy-response pipeline with variance-gated rewards lets a 7B LLM beat GPT-4o at social goal completion by 7.32 percent.

desk verdict Solid, modestly novel social-agent RL paper with a load-bearing evaluation-cleanliness problem: hyperparameters tuned on the same test set that carries the headline result. read the letter →

arxiv 2608.05832 v1 pith:VYEGCBLO submitted 2026-08-06 cs.CL

Xiaofeng Wang , Kakam Chong , Shuai Xiao , DeXin Kong , Qingyuan Tian , Chen Ju , Xu Yan , Shuai Zhao
show 4 more authors
Fei Huang Rui Wang Shuguang Han jufeng chen
This is my paper · ORCID
classification cs.CL
keywords socialintelligencelargelanguagemodelshierarchicalreinforcementlearningvariance-gatedrewardroutingstrategyplanningdialoguegenerationSOTOPIAbenchmarkgoal-completion
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper claims that social dialogue for LLMs is better trained as a two-stage process—first produce a strategy, then produce the utterance—rather than optimizing every turn against a single goal-completion reward. To train such a generator, it introduces LHRL-VGR, which routes each turn's reward between goal completion and strategy adherence based on how much the goal-completion score varies across repeated samples: when scores are stable, strategy adherence is rewarded; when uncertain, goal completion is. The claim is that this variance-gated routing prevents the model from learning shortcut utterances and instead fosters human-like social inference. On SOTOPIA and SOTOPIA-Hard, a Qwen2.5-7B agent trained this way surpasses GPT-4o in goal completion by the reported relative gain of 7.32 percent, which the paper takes as evidence that hierarchical reasoning plus utterance-level reward routing is the right direction for social intelligence.

What carries the argument

The central objects are the two-stage TSR generation pipeline and the variance-gated reward mechanism (VGRM). In TSR, a strategy model first outputs a 'think' rationale and a strategy (behavior guidance plus an utterance-level objective), and a response model then generates the utterance conditioned only on the strategy, not on the raw thinking. VGRM decides the response reward: compute the variance $\mathrm{Var}(r_\phi(s,a_1),\ldots,r_\phi(s,a_k))$ of goal-completion scores across $k$ sampled responses; if the variance exceeds a threshold $\theta$, use the boundary-scaled goal-completion reward times a non-repetition penalty; otherwise use a strategy-followed reward times the same penalty. Training uses a two-stage group-relative policy optimization in which the stage-1 advantage is $\alpha A_s + (1-\alpha)A_{tg}$ and the stage-2 advantage is the normalized routing reward within a group of $G_2$ samples.

What would settle it

Fix the thresholds at $\theta=0.3$ and $\alpha=0.4$, then run LHRL-VGR on a held-out set of SOTOPIA scenarios using a different reward model than Claude-3.5-Sonnet (say GPT-4o as judge) and also on a different negotiation benchmark; if the goal-completion advantage over the GPT-4o baseline disappears, or if performance returns only after retuning $\theta$ and $\alpha$, then variance-gated routing is not the load-bearing mechanism claimed.

Watch

Extended reading notes

Core claim

The central claim is that the Think-Strategy-Response (TSR) framework—decomposing each dialogue turn into a hidden strategic plan and a surface response—together with the LHRL-VGR training algorithm gives a 7B-parameter agent state-of-the-art social negotiation ability. The decisive mechanism is the variance-gated reward router: for each turn, the algorithm samples several responses, computes an LLM-assigned goal-completion score for each, and uses the variance of those scores to choose whether to reinforce goal completion or strategy adherence. A second mechanism is the composite stage-1 advantage, which blends a strategy-content reward (predicted outcome plus completeness) with a transferred goal-completion reward from the actual response, weighted by a scalar $\alpha$. With these, the fine-tuned Qwen2.5-7B scores 8.78 GOAL on SOTOPIA and 7.65 on SOTOPIA-Hard, against 8.19 and 6.97 for GPT-4o, with the paper reporting the differences as statistically significant over four evaluation runs.

Load-bearing premise

The whole result rests on the assumption that the run-to-run variation in the reward model's goal-completion scores is a trustworthy signal for choosing which reward to train on, and that the two fixed settings ($\theta=0.3$, $\alpha=0.4$) chosen on one benchmark will keep working elsewhere without retuning.

Editorial extensions

If this is right

  • A 7B open-weight model can outscore GPT-4o on goal completion in social negotiation, suggesting that focused reward design can substitute for model scale in this domain.
  • Reward routing per utterance should depend on measurement stability: use goal-completion rewards when goal scores are noisy, and strategy-adherence rewards when they are stable, rather than applying one uniform reward to every turn.
  • Joint reinforcement learning over both the strategy and response stages outperforms training either stage alone, indicating that strategy quality and execution quality are coupled and should be optimized together.
  • The two-stage generation produces explicitly interpretable social behavior—visible strategies and rationales—making the agent's choices more auditable than end-to-end dialogue models.
  • The paper's variance analysis of the LLM evaluator shows that even semantically equivalent responses receive unstable goal scores, so reward design must account for judge noise instead of treating judge scores as ground truth.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The variance-gating principle is a general recipe: any RL pipeline with a noisy learned evaluator could route between outcome rewards and process rewards by measuring evaluator variance, which may transfer beyond dialogue to code generation or mathematical reasoning.
  • Because the thresholds $\theta=0.3$ and $\alpha=0.4$ were tuned on SOTOPIA-Hard, the most direct out-of-paper test is whether the same thresholds transfer to other negotiation datasets or to a different reward model without retuning; if they do, the method is robust, and if not, a principled threshold-selection rule would be needed.
  • The paper's case studies suggest the failure mode its rewards target is dialogue stall (looping on pleasantries). A natural extension is to measure per-turn novelty and goal progress in open-domain agents, predicting that variance-gated strategy-followed rewards prevent such stalls in conversations longer than the 20-turn SOTOPIA cap.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 6 minor

Summary. The paper proposes the Think-Strategy-Response (TSR) framework, which decomposes social dialogue generation into a high-level strategy stage and a low-level response stage, and the Linearized Hierarchical Reinforcement Learning with Variance-Gated Rewards (LHRL-VGR) algorithm, which routes between goal-completion and strategy-followed rewards based on the variance of an LLM judge's goal-completion scores. The method is instantiated with Qwen2.5-7B-Instruct, trained on SOTOPIA-π, and evaluated on SOTOPIA and SOTOPIA-Hard. The central claim is that the trained 7B agent surpasses GPT-4o by 7.32% in goal-completion success, establishing state-of-the-art performance in multi-agent social negotiation.

Significance. If the empirical claims survive scrutiny, this is a meaningful contribution: the TSR decomposition is a natural and interpretable way to inject strategic planning into social dialogue, and the variance-gated reward routing is a novel mechanism for deciding when to trust a goal-completion reward versus a strategy-adherence reward. The manuscript is unusually detailed on prompts, hyperparameters, training data construction, and reward rubrics, and it releases code and training data, which is a concrete reproducibility asset. The ablation study and the case-study tables in Appendix L give useful qualitative evidence that the individual reward components affect behavior in plausible ways. However, the headline quantitative claim is currently not supported by a clean evaluation protocol, because key hyperparameters are tuned on the same benchmark used for the main result and the reported improvement percentage is not reproducible from the table. The significance of the work therefore depends on resolving these evaluation issues.

major comments (4)
  1. [Appendix G, Table 1] The variance threshold θ (Eq. 7) and the stage-weight α (Eq. 11) are selected by maximizing GOAL on SOTOPIA-Hard, which is the same test set used in Table 1 to demonstrate the headline result. With only 70 hard tasks and a grid of candidate values, this is test-set selection, not model selection: the reported 7.65 GOAL on SOTOPIA-Hard is inflated relative to what a pre-specified protocol would yield. No held-out validation, nested selection, or correction for multiple comparisons is described. The authors should either report results from a pre-registered hyperparameter choice, perform selection on a validation split, or provide an analysis of sensitivity that does not use the final test set.
  2. [Abstract, Table 1, footnote 2] The abstract claims a 7.32% improvement in goal-completion success over GPT-4o, but this number is not derivable from Table 1. The relative GOAL gains over GPT-4o are +7.20% on SOTOPIA (8.78 vs. 8.19) and +9.76% on SOTOPIA-Hard (7.65 vs. 6.97); no weighted or unweighted combination of these values yields 7.32%. Footnote 2 states that the figure is a relative gain computed from raw scores, but it does not specify the aggregation that produces 7.32%. The authors should state the exact formula, correct the abstract, or remove the unverifiable percentage.
  3. [Section 5.1, Table 1] The claim that the main results are 'statistically significant with p < 0.05' is unsupported: the table reports only means, with no standard deviations, confidence intervals, or statistical test details. The evaluation is averaged over four runs of the SOTOPIA tasks, which gives very limited information about the distribution of scores for a 450-task / 70-task evaluation. The authors should report per-run scores, the variance across runs, the test used, and the resulting p-values; without this, the significance claim cannot be checked and may not survive a paired test with only four replicates.
  4. [Section 4.1, Eq. (7), Appendix D] The variance-gated routing mechanism assumes that the variance of the LLM evaluator's goal-completion scores across sampled responses is a reliable signal for choosing between goal-completion and strategy-followed rewards. Appendix D demonstrates that the evaluator exhibits non-negligible variance even for semantically equivalent inputs, but it does not establish that variance is a valid proxy for reward reliability, nor that a threshold of θ=0.3 is robust. Since this assumption is the core of the proposed algorithm and is coupled with the test-set tuning in Appendix G, the paper should provide a derivation or an independent validation of this design choice, for example by showing that the routing decision correlates with better outcomes on a held-out split.
minor comments (6)
  1. [Section 2] In the related-work paragraph, the proposed method is referred to as 'LRVR-VGR'; this should be 'LHRL-VGR'.
  2. [Section 4.1.1, Eq. (6)] Equation (6) defines the non-repetition reward using R2(a_i, a_{i-2}), but the text says the reward penalizes overlap with the model's 'prior turn.' Under standard user-assistant alternation, the model's prior turn would be a_{i-2}, but this indexing should be stated explicitly to avoid confusion.
  3. [Appendix J.2, Prompt 5] The scoring instruction in Prompt 5 says 'score must be an integer from 0–10', while both the rubric and the JSON template specify a 0–5 scale. This inconsistency should be corrected.
  4. [Figure 2] The figure contains the label 'satge-1outputs'; this is a typo for 'stage-1 outputs'.
  5. [References] The EPO paper is cited twice as Liu et al. 2025a and Liu et al. 2025b with the same arXiv identifier; a single citation should be used.
  6. [Table 3] The hyperparameter table lists 'KL Loss Coef' and 'KL Coef' as separate entries with the same value; one of them appears to be redundant or mislabeled.

Circularity Check

1 steps flagged · score 5.0 of 10

SOTOPIA-Hard gains are partly self-confirming: Appendix G tunes the routing threshold and stage weight on the same SOTOPIA-Hard GOAL metric that Table 1 then reports as the headline result.

  1. fitted input called prediction [Appendix G (Eq. 7 and Eq. 11) versus main results in Table 1]
    "In this section, we conduct experiments to investigate the influence of hyperparameters in Equation 7 and Equation 11. We employ qwen2.5-7b-instruct as the base model and evaluate performance on the SOTOPIA-Hard task. ... Optimal performance is achieved at θ=0.3, while GOAL performance decreases monotonically as θ increases. ... The experimental results show that setting α=0.4 can achieve a more balanced advantage in stage-1, leading to an outstanding performance in GOAL."

    The two hyperparameters that define the proposed mechanism are the variance-gating threshold θ in Eq. 7 and the stage-balancing weight α in Eq. 11. Appendix G selects both by maximizing GOAL on SOTOPIA-Hard, and Table 1 then reports the SOTOPIA-Hard GOAL of 7.65 as the central evidence that LHRL-VGR surpasses GPT-4o. The selection and the headline evaluation use the same benchmark and the same metric, with no described validation split, nested selection, or correction for multiple comparisons. The reported SOTOPIA-Hard advantage is therefore not an independent prediction; it is partly a fitted result, forced to be at least as high as the other θ and α values tried on that exact test set.

full rationale

Most of the derivation chain is self-contained: the TSR framework is a prompt-based decomposition, LHRL-VGR is a concrete RL objective in Eq. 13, and the rewards are defined through LLM evaluators rather than through the final metric by construction. However, the paper's headline empirical claim is compromised by one fitted-input step: θ and α are tuned directly on SOTOPIA-Hard GOAL in Appendix G, and the same SOTOPIA-Hard GOAL score is then used in Table 1 to claim state-of-the-art performance. This makes the SOTOPIA-Hard result partly self-confirming rather than a clean out-of-sample prediction. Separately, the abstract's '+7.32% improvement' is not reproducible from Table 1: the relative GOAL gains over GPT-4o are +7.20% on SOTOPIA and +9.76% on SOTOPIA-Hard, and no aggregation yields 7.32%; the footnote does not resolve this. The core algorithmic idea—variance-gated routing—still has independent content supported by ablations, alternative-routing comparisons, and a small human evaluation, so the paper is not wholly circular. The appropriate circularity score is 5.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The central claim rests on the validity of the SOTOPIA benchmark and LLM-as-judge evaluations, plus an ad hoc assumption that variance in LLM goal scores is a reliable routing signal. The two free parameters (theta, alpha) are tuned on the target test set, which is the main circularity concern.

free parameters (3)
  • variance threshold theta = 0.3
    Routing threshold in Eq. 7; chosen by sweeping on SOTOPIA-Hard (Appendix G), the same benchmark used for the main results.
  • stage-1 advantage weight alpha = 0.4
    Weight between strategy-content and transferred-goal advantages in Eq. 11; selected on SOTOPIA-Hard (Appendix G).
  • goal-completion data threshold = 8
    Score below which a turn is labeled incomplete for training-data categorization (Appendix A.2); hand-set, affects which turns enter RL.
assumptions (4)
  • domain assumption SOTOPIA and its GOAL score are a valid measure of social intelligence.
    The paper evaluates on SOTOPIA and SOTOPIA-Hard (Section 5.1) and treats GOAL as the primary metric; if the benchmark is not a good proxy, the head-to-head claim loses force.
  • domain assumption GPT-4o and Claude-3.5-Sonnet LLM judgments correlate with human judgments and are stable enough for training and evaluation.
    GPT-4o is used as judge and partner, Claude-3.5-Sonnet as reward model (Sections 5.1, A.2); the paper cites prior correlation studies but does not re-validate for its setups.
  • ad hoc to paper The variance of goal-completion scores across sampled responses indicates whether to prefer goal-completion or strategy-followed rewards.
    Eq. 7 routes on Var(r_phi) with threshold theta; Appendix D only shows that the evaluator is noisy, not that variance is a reliable reliability signal.
  • domain assumption Generating a strategy before a response improves goal achievement over end-to-end generation.
    The TSR framework (Section 3) assumes this; TSR vanilla (6.74) improves only slightly over Qwen2.5-7B-Instruct (6.71), so the assumption is weakly supported.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Enhancing Social Intelligence in LLMs with Hierarchical Reasoning and Utterance-Level Goal Rewarding." pith.science (2026). https://pith.science/paper/VYEGCBLO

@misc{pith2026260805832,
  author       = {Pith},
  title        = {Pith review of: Enhancing Social Intelligence in LLMs with Hierarchical Reasoning and Utterance-Level Goal Rewarding},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/VYEGCBLO}},
  note         = {Machine review of arXiv:2608.05832}
}
read the original abstract

Large language models (LLMs) excel in structured tasks but struggle with dynamic social interactions, where success requires long-term goal coordination and rapid adaptation. Current methods often apply uniform goal-based rewards to every utterance, overlooking the specificity of objectives at each dialogue turn and failing to account for the rationale of potential strategies. Inspired by the Theory of Planned Behavior, we propose the Think-Strategy-Response (TSR) framework, which decomposes social dialogue into two hierarchical stages: high-level strategic planning and low-level linguistic execution. To optimize TSR, we introduce Linearized Hierarchical Reinforcement Learning with Variance-Gated Rewards (LHRL-VGR), a novel algorithm that dynamically routes rewards - balancing goal completion and strategy adherence - based on the variance of goal achievement scores. Experiments on the SOTOPIA benchmark show that our approach fine-tunes a Qwen2.5-7B agent to surpass the GPT-4o baseline by 7.32% in goal completion success, demonstrating state-of-the-art performance in multi-agent social negotiation tasks.

Figures

Figures reproduced from arXiv: 2608.05832 by the authors.

Figure 1
Figure 1. An exemplary scenario demonstrating the TSR framework. The social task in this work [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. The pipline of LHRL-VGR. Therefore, at turn i our framework follows the following output pipeline: ti , si ∼ πstrategy(· | sstrategy, p, g, h1:i−1), (2) ai ∼ πresponse(· | sresponse, p, si , g, h1:i−1). (3) where sstrategy and sresponse denote the prompt contents for the two stages; see Appendix J.1 for details. This two-stage decomposition enables models to concentrate on specific components of generation. It also … view at source ↗
Figure 3
Figure 3. Performance Comparison of different Training Paradigm. WIN-S and WIN-R are [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Performance heatmap of using different models for Stage 1 and Stage 2. To validate the design, we conducted ablation studies on the various reward components, with results summarized in Ta￾ble 2. Notably, the removal of the strategy completeness reward leads to a signi…
Figure 5
Figure 5. Figure 5: Variance Distribution Comparison. was calculated for each of the 50 dialogue sets. The distribution of these variances is shown in Fig￾ure 5. The results indicate that even for semantically equivalent or highly similar content, the LLM evaluator exhibits non-negligible…
Figure 6
Figure 6. Figure 6: GOAL score across Different Model Pairings. The horizontal axis corresponds to the [PITH_FULL_IMAGE:figures/full_fig_p017_6.png]
Figure 7
Figure 7. Figure 7: The GOAL performance under differ￾ent θ. 0.0 0.2 0.4 0.6 0.8 7.1 7.2 7.3 7.4 7.5 7.6 7.7 GOAL [PITH_FULL_IMAGE:figures/full_fig_p018_7.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

79 extracted references · 53 canonical work pages

  1. [1]

    The theory of planned behavior

    Icek Ajzen. The theory of planned behavior. Organizational behavior and human decision processes, 50 0 (2): 0 179--211, 1991

    work page 1991

  2. [2]

    Claude 3.5 sonnet model card addendum

    AI Anthropic. Claude 3.5 sonnet model card addendum. Claude-3.5-Sonnet Model Card, 2024. URL https://www.anthropic.com/claude/sonnet

    work page 2024

  3. [3]

    Efficacy of the theory of planned behaviour: A meta-analytic review

    Christopher J Armitage and Mark Conner. Efficacy of the theory of planned behaviour: A meta-analytic review. British journal of social psychology, 40 0 (4): 0 471--499, 2001

    work page 2001

  4. [4]

    Social foundations of thought and action

    Albert Bandura et al. Social foundations of thought and action. Englewood Cliffs, NJ, 1986 0 (23-28): 0 2, 1986

    work page 1986

  5. [5]

    Theory of planned behavior

    Mark Conner. Theory of planned behavior. Handbook of sport psychology, pp.\ 1--18, 2020

    work page 2020

  6. [6]

    Gemini 2.0 flash thinking

    Google DeepMind. Gemini 2.0 flash thinking. Gemini 2.0 flash thinking System Card, 2024. URL https://deepmind.google/technologies/ gemini/flash-thinking/

    work page 2024

  7. [7]

    Plug-and-play policy planner for large language model powered dialogue agents

    Yang Deng, Wenxuan Zhang, Wai Lam, See-Kiong Ng, and Tat-Seng Chua. Plug-and-play policy planner for large language model powered dialogue agents. In Proceedings of ICLR, 2024. URL https://openreview.net/forum?id=MCNqgUFTHI

    work page 2024

  8. [8]

    Frames of mind: The theory of multiple intelligences

    Howard E Gardner. Frames of mind: The theory of multiple intelligences. Basic books, 2011

    2011

Show all 79 references
  1. [10]

    Large language model based multi-agents: a survey of progress and challenges

    Taicheng Guo, Xiuying Chen, Yaqi Wang, Ruidi Chang, Shichao Pei, Nitesh V Chawla, Olaf Wiest, and Xiangliang Zhang. Large language model based multi-agents: a survey of progress and challenges. In Proceedings of IJCAI, pp.\ 8048--8057, 2024. URL https://dl.acm.org/doi/abs/10.2...

    2024 doi

  2. [11]

    Planning like human: A dual-process framework for dialogue planning

    Tao He, Lizi Liao, Yixin Cao, Yuanxing Liu, Ming Liu, Zerui Chen, and Bing Qin. Planning like human: A dual-process framework for dialogue planning. In Proceedings of ACL, pp.\ 4768--4791, 2024. URL https://aclanthology.org/2024.acl-long.262/

    2024

  3. [14]

    Social intelligence

    John F Kihlstrom and Nancy Cantor. Social intelligence. In R. J. Sternberg (ed.), Handbook of Intelligence, pp.\ 359--379. Cambridge University Press, Cambridge, 2000

    2000

  4. [15]

    Interactional foundations of language: The interaction engine hypothesis

    Stephen C Levinson. Interactional foundations of language: The interaction engine hypothesis. In Human language: From genes and brain to behavior, pp.\ 189--200. MIT Press, 2019

    2019

  5. [17]

    Rouge: A package for automatic evaluation of summaries

    Chin-Yew Lin. Rouge: A package for automatic evaluation of summaries. In Text summarization branches out, pp.\ 74--81, 2004

    2004

  6. [21]

    Openai o3-mini system card

    OpenAI. Openai o3-mini system card. OpenAI o3-mini System Card, 2025. URL https://openai.com/index/o3-mini-system-card/

    2025

  7. [22]

    Hierarchical reinforcement learning: A comprehensive survey

    Shubham Pateria, Budhitama Subagdja, Ah-hwee Tan, and Chai Quek. Hierarchical reinforcement learning: A comprehensive survey. ACM Computing Surveys (CSUR), 54 0 (5): 0 1--35, 2021

    2021

  8. [23]

    A framework for studying the neurobiology of value-based decision making

    Antonio Rangel, Colin Camerer, and P Read Montague. A framework for studying the neurobiology of value-based decision making. Nature reviews neuroscience, 9 0 (7): 0 545--556, 2008

    2008

  9. [24]

    A simplest systematics for the organization of turn-taking for conversation

    Harvey Sacks, Emanuel A Schegloff, and Gail Jefferson. A simplest systematics for the organization of turn-taking for conversation. language, 50 0 (4): 0 696--735, 1974. URL https://muse.jhu.edu/pub/24/article/452679/summary

    1974

  10. [26]

    Relevance: Communication and cognition, volume 142

    Dan Sperber and Deirdre Wilson. Relevance: Communication and cognition, volume 142. Harvard University Press Cambridge, MA, 1986

    1986

  11. [27]

    Feudal networks for hierarchical reinforcement learning, 2017

    Alexander Sasha Vezhnevets, Simon Osindero, Tom Schaul, Nicolas Heess, Max Jaderberg, David Silver, and Koray Kavukcuoglu. Feudal networks for hierarchical reinforcement learning, 2017. URL https://arxiv.org/abs/1703.01161

    2017 arXiv

  12. [28]

    Rema: Learning to meta-think for llms with multi-agent reinforcement learning, 2025

    Ziyu Wan, Yunxiang Li, Xiaoyu Wen, Yan Song, Hanjing Wang, Linyi Yang, Mark Schmidt, Jun Wang, Weinan Zhang, Shuyue Hu, and Ying Wen. Rema: Learning to meta-think for llms with multi-agent reinforcement learning, 2025. URL https://arxiv.org/abs/2503.09501

    2025 arXiv

  13. [29]

    Leave no document behind: Benchmarking long-context LLM s with extended multi-doc QA

    Minzheng Wang, Longze Chen, Fu Cheng, Shengyi Liao, Xinghua Zhang, Bingli Wu, Haiyang Yu, Nan Xu, Lei Zhang, Run Luo, Yunshui Li, Min Yang, Fei Huang, and Yongbin Li. Leave no document behind: Benchmarking long-context LLM s with extended multi-doc QA . In Proceedings of EMNLP...

    2024

  14. [32]

    Adaptive thinking via mode policy optimization for social language agents

    Minzheng Wang, Yongbin Li, Haobo Wang, Xinghua Zhang, Nan Xu, Bingli Wu, Fei Huang, Haiyang Yu, and Wenji Mao. Adaptive thinking via mode policy optimization for social language agents. arXiv preprint arXiv:2505.02156, 2025 b . URL https://arxiv.org/abs/2505.02156

    2025

  15. [33]

    Sotopia- : Interactive learning of socially intelligent language agents

    Ruiyi Wang, Haofei Yu, Wenxin Zhang, Zhengyang Qi, Maarten Sap, Yonatan Bisk, Graham Neubig, and Hao Zhu. Sotopia- : Interactive learning of socially intelligent language agents. In Proceedings of ACL, pp.\ 12912--12940, 2024 c . URL https://aclanthology.org/2024.luhme-long.698/

    2024

  16. [35]

    Debt collection negotiations with large language models: An evaluation system and optimizing decision making with multi-agent, 2025 d

    Xiaofeng Wang, Zhixin Zhang, Jinguang Zheng, Yiming Ai, and Rui Wang. Debt collection negotiations with large language models: An evaluation system and optimizing decision making with multi-agent, 2025 d . URL https://arxiv.org/abs/2502.18228

    2025 arXiv

  17. [36]

    Does changing behavioral intentions engender behavior change? a meta-analysis of the experimental evidence

    Thomas L Webb and Paschal Sheeran. Does changing behavioral intentions engender behavior change? a meta-analysis of the experimental evidence. Psychological bulletin, 132 0 (2): 0 249, 2006

    2006

  18. [38]

    React: Synergizing reasoning and acting in language models

    Shunyu Yao, Jeffrey Zhao, Dian Yu, Nan Du, Izhak Shafran, Karthik R Narasimhan, and Yuan Cao. React: Synergizing reasoning and acting in language models. In The Eleventh International Conference on Learning Representations, 2023. URL https://openreview.net/forum?id=WE_vluYUL-X

    2023

  19. [39]

    Sotopia-rl: Reward design for social intelligence, 2025

    Haofei Yu, Zhengyang Qi, Yining Zhao, Kolby Nottingham, Keyang Xuan, Bodhisattwa Prasad Majumder, Hao Zhu, Paul Pu Liang, and Jiaxuan You. Sotopia-rl: Reward design for social intelligence, 2025. URL https://arxiv.org/abs/2508.03905

    2025

  20. [42]

    A collaborative reasoning framework powered by reinforcement learning and large language models for complex questions answering over knowledge graph

    Zhiqiang Zhang and Wen Zhao. A collaborative reasoning framework powered by reinforcement learning and large language models for complex questions answering over knowledge graph. In Owen Rambow, Leo Wanner, Marianna Apidianaki, Hend Al-Khalifa, Barbara Di Eugenio, and Steven S...

    2025

  21. [43]

    Judging llm-as-a-judge with mt-bench and chatbot arena

    Lianmin Zheng, Wei-Lin Chiang, Ying Sheng, Siyuan Zhuang, Zhanghao Wu, Yonghao Zhuang, Zi Lin, Zhuohan Li, Dacheng Li, Eric Xing, et al. Judging llm-as-a-judge with mt-bench and chatbot arena. In Proceedings of NeurIPS, volume 36, pp.\ 46595--46623, 2023. URL https://openrevie...

    2023

  22. [44]

    L lama F actory: Unified efficient fine-tuning of 100+ language models

    Yaowei Zheng, Richong Zhang, Junhao Zhang, Yanhan Ye, and Zheyan Luo. L lama F actory: Unified efficient fine-tuning of 100+ language models. In Proceedings of ACL, System Demonstrations, pp.\ 400--410, 2024. URL https://aclanthology.org/2024.acl-demos.38

    2024

  23. [45]

    SOTOPIA : Interactive evaluation for social intelligence in language agents

    Xuhui Zhou, Hao Zhu, Leena Mathur, Ruohong Zhang, Haofei Yu, Zhengyang Qi, Louis-Philippe Morency, Yonatan Bisk, Daniel Fried, Graham Neubig, and Maarten Sap. SOTOPIA : Interactive evaluation for social intelligence in language agents. In Proceedings of ICLR, 2024. URL https:/...

    2024

  24. [46]

    2025 , eprint=

    ReMA: Learning to Meta-think for LLMs with Multi-Agent Reinforcement Learning , author=. 2025 , eprint=

    2025

  25. [47]

    A Collaborative Reasoning Framework Powered by Reinforcement Learning and Large Language Models for Complex Questions Answering over Knowledge Graph

    Zhang, Zhiqiang and Zhao, Wen. A Collaborative Reasoning Framework Powered by Reinforcement Learning and Large Language Models for Complex Questions Answering over Knowledge Graph. Proceedings of the 31st International Conference on Computational Linguistics. 2025

    2025

  26. [48]

    2025 , eprint=

    Adaptive Thinking via Mode Policy Optimization for Social Language Agents , author=. 2025 , eprint=

    2025

  27. [49]

    2025 , eprint=

    Sotopia-RL: Reward Design for Social Intelligence , author=. 2025 , eprint=

    2025

  28. [50]

    2011 , publisher=

    Frames of mind: The theory of multiple intelligences , author=. 2011 , publisher=

    2011

  29. [51]

    Handbook of Intelligence , editor =

    Kihlstrom, John F and Cantor, Nancy , title =. Handbook of Intelligence , editor =. 2000 , pages =

    2000

  30. [52]

    Englewood Cliffs, NJ , volume=

    Social foundations of thought and action , author=. Englewood Cliffs, NJ , volume=

  31. [53]

    Planning Like Human: A Dual-process Framework for Dialogue Planning

    He, Tao and Liao, Lizi and Cao, Yixin and Liu, Yuanxing and Liu, Ming and Chen, Zerui and Qin, Bing. Planning Like Human: A Dual-process Framework for Dialogue Planning. Proceedings of ACL. 2024

    2024

  32. [54]

    2025 , journal=

    Adaptive Thinking via Mode Policy Optimization for Social Language Agents , author=. 2025 , journal=

    2025

  33. [55]

    2020 , eprint=

    Towards Scalable Multi-domain Conversational Agents: The Schema-Guided Dialogue Dataset , author=. 2020 , eprint=

    2020

  34. [56]

    2025 , eprint=

    EPO: Explicit Policy Optimization for Strategic Reasoning in LLMs via Reinforcement Learning , author=. 2025 , eprint=

    2025

  35. [57]

    Text summarization branches out , pages=

    Rouge: A package for automatic evaluation of summaries , author=. Text summarization branches out , pages=

  36. [58]

    1986 , publisher=

    Relevance: Communication and cognition , author=. 1986 , publisher=

    1986

  37. [59]

    Human language: From genes and brain to behavior , pages=

    Interactional foundations of language: The interaction engine hypothesis , author=. Human language: From genes and brain to behavior , pages=. 2019 , publisher=

    2019

  38. [60]

    Proceedings of ICLR , year=

    Plug-and-Play Policy Planner for Large Language Model Powered Dialogue Agents , author=. Proceedings of ICLR , year=

  39. [61]

    arXiv preprint arXiv:2406.11978 , url=

    Dialogue Action Tokens: Steering Language Models in Goal-Directed Dialogue with a Multi-Turn Planner , author=. arXiv preprint arXiv:2406.11978 , url=

    arXiv

  40. [62]

    arXiv preprint arXiv:2502.15538 , year=

    SOTOPIA- \ Omega \ : Dynamic Strategy Injection Learning and Social Instrucion Following Evaluation for Social Agents , author=. arXiv preprint arXiv:2502.15538 , year=

    arXiv

  41. [63]

    2024 , eprint=

    DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Language Models , author=. 2024 , eprint=

    2024

  42. [64]

    arXiv preprint arXiv:2412.15115 , url=

    Qwen2.5 technical report , author=. arXiv preprint arXiv:2412.15115 , url=

    arXiv

  43. [65]

    OpenAI o3-mini System Card , url=

    OpenAI o3-mini System Card , author=. OpenAI o3-mini System Card , url=

  44. [66]

    Gemini 2.0 flash thinking System Card , url=

    Gemini 2.0 flash thinking , author=. Gemini 2.0 flash thinking System Card , url=

  45. [67]

    arXiv preprint arXiv:2412.16720 , year=

    Openai o1 system card , author=. arXiv preprint arXiv:2412.16720 , year=

    arXiv

  46. [68]

    arXiv preprint arXiv:2501.12948 , url=

    Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning , author=. arXiv preprint arXiv:2501.12948 , url=

    arXiv

  47. [69]

    2024 , url=

    Xuhui Zhou and Hao Zhu and Leena Mathur and Ruohong Zhang and Haofei Yu and Zhengyang Qi and Louis-Philippe Morency and Yonatan Bisk and Daniel Fried and Graham Neubig and Maarten Sap , booktitle=. 2024 , url=

    2024

  48. [70]

    arXiv preprint arXiv:2402.03300 , year=

    Deepseekmath: Pushing the limits of mathematical reasoning in open language models , author=. arXiv preprint arXiv:2402.03300 , year=

    arXiv

  49. [71]

    arXiv preprint arXiv:2409.14195 , url=

    The Imperative of Conversation Analysis in the Era of LLMs: A Survey of Tasks, Techniques, and Trends , author=. arXiv preprint arXiv:2409.14195 , url=

    arXiv

  50. [72]

    arXiv preprint arXiv:2502.12486 , url=

    EPO: Explicit Policy Optimization for Strategic Reasoning in LLMs via Reinforcement Learning , author=. arXiv preprint arXiv:2502.12486 , url=

    arXiv

  51. [73]

    2025 , eprint=

    Debt Collection Negotiations with Large Language Models: An Evaluation System and Optimizing Decision Making with Multi-Agent , author=. 2025 , eprint=

    2025

  52. [74]

    arXiv preprint arXiv:2410.21276 , URL =

    Gpt-4o system card , author=. arXiv preprint arXiv:2410.21276 , URL =

    arXiv

  53. [75]

    Claude-3.5-Sonnet Model Card , url=

    Claude 3.5 Sonnet Model Card Addendum , author=. Claude-3.5-Sonnet Model Card , url=

  54. [76]

    arXiv preprint arXiv:2412.19437 , url=

    Deepseek-v3 technical report , author=. arXiv preprint arXiv:2412.19437 , url=

    arXiv

  55. [77]

    Hugging Face , url=

    Qwq: Reflect deeply on the boundaries of the unknown , author=. Hugging Face , url=

  56. [78]

    Proceedings of NeurIPS , url=

    Judging llm-as-a-judge with mt-bench and chatbot arena , author=. Proceedings of NeurIPS , url=

  57. [79]

    Leave No Document Behind: Benchmarking Long-Context LLM s with Extended Multi-Doc QA

    Wang, Minzheng and Chen, Longze and Cheng, Fu and Liao, Shengyi and Zhang, Xinghua and Wu, Bingli and Yu, Haiyang and Xu, Nan and Zhang, Lei and Luo, Run and Li, Yunshui and Yang, Min and Huang, Fei and Li, Yongbin. Leave No Document Behind: Benchmarking Long-Context LLM s wit...

    2024

  58. [80]

    Proceedings of ACL , url=

    SOTOPIA- : Interactive Learning of Socially Intelligent Language Agents , author=. Proceedings of ACL , url=

  59. [81]

    ACM Computing Surveys (CSUR) , volume=

    Hierarchical reinforcement learning: A comprehensive survey , author=. ACM Computing Surveys (CSUR) , volume=. 2021 , publisher=

    2021

  60. [82]

    2017 , eprint=

    FeUdal Networks for Hierarchical Reinforcement Learning , author=. 2017 , eprint=

    2017

  61. [83]

    Handbook of sport psychology , pages=

    Theory of planned behavior , author=. Handbook of sport psychology , pages=. 2020 , publisher=

    2020

  62. [84]

    Organizational behavior and human decision processes , volume=

    The theory of planned behavior , author=. Organizational behavior and human decision processes , volume=. 1991 , publisher=

    1991

  63. [85]

    British journal of social psychology , volume=

    Efficacy of the theory of planned behaviour: A meta-analytic review , author=. British journal of social psychology , volume=. 2001 , publisher=

    2001

  64. [86]

    , author=

    Does changing behavioral intentions engender behavior change? A meta-analysis of the experimental evidence. , author=. Psychological bulletin , volume=. 2006 , publisher=

    2006

  65. [87]

    The Eleventh International Conference on Learning Representations , year=

    ReAct: Synergizing Reasoning and Acting in Language Models , author=. The Eleventh International Conference on Learning Representations , year=

  66. [88]

    Proceedings of IJCAI , url=

    Large language model based multi-agents: a survey of progress and challenges , author=. Proceedings of IJCAI , url=

  67. [89]

    arXiv preprint arXiv:2412.04905 , url=

    Reframing Dialogue Interaction with Fine-grained Element Modeling , author=. arXiv preprint arXiv:2412.04905 , url=

    arXiv

  68. [90]

    arXiv preprint arXiv:2506.06122 , year=

    Reinforcement Learning Optimization for Large-Scale Learning: An Efficient and User-Friendly Scaling Library , author=. arXiv preprint arXiv:2506.06122 , year=

    arXiv

  69. [91]

    L lama F actory: Unified Efficient Fine-Tuning of 100+ Language Models

    Zheng, Yaowei and Zhang, Richong and Zhang, Junhao and Ye, Yanhan and Luo, Zheyan. L lama F actory: Unified Efficient Fine-Tuning of 100+ Language Models. Proceedings of ACL, System Demonstrations. 2024

    2024

  70. [92]

    language , volume=

    A simplest systematics for the organization of turn-taking for conversation , author=. language , volume=. 1974 , publisher=

    1974

  71. [93]

    Nature reviews neuroscience , volume=

    A framework for studying the neurobiology of value-based decision making , author=. Nature reviews neuroscience , volume=. 2008 , publisher=

    2008

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.