pith. sign in

arxiv: 2606.05557 · v1 · pith:GVZ63FCHnew · submitted 2026-06-04 · 💻 cs.CL

AURA: Intent-Directed Probing for Implicit-Need Surfacing in Situated LLM Agents

Yang Li , Jiaxiang Liu , Jiang Cai , Mingkun Xu This is my paper

Pith reviewed 2026-06-28 02:04 UTC · model grok-4.3

classification 💻 cs.CL
keywords LLM agentsimplicit needstool probingIntentFramegap scoresituated queriesReAct baseline
0
0 comments X

The pith

AURA uses an IntentFrame with gap score to improve implicit-need coverage in situated LLM agents.

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

The paper introduces AURA to address how situated queries often imply more than their literal meaning, such as wanting to know if someone is available. It inserts an inference step that generates an IntentFrame estimating the implicit need and a gap score to manage probing. This results in better coverage of implicit needs compared to standard agents on a dedicated benchmark. The method also trades some accuracy for fewer probes on factual tasks.

Core claim

AURA inserts an inference step between scene perception and tool use that produces an IntentFrame: a structured estimate of the implicit need with a scalar gap score that controls per-query probe budget and tool selection. On the benchmark this yields improved implicit-need coverage over ReAct-style probing.

What carries the argument

The IntentFrame, a structured estimate of the implicit need with a scalar gap score controlling probe budget and tool selection.

If this is right

  • It achieves a 0.07 higher implicit-need coverage on the 100-query benchmark.
  • The improvement reproduces across backbones and is due to gap calibration.
  • It reduces probes by 82 percent on factual lookup while avoiding forbidden tools.
  • Three of the four scenes show individually significant gains.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • This calibration technique might apply to other areas of agent decision making where estimating information value is key.
  • In practice, it could lead to more efficient and privacy-aware interactions in real user scenarios.
  • Future work could test if the gap score can be learned rather than prompted.

Load-bearing premise

The 100-query four-scene benchmark accurately represents real-world implicit needs and the gap score measures probe value without selection bias.

What would settle it

Observing no coverage improvement when testing on queries outside the original four scenes or with a different set of 100 queries would falsify the performance gain claim.

Figures

Figures reproduced from arXiv: 2606.05557 by Jiang Cai, Jiaxiang Liu, Mingkun Xu, Yang Li.

Figure 1
Figure 1. Figure 1: AURA pipeline. Sense receives the user query; Scene and Memory assemble a deterministic context preview. The IntentInferrer (⋆) outputs an IntentFrame whose scalar gap field routes probe budget and tool selection in Explore (Section 3.3). Reason plans over the enriched context; Act and Interact emit the response. The gap g is the control input: downstream stages use a determinis￾tic map B(g) = 0, 1, 2, 3, … view at source ↗
Figure 2
Figure 2. Figure 2: Factual-grounding access Pareto. Left: factual lookup ( [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Per-subcategory mean of IntentFrame.gap vs. probes actually issued by the LLM (5 queries × 3 seeds per marker; explore_max_steps=3). The dashed line is the deterministic ceiling map B(g) from Section 3.3, capped at the configured ceiling 3. Probe count varies 2.8× (0.80–2.20) under identical configuration; all four budget values {0, 1, 2, 3} appear; Pearson r(g, probes)=0.19. The second_order subcategory h… view at source ↗
Figure 4
Figure 4. Figure 4: Per-subcategory implicit-need surfacing on the implicit-intent benchmark (25 queries, 3 seeds, N=15 per cell). AURA Intent improves most on lexically decoupled private-state queries (e.g. availability); on second_order queries (“does X think Y?”) the surface already cues belief re￾trieval and NoIntent ties. ing simply becomes a fixed per-query cost. We therefore measure the effective number of probes issue… view at source ↗
Figure 5
Figure 5. Figure 5: AURATown — 60×60 grid with 20 named locations at their real coordinates (from demo/town/assets/town_map.json), coloured by type (home 5, commerce 6, civic 4, open/park 5). Rings labelled L/Z/C/Y/W mark the five agents at their starting homes per demo/town/agents.py. Pub￾lic state (location, action, nearby agents) is visible in the scene snapshot; private state (availability, emotional_state, unspoken_goal,… view at source ↗
Figure 6
Figure 6. Figure 6: Motivating contrast on “Where is Lin Wei?”. Vanilla (left) vs. AURA (right) on the same scene state. IntentFrame example ( [PITH_FULL_IMAGE:figures/full_fig_p015_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: The IntentFrame produced for a single user query. Long-context prompt-content sketch ( [PITH_FULL_IMAGE:figures/full_fig_p015_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: AURA versus a long-context approach on AURATown. A long-context model pays the full per-agent public+private+memory cost per query (panel A); AURA’s IntentFrame routes se￾lective probing so the prompt grows with the in￾ferred gap, not the number of agents (panel B). To￾ken estimates are from the AURATown prompt log; FA / latency from Tables 1, 2. Static-Context (the long-context lower-bound in our suite) r… view at source ↗
Figure 10
Figure 10. Figure 10: Per-rater ∆ (AURA − Vanilla) across the 4 dimensions (N=8 raters, 50 scenarios each). Each coloured dot is one rater’s mean ∆ across the 50 scenarios; the grey violin is the pooled ∆ dis￾tribution; the black diamond is the pooled mean. Letters identify raters (M is the one whose ∆ falls below zero on factual accuracy and is the basis for the N=7 sensitivity analysis). 7/8 raters’ means lie above zero on e… view at source ↗
Figure 9
Figure 9. Figure 9: SOTOPIA 7-dimension social evalua￾tion (AURA full, 200-step run). Each axis is nor￾malised to [0, 1] of its native range (shown next to each label); raw values annotated. Strongest dimensions: believability (9.0/10), goal (9.5/10); weakest: social rules (−2.13) and secret (−0.67) (both negative-only scales). The shape is lopsided￾positive: AURA agents excel on individual-level dimensions and underperform o… view at source ↗
Figure 11
Figure 11. Figure 11: Probe-budget sweep, 3 seeds, 500 GA judgments per (B,seed) cell. Pareto frontier: {B=0, B=1} (red rings); B⋆=1 adds +0.012 GA at +4.1 s, every B ≥ 2 is dominated. The same single probe buys +13.7 pp implicit score in the social sub-regime ( [PITH_FULL_IMAGE:figures/full_fig_p024_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Cross-domain results (all single-seed; sample sizes per panel vary, see text). Bar height is the raw metric in each domain’s na￾tive units (accuracy, payoff/round, partial-success score). AURA full beats Vanilla on Puzzle (+5 pp absolute, n=20 episodes), ties on Math (n=20, judge-saturated), and underperforms on Trust (n=6 games/cond. — underpowered, de￾scriptive only), MemoryArena (n=1 paper × 5 sub￾task… view at source ↗
read the original abstract

A situated query like "where is Lin Wei?" often encodes more than its literal content: the user may also want to know whether Lin Wei is free, in a good mood, or worth interrupting now. Standard tool-use agents answer the literal question and stop. AURA inserts an inference step between scene perception and tool use that produces an IntentFrame: a structured estimate of the implicit need with a scalar gap score that controls per-query probe budget and tool selection. On a 100-query four-scene implicit-intent benchmark, AURA improves implicit-need coverage over ReAct-style probing (Delta = +0.07, p < 10^-6); three of four scenes are individually significant, the gain reproduces on a second backbone, and a prompt ablation attributes the lift to gap calibration rather than answer memorisation. On factual lookup the controller trades raw accuracy for 82% fewer probes and zero forbidden-tool violations on a privacy-sensitive slice; scope conditions are detailed in Limitations. Code, simulator, and benchmark are released at https://github.com/innovation64/AURA.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 1 minor

Summary. The manuscript introduces AURA, an augmentation for situated LLM agents that inserts an IntentFrame inference step after scene perception. The IntentFrame yields a structured implicit-need estimate together with a scalar gap score that governs per-query probe budget and tool selection. On a custom 100-query four-scene benchmark, AURA reports a +0.07 gain in implicit-need coverage over ReAct-style probing (p < 10^{-6}), with three scenes individually significant; the result reproduces on a second backbone and an ablation attributes the improvement to gap calibration rather than memorization. On factual lookup the controller reduces probes by 82% while incurring zero forbidden-tool violations on a privacy-sensitive slice. Code, simulator, and benchmark are released.

Significance. If the benchmark construction proves free of selection bias and the gap score is shown to be independent of the evaluation data, the approach would supply a concrete, controllable mechanism for surfacing implicit needs without excessive tool calls. The public release of the full experimental artifacts is a clear methodological strength that directly addresses reproducibility concerns.

major comments (2)
  1. [experimental evaluation section] Benchmark construction (experimental evaluation section): The manuscript supplies no description of how the 100 queries were sampled, how the four scenes and their implicit-need labels were generated, or whether any aspect of the IntentFrame logic influenced scene or query design. Because the benchmark is entirely author-constructed and the headline delta rests on coverage measured against these labels, the absence of construction details leaves open the possibility that the reported gain (+0.07) and the ablation attribution to gap calibration are partly artifacts of benchmark-specific tuning.
  2. [ablation and statistical reporting] Hyperparameter independence (ablation and statistical reporting): The abstract states that the gap score controls probe budget, yet provides no information on whether the gap threshold or related parameters were tuned or validated on the same 100-query set used for the final coverage and significance calculations. If any such overlap exists, the p < 10^{-6} result and the claim that the lift is due to calibration rather than memorization cannot be treated as independent evidence.
minor comments (1)
  1. [abstract] The Limitations section is referenced for scope conditions but is not quoted or summarized in the abstract; expanding the abstract's mention of scope conditions would improve standalone readability.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive comments regarding benchmark construction and hyperparameter independence. We address each point below and will revise the manuscript to provide the requested details and clarifications.

read point-by-point responses

  1. Referee: [experimental evaluation section] Benchmark construction (experimental evaluation section): The manuscript supplies no description of how the 100 queries were sampled, how the four scenes and their implicit-need labels were generated, or whether any aspect of the IntentFrame logic influenced scene or query design. Because the benchmark is entirely author-constructed and the headline delta rests on coverage measured against these labels, the absence of construction details leaves open the possibility that the reported gain (+0.07) and the ablation attribution to gap calibration are partly artifacts of benchmark-specific tuning.

    Authors: We agree the manuscript provides insufficient detail on benchmark construction. In revision we will add a dedicated subsection under Experimental Evaluation that describes the query sampling procedure (stratified random sampling across four scene templates), the process for generating scenes and implicit-need labels (via independent human annotation with inter-annotator agreement reported), and an explicit statement that IntentFrame logic played no role in scene or query design. The released benchmark repository already contains the full generation scripts and label files, enabling direct verification. revision: yes

  2. Referee: [ablation and statistical reporting] Hyperparameter independence (ablation and statistical reporting): The abstract states that the gap score controls probe budget, yet provides no information on whether the gap threshold or related parameters were tuned or validated on the same 100-query set used for the final coverage and significance calculations. If any such overlap exists, the p < 10^{-6} result and the claim that the lift is due to calibration rather than memorization cannot be treated as independent evidence.

    Authors: We acknowledge the manuscript does not explicitly address hyperparameter provenance. In the revised ablation section we will state that the gap threshold and controller parameters were fixed using a disjoint 20-query development set collected prior to the evaluation benchmark; no tuning or cross-validation occurred on the 100-query test set. This separation preserves independence of the reported p-value and ablation results. The development-set construction details and parameter-selection protocol will be added for full transparency. revision: yes

Circularity Check

0 steps flagged

No circularity in derivation or evaluation chain

full rationale

The provided abstract and context describe AURA's IntentFrame and gap-score mechanism, followed by an empirical delta on an author-constructed 100-query benchmark against an external ReAct baseline. No equations, fitted parameters, self-citations, or uniqueness theorems are quoted that would reduce the reported coverage gain or ablation result to a quantity defined by the method's own inputs or prior author work. The benchmark functions as an independent test set rather than a self-referential construction, satisfying the criteria for a self-contained evaluation with no load-bearing circular steps.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract-only review supplies no explicit free parameters, axioms, or invented entities beyond the high-level description of IntentFrame and gap score; therefore the ledger is empty and the rationale records the information limit.

pith-pipeline@v0.9.1-grok · 5725 in / 1209 out tokens · 24236 ms · 2026-06-28T02:04:13.087920+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

62 extracted references · 16 canonical work pages

  1. [1]

    O'Brien, Carrie J

    Park, Joon Sung and O'Brien, Joseph and Cai, Carrie Jun and Morris, Meredith Ringel and Liang, Percy and Bernstein, Michael S. , title =. Proceedings of the 36th Annual ACM Symposium on User Interface Software and Technology , articleno =. 2023 , isbn =. doi:10.1145/3586183.3606763 , abstract =

    work page doi:10.1145/3586183.3606763 2023

  2. [2]

    Patil and Kevin Lin and Sarah Wooders and Joseph E

    Charles Packer and Vivian Fang and Shishir G. Patil and Kevin Lin and Sarah Wooders and Joseph E. Gonzalez , title=. CoRR , volume=. 2023 , cdate=

    2023

  3. [3]

    Transactions on Machine Learning Research , issn=

    Voyager: An Open-Ended Embodied Agent with Large Language Models , author=. Transactions on Machine Learning Research , issn=. 2024 , url=

    2024

  4. [4]

    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=

  5. [5]

    Toolformer: language models can teach themselves to use tools , year =

    Schick, Timo and Dwivedi-Yu, Jane and Dess\'. Toolformer: language models can teach themselves to use tools , year =. Proceedings of the 37th International Conference on Neural Information Processing Systems , articleno =

  6. [6]

    and Le, Quoc V

    Wei, Jason and Wang, Xuezhi and Schuurmans, Dale and Bosma, Maarten and Ichter, Brian and Xia, Fei and Chi, Ed H. and Le, Quoc V. and Zhou, Denny , title =. Proceedings of the 36th International Conference on Neural Information Processing Systems , articleno =. 2022 , isbn =

    2022

  7. [7]

    Proceedings of the 37th International Conference on Neural Information Processing Systems , articleno =

    Shinn, Noah and Cassano, Federico and Gopinath, Ashwin and Narasimhan, Karthik and Yao, Shunyu , title =. Proceedings of the 37th International Conference on Neural Information Processing Systems , articleno =. 2023 , publisher =

    2023

  8. [8]

    Transactions on Machine Learning Research , issn=

    Cognitive Architectures for Language Agents , author=. Transactions on Machine Learning Research , issn=. 2024 , url=

    2024

  9. [9]

    GPTSwarm: language agents as optimizable graphs , year =

    Zhuge, Mingchen and Wang, Wenyi and Kirsch, Louis and Faccio, Francesco and Khizbullin, Dmitrii and Schmidhuber, J\". GPTSwarm: language agents as optimizable graphs , year =. Proceedings of the 41st International Conference on Machine Learning , articleno =

  10. [10]

    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

  11. [11]

    AgentBench: Evaluating

    Xiao Liu and Hao Yu and Hanchen Zhang and Yifan Xu and Xuanyu Lei and Hanyu Lai and Yu Gu and Hangliang Ding and Kaiwen Men and Kejuan Yang and Shudan Zhang and Xiang Deng and Aohan Zeng and Zhengxiao Du and Chenhui Zhang and Sheng Shen and Tianjun Zhang and Yu Su and Huan Sun and Minlie Huang and Yuxiao Dong and Jie Tang , booktitle=. AgentBench: Evaluat...

    2024

  12. [12]

    Science China Information Sciences , volume =

    The Rise and Potential of Large Language Model Based Agents: A Survey , author =. Science China Information Sciences , volume =. 2025 , doi =

    2025

  13. [13]

    Sirui Hong and Mingchen Zhuge and Jonathan Chen and Xiawu Zheng and Yuheng Cheng and Jinlin Wang and Ceyao Zhang and Zili Wang and Steven Ka Shing Yau and Zijuan Lin and Liyang Zhou and Chenyu Ran and Lingfeng Xiao and Chenglin Wu and J. Meta. The Twelfth International Conference on Learning Representations , year=

  14. [14]

    Zhang, Ceyao and Yang, Kaijie and Hu, Siyi and Wang, Zihao and Li, Guanghe and Sun, Yihang and Zhang, Cheng and Zhang, Zhaowei and Liu, Anji and Zhu, Song-Chun and Chang, Xiaojun and Zhang, Junge and Yin, Feng and Liang, Yitao and Yang, Yaodong , title =. Proceedings of the Thirty-Eighth AAAI Conference on Artificial Intelligence and Thirty-Sixth Conferen...

    work page doi:10.1609/aaai.v38i16.29710 2024

  15. [15]

    arXiv preprint arXiv:2312.10997 , volume=

    Retrieval-augmented generation for large language models: A survey , author=. arXiv preprint arXiv:2312.10997 , volume=

    Pith/arXiv arXiv

  16. [16]

    Retrieval-augmented generation for knowledge-intensive NLP tasks , year =

    Lewis, Patrick and Perez, Ethan and Piktus, Aleksandra and Petroni, Fabio and Karpukhin, Vladimir and Goyal, Naman and K\". Retrieval-augmented generation for knowledge-intensive NLP tasks , year =. Proceedings of the 34th International Conference on Neural Information Processing Systems , articleno =

  17. [17]

    arXiv preprint arXiv:2205.00445 , year=

    MRKL Systems: A modular, neuro-symbolic architecture that combines large language models, external knowledge sources and discrete reasoning , author=. arXiv preprint arXiv:2205.00445 , year=

    Pith/arXiv arXiv

  18. [18]

    Frontiers of Computer Science , year=

    A survey on large language model based autonomous agents , author=. Frontiers of Computer Science , year=

  19. [19]

    LLM-powered Autonomous Agents

    Weng, Lilian. LLM-powered Autonomous Agents. lilianweng.github.io. 2023

    2023

  20. [20]

    2023 , url=

    Guohao Li and Hasan Abed Al Kader Hammoud and Hani Itani and Dmitrii Khizbullin and Bernard Ghanem , booktitle=. 2023 , url=

    2023

  21. [21]

    arXiv preprint arXiv:2310.02172 , year=

    Lyfe agents: Generative agents for low-cost real-time social interactions , author=. arXiv preprint arXiv:2310.02172 , year=

    arXiv

  22. [22]

    C hat D ev: Communicative Agents for Software Development

    Qian, Chen and Liu, Wei and Liu, Hongzhang and Chen, Nuo and Dang, Yufan and Li, Jiahao and Yang, Cheng and Chen, Weize and Su, Yusheng and Cong, Xin and Xu, Juyuan and Li, Dahai and Liu, Zhiyuan and Sun, Maosong. C hat D ev: Communicative Agents for Software Development. Proceedings of the 62nd Annual Meeting of the Association for Computational Linguist...

    work page doi:10.18653/v1/2024.acl-long.810 2024

  23. [23]

    The Twelfth International Conference on Learning Representations , year=

    AgentVerse: Facilitating Multi-Agent Collaboration and Exploring Emergent Behaviors , author=. The Twelfth International Conference on Learning Representations , year=

  24. [24]

    Organization of memory , volume=

    Episodic and semantic memory , author=. Organization of memory , volume=. 1972 , publisher=

    1972

  25. [25]

    Psychology of learning and motivation , volume=

    Human memory: A proposed system and its control processes , author=. Psychology of learning and motivation , volume=. 1968 , publisher=

    1968

  26. [26]

    Ebbinghaus, Hermann , year=

  27. [27]

    Zhong, Wanjun and Guo, Lianghong and Gao, Qiqi and Ye, He and Wang, Yanlin , title =. Proceedings of the Thirty-Eighth AAAI Conference on Artificial Intelligence and Thirty-Sixth Conference on Innovative Applications of Artificial Intelligence and Fourteenth Symposium on Educational Advances in Artificial Intelligence , articleno =. 2024 , isbn =. doi:10....

    work page doi:10.1609/aaai.v38i17.29946 2024

  28. [28]

    2014 , publisher=

    The ecological approach to visual perception: classic edition , author=. 2014 , publisher=

    2014

  29. [29]

    Artificial intelligence , volume=

    Intelligence without representation , author=. Artificial intelligence , volume=. 1991 , publisher=

    1991

  30. [30]

    arXiv preprint arXiv:2401.03568 , year=

    Agent ai: Surveying the horizons of multimodal interaction , author=. arXiv preprint arXiv:2401.03568 , year=

    Pith/arXiv arXiv

  31. [31]

    A gent G ym: Evaluating and Training Large Language Model-based Agents across Diverse Environments

    Xi, Zhiheng and Ding, Yiwen and Chen, Wenxiang and Hong, Boyang and Guo, Honglin and Wang, Junzhe and Guo, Xin and Yang, Dingwen and Liao, Chenyang and He, Wei and Gao, Songyang and Chen, Lu and Zheng, Rui and Zou, Yicheng and Gui, Tao and Zhang, Qi and Qiu, Xipeng and Huang, Xuanjing and Wu, Zuxuan and Jiang, Yu-Gang. A gent G ym: Evaluating and Training...

    work page doi:10.18653/v1/2025.acl-long.1355 2025

  32. [32]

    arXiv preprint arXiv:2312.03664 , year=

    Generative agent-based modeling with actions grounded in physical, social, or digital space using Concordia , author=. arXiv preprint arXiv:2312.03664 , year=

    arXiv

  33. [33]

    M ulti A gent B ench : Evaluating the Collaboration and Competition of LLM agents

    Zhu, Kunlun and Du, Hongyi and Hong, Zhaochen and Yang, Xiaocheng and Guo, Shuyi and Wang, Zhe and Wang, Zhenhailong and Qian, Cheng and Tang, Xiangru and Ji, Heng and You, Jiaxuan. M ulti A gent B ench : Evaluating the Collaboration and Competition of LLM agents. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Vol...

    work page doi:10.18653/v1/2025.acl-long.421 2025

  34. [34]

    MA g IC : Investigation of Large Language Model Powered Multi-Agent in Cognition, Adaptability, Rationality and Collaboration

    Xu, Lin and Hu, Zhiyuan and Zhou, Daquan and Ren, Hongyu and Dong, Zhen and Keutzer, Kurt and Ng, See-Kiong and Feng, Jiashi. MA g IC : Investigation of Large Language Model Powered Multi-Agent in Cognition, Adaptability, Rationality and Collaboration. Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing. 2024. doi:10.186...

    work page doi:10.18653/v1/2024.emnlp-main.416 2024

  35. [35]

    T -Eval: Evaluating the Tool Utilization Capability of Large Language Models Step by Step

    Chen, Zehui and Du, Weihua and Zhang, Wenwei and Liu, Kuikun and Liu, Jiangning and Zheng, Miao and Zhuo, Jingming and Zhang, Songyang and Lin, Dahua and Chen, Kai and Zhao, Feng. T -Eval: Evaluating the Tool Utilization Capability of Large Language Models Step by Step. Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistic...

    work page doi:10.18653/v1/2024.acl-long.515 2024

  36. [36]

    Yujia Qin and Shihao Liang and Yining Ye and Kunlun Zhu and Lan Yan and Yaxi Lu and Yankai Lin and Xin Cong and Xiangru Tang and Bill Qian and Sihan Zhao and Lauren Hong and Runchu Tian and Ruobing Xie and Jie Zhou and Mark Gerstein and dahai li and Zhiyuan Liu and Maosong Sun , booktitle=. Tool. 2024 , url=

    2024

  37. [37]

    From Text to Tactic: Evaluating

    Jonathan Light and Min Cai and Sheng Shen and Ziniu Hu , booktitle=. From Text to Tactic: Evaluating. 2023 , url=

    2023

  38. [38]

    2024 , url=

    Jinhao Duan and Renming Zhang and James Diffenderfer and Bhavya Kailkhura and Lichao Sun and Elias Stengel-Eskin and Mohit Bansal and Tianlong Chen and Kaidi Xu , booktitle=. 2024 , url=

    2024

  39. [39]

    S ocial B ench: Sociality Evaluation of Role-Playing Conversational Agents

    Chen, Hongzhan and Chen, Hehong and Yan, Ming and Xu, Wenshen and Xing, Gao and Shen, Weizhou and Quan, Xiaojun and Li, Chenliang and Zhang, Ji and Huang, Fei. S ocial B ench: Sociality Evaluation of Role-Playing Conversational Agents. Findings of the Association for Computational Linguistics: ACL 2024. 2024. doi:10.18653/v1/2024.findings-acl.125

    work page doi:10.18653/v1/2024.findings-acl.125 2024

  40. [40]

    URLhttps://aclanthology.org/2024.tacl-1.9/

    Liu, Nelson F. and Lin, Kevin and Hewitt, John and Paranjape, Ashwin and Bevilacqua, Michele and Petroni, Fabio and Liang, Percy. Lost in the Middle: How Language Models Use Long Contexts. Transactions of the Association for Computational Linguistics. 2024. doi:10.1162/tacl_a_00638

    work page doi:10.1162/tacl_a_00638 2024

  41. [41]

    The Fourteenth International Conference on Learning Representations , year=

    An Information Theoretic Perspective on Agentic System Design , author=. The Fourteenth International Conference on Learning Representations , year=

  42. [42]

    Found in the middle: Calibrating Positional Attention Bias Improves Long Context Utilization

    Hsieh, Cheng-Yu and Chuang, Yung-Sung and Li, Chun-Liang and Wang, Zifeng and Le, Long and Kumar, Abhishek and Glass, James and Ratner, Alexander and Lee, Chen-Yu and Krishna, Ranjay and Pfister, Tomas. Found in the middle: Calibrating Positional Attention Bias Improves Long Context Utilization. Findings of the Association for Computational Linguistics: A...

    work page doi:10.18653/v1/2024.findings-acl.890 2024

  43. [43]

    ContextAgent: Context-Aware Proactive

    Bufang Yang and Lilin Xu and Liekang Zeng and Kaiwei Liu and Siyang Jiang and Wenrui Lu and Hongkai Chen and Xiaofan Jiang and Guoliang Xing and Zhenyu Yan , booktitle=. ContextAgent: Context-Aware Proactive. 2026 , url=

    2026

  44. [44]

    arXiv preprint arXiv:2512.06721 , year=

    ProAgent: Harnessing On-Demand Sensory Contexts for Proactive LLM Agent Systems , author=. arXiv preprint arXiv:2512.06721 , year=

    Pith/arXiv arXiv

  45. [45]

    arXiv preprint arXiv:2506.12508 , year=

    AgentOrchestra: Orchestrating Multi-Agent Intelligence with the Tool-Environment-Agent (TEA) Protocol , author=. arXiv preprint arXiv:2506.12508 , year=

    Pith/arXiv arXiv

  46. [46]

    arXiv e-prints , pages=

    ProAgentBench: Evaluating llm agents for proactive assistance with real-world data , author=. arXiv e-prints , pages=

  47. [47]

    arXiv preprint arXiv:2603.25723 , year=

    Natural-language agent harnesses , author=. arXiv preprint arXiv:2603.25723 , year=

    Pith/arXiv arXiv

  48. [48]

    arXiv preprint arXiv:2510.04618 , year=

    Agentic context engineering: Evolving contexts for self-improving language models , author=. arXiv preprint arXiv:2510.04618 , year=

    Pith/arXiv arXiv

  49. [49]

    arXiv preprint arXiv:2510.19771 , year=

    Beyond reactivity: Measuring proactive problem solving in llm agents , author=. arXiv preprint arXiv:2510.19771 , year=

    arXiv

  50. [50]

    Gonzalez and Ion Stoica , booktitle=

    Lianmin Zheng and Wei-Lin Chiang and Ying Sheng and Siyuan Zhuang and Zhanghao Wu and Yonghao Zhuang and Zi Lin and Zhuohan Li and Dacheng Li and Eric Xing and Hao Zhang and Joseph E. Gonzalez and Ion Stoica , booktitle=. Judging. 2023 , url=

    2023

  51. [51]

    Proactive Agent: Shifting

    Yaxi Lu and Shenzhi Yang and Cheng Qian and Guirong Chen and Qinyu Luo and Yesai Wu and Huadong Wang and Xin Cong and Zhong Zhang and Yankai Lin and Weiwen Liu and Yasheng Wang and Zhiyuan Liu and Fangming Liu and Maosong Sun , booktitle=. Proactive Agent: Shifting. 2025 , url=

    2025

  52. [52]

    arXiv preprint arXiv:2302.02083 , volume=

    Theory of mind may have spontaneously emerged in large language models , author=. arXiv preprint arXiv:2302.02083 , volume=

    arXiv

  53. [53]

    Neural Theory-of-Mind? On the Limits of Social Intelligence in Large LM s

    Sap, Maarten and Le Bras, Ronan and Fried, Daniel and Choi, Yejin. Neural Theory-of-Mind? On the Limits of Social Intelligence in Large LM s. Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing. 2022. doi:10.18653/v1/2022.emnlp-main.248

    work page doi:10.18653/v1/2022.emnlp-main.248 2022

  54. [54]

    Minding Language Models' (Lack of) Theory of Mind: A Plug-and-Play Multi-Character Belief Tracker

    Sclar, Melanie and Kumar, Sachin and West, Peter and Suhr, Alane and Choi, Yejin and Tsvetkov, Yulia. Minding Language Models' (Lack of) Theory of Mind: A Plug-and-Play Multi-Character Belief Tracker. Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2023. doi:10.18653/v1/2023.acl-long.780

    work page doi:10.18653/v1/2023.acl-long.780 2023

  55. [55]

    arXiv preprint arXiv:2302.08399 , year=

    Large language models fail on trivial alterations to theory-of-mind tasks , author=. arXiv preprint arXiv:2302.08399 , year=

    arXiv

  56. [56]

    FANT o M : A Benchmark for Stress-testing Machine Theory of Mind in Interactions

    Kim, Hyunwoo and Sclar, Melanie and Zhou, Xuhui and Bras, Ronan and Kim, Gunhee and Choi, Yejin and Sap, Maarten. FANT o M : A Benchmark for Stress-testing Machine Theory of Mind in Interactions. Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing. 2023. doi:10.18653/v1/2023.emnlp-main.890

    work page doi:10.18653/v1/2023.emnlp-main.890 2023

  57. [57]

    Evaluating Very Long-Term Conversational Memory of LLM Agents

    Maharana, Adyasha and Lee, Dong-Ho and Tulyakov, Sergey and Bansal, Mohit and Barbieri, Francesco and Fang, Yuwei. Evaluating Very Long-Term Conversational Memory of LLM Agents. Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2024. doi:10.18653/v1/2024.acl-long.747

    work page doi:10.18653/v1/2024.acl-long.747 2024

  58. [58]

    Plan-and-Solve Prompting: Improving Zero-Shot Chain-of-Thought Reasoning by Large Language Models

    Wang, Lei and Xu, Wanyu and Lan, Yihuai and Hu, Zhiqiang and Lan, Yunshi and Lee, Roy Ka-Wei and Lim, Ee-Peng. Plan-and-Solve Prompting: Improving Zero-Shot Chain-of-Thought Reasoning by Large Language Models. Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2023. doi:10.18653/v1/2023.acl-long.147

    work page doi:10.18653/v1/2023.acl-long.147 2023

  59. [59]

    arXiv preprint arXiv:2601.20465 , year=

    BMAM: Brain-inspired Multi-Agent Memory Framework , author=. arXiv preprint arXiv:2601.20465 , year=

    arXiv

  60. [60]

    2024 , note=

    Building Effective Agents , author=. 2024 , note=

    2024

  61. [61]

    2026 , month =

    Interaction Models: A Scalable Approach to. 2026 , month =

    2026

  62. [62]

    Joe Qin , booktitle=

    Zongxi Li and Yang Li and Haoran Xie and S. Joe Qin , booktitle=. CondAmbig. 2025 , url=

    2025