pith. sign in

arxiv: 2606.28733 · v1 · pith:VTTEEEYZnew · submitted 2026-06-27 · 💻 cs.AI

Agentic Abstention: Do Agents Know When to Stop Instead of Act?

Han Luo , Bingbing Wen , Lucy Lu Wang This is my paper

Pith reviewed 2026-06-30 09:51 UTC · model grok-4.3

classification 💻 cs.AI
keywords agentic abstentionLLM agentstimely abstentionstopping rulescontext engineeringweb shoppingterminal environmentsquestion answering
0
0 comments X

The pith

LLM agents struggle more with when to abstain than whether they can abstain from further actions.

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

LLM agents operate over multiple turns using tools like search and terminals, yet many user goals are underspecified or impossible to achieve in the given environment. The paper establishes that agents must decide sequentially whether to continue, answer, or abstain, and that the critical issue is the timing of abstention rather than the capacity to recognize uncertainty. Evaluation of thirteen systems across more than twenty-eight thousand tasks in web shopping, terminal, and QA settings shows wide variation, with some agents never abstaining appropriately and others doing so only after many wasteful steps. Larger models and different scaffolds affect timing in inconsistent ways. CONVOLVE improves outcomes by distilling complete trajectories into reusable stopping rules that can be injected into context without any model parameter updates.

Core claim

Agentic abstention is the sequential problem of deciding at each turn to answer, abstain, or gather more information, where the need to stop may become clear only after environmental interaction reveals that no valid result matches the instruction. Current agents exhibit large gaps in timely abstention, especially on tasks that initially appear feasible. Model scale, reasoning capability, and scaffolding influence abstention timing differently, with larger models sometimes performing worse. CONVOLVE addresses this by engineering context through distillation of full interaction trajectories into reusable stopping rules, raising Llama-3.3-70B timely recall rate from 26.7 to 57.4 on WebShop wit

What carries the argument

Agentic Abstention as a multi-turn sequential decision process, carried by CONVOLVE which distills full interaction trajectories into reusable stopping rules for context injection.

If this is right

  • Model scale does not reliably improve timely abstention and can sometimes reduce it.
  • Different agent scaffolds produce distinct patterns of abstention timing.
  • The largest gaps between agents appear on tasks where the environment must be queried to reveal that the instruction cannot be satisfied.
  • CONVOLVE raises timely recall without requiring model parameter updates.

Where Pith is reading between the lines

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

  • Agents equipped with such stopping rules could reduce wasted computation in deployed systems by halting earlier on unachievable goals.
  • The distillation approach could be adapted to other sequential agent decisions such as when to reformulate a query or recover from an error.
  • Testing the same method on environments with continuous rather than discrete state spaces would reveal whether the stopping rules remain effective.

Load-bearing premise

The constructed tasks across the three environments accurately represent real cases where abstention is the correct response and the timely recall rate metric measures desired stopping behavior without bias from task design.

What would settle it

If CONVOLVE produces no increase in timely recall rate when applied to a new collection of tasks whose uncertainty patterns differ from the original web shopping, terminal, and QA sets, the claim of general improvement would be refuted.

Figures

Figures reproduced from arXiv: 2606.28733 by Bingbing Wen, Han Luo, Lucy Lu Wang.

Figure 1
Figure 1. Figure 1: This is an Environment-based Abstention example in a web shopping scenario, where the [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: (a) Each adapted task in TerminalBench 2.0 consists of four core components: (1) a containerized environment initialized with the relevant packages and files, (2) an instruction describing the task to be completed, (3) a set of tests for verifying completion, and (4) a manually written reference solution. For our abstention setting, we rewrite the original instruction to construct abstention-warranted vari… view at source ↗
Figure 3
Figure 3. Figure 3: Abstention is hard for agents, especially timely abstention. Abstention Recall increases with larger K, but early abstention (e.g., AbsRec@1) remains low across settings and systems. This suggests that agents often abstain only after unnecessary interaction, rather than when abstention first becomes warranted. clearly match an existing section are assigned to an “other” section. This keeps the evolving con… view at source ↗
Figure 4
Figure 4. Figure 4: AbsRec@K across Abstention Categories in Web, Terminal, and QA scenarios. From top to bottom, the rows show Web, Terminal, and QA results. Missing Target in Web, Underspecified Intent in Terminal, and False Premise and Underspecified Intent in QA are the most difficult cases across models. Performance for the same model varies significantly across abstention categories. suggests that larger search budgets … view at source ↗
Figure 5
Figure 5. Figure 5: More reasoning leads to some improvement in [PITH_FULL_IMAGE:figures/full_fig_p008_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Cumulative over-abstention rate by turn in Web and [PITH_FULL_IMAGE:figures/full_fig_p009_6.png] view at source ↗
Figure 8
Figure 8. Figure 8: (a) Length distributions show similar token counts between original and rewritten WebShop instructions. (b) t-SNE visualization of instruction embeddings reveals substantial semantic overlap, demonstrating that original and rewritten WebShop instructions are semantically indistinguishable. The similarity breakdown across the three abstention scenarios is shown in [PITH_FULL_IMAGE:figures/full_fig_p018_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: t-SNE visualizations of original and rewritten WebShop instructions by abstention category: [PITH_FULL_IMAGE:figures/full_fig_p018_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Composition of the selected AbstentionBench subset across datasets and scenarios. The [PITH_FULL_IMAGE:figures/full_fig_p022_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Example playbook learned by Llama-3.3-70B. The playbook contains general abstention principles and trajectory-derived guidelines distilled from Llama-3.3-70B interaction trajectories. These entries summarize when the agent should stop searching or clicking, especially when the environment indicates that the request is subjective, underspecified, or unfulfillable. 36 [PITH_FULL_IMAGE:figures/full_fig_p036… view at source ↗
Figure 12
Figure 12. Figure 12: Agent scaffolds matter beyond the base model. With the same base model (GPT￾5.4-mini), Codex CLI consistently achieves higher abstention recall than Terminus 2 across both request-based and environment-based task. 37 [PITH_FULL_IMAGE:figures/full_fig_p037_12.png] view at source ↗
read the original abstract

LLM agents are expected to act over multiple turns, using search, browsing interfaces, and terminal tools to complete user goals. Yet not every goal is well specified or achievable in the available environment. In such cases, a reliable agent should recognize that further interaction is unlikely to help and abstain from additional tool calls. We define Agentic Abstention, the problem of deciding when an agent should stop acting under uncertainty. Unlike standard LLM abstention, which is usually evaluated as a single-turn answer-or-abstain decision, agentic abstention is a sequential decision problem: an agent can answer, abstain, or gather more information at each turn, and the need to abstain may only become clear after interacting with the environment. We study this problem across web shopping, terminal environments, and question answering, evaluating 13 LLM-as-agent systems and 2 agent scaffolds on more than 28,000 tasks. Our results show that the main challenge is not only whether agents can abstain, but also when they abstain. Some agents never abstain when they should, while others do so only after many unnecessary interactions. This gap is especially large on tasks where the instruction appears feasible until the environment reveals otherwise (e.g., no valid result matches the instruction). We further find that model scale, reasoning, and agent scaffolding affect abstention in different ways, where larger or more capable models sometimes perform worse at timely abstention. Finally, we introduce CONVOLVE, a context engineering method for improving agentic abstention that distills full interaction trajectories into reusable stopping rules. On WebShop, CONVOLVE substantially improves timely abstention without updating model parameters, raising Llama-3.3-70B's timely recall rate from 26.7 to 57.4. Our dataset and code are available at https://lhannnn.github.io/agentic-abstention

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

3 major / 3 minor

Summary. The paper defines 'agentic abstention' as a sequential decision problem in which LLM agents must decide at each turn whether to act, gather information, or abstain when a user goal is ill-specified or unachievable in environments such as WebShop, terminal shells, and QA. It evaluates 13 LLM-as-agent systems and 2 scaffolds on more than 28,000 tasks, reports that agents differ markedly in the timing of abstention (some never abstain when warranted, others only after many turns), shows that scale and scaffolding affect timely abstention in non-monotonic ways, and introduces CONVOLVE, a context-engineering method that distills interaction trajectories into reusable stopping rules. On WebShop, CONVOLVE raises Llama-3.3-70B's timely recall rate from 26.7 to 57.4 without parameter updates; the dataset and code are released.

Significance. If the task labeling and 'timely recall rate' metric are shown to be robust, the work identifies a practically important gap in agent reliability and supplies a parameter-free, reproducible improvement technique together with a large public benchmark. The release of the full dataset and code is a clear strength that enables direct verification and extension.

major comments (3)
  1. [§3, §4.1] §3 (Problem Definition) and §4.1 (Task Construction): the criteria used to label the 28k tasks as requiring abstention (e.g., 'no valid result matches the instruction' in WebShop) are described at a high level; without an explicit decision procedure, inter-annotator agreement, or pre-registration of the ambiguity-injection process, it is impossible to rule out that the baseline gap and the CONVOLVE improvement are partly artifacts of how the evaluation tasks were constructed.
  2. [§4.2] §4.2 (Evaluation Metrics): the 'timely recall rate' is introduced as the headline metric yet no closed-form definition, weighting between abstention accuracy and interaction length, or statistical controls for task difficulty are supplied; the reported jump from 26.7 to 57.4 therefore cannot be assessed for sensitivity to the precise operationalization of 'timely'.
  3. [§5] §5 (CONVOLVE Experiments): the claim that CONVOLVE improves abstention 'without updating model parameters' is load-bearing, but the paper does not report whether the distilled stopping rules were tuned on a held-out portion of the same 28k tasks or whether the improvement holds under a strict train/test split of the newly constructed environments.
minor comments (3)
  1. The abstract states results for 'Llama-3.3-70B' while the main text occasionally uses 'Llama-3-70B'; consistent naming and version numbers should be used throughout.
  2. Table captions and axis labels for the interaction-length histograms are not described in sufficient detail to interpret the 'timely' dimension visually.
  3. A short related-work subsection contrasting agentic abstention with single-turn LLM abstention and with existing 'know-when-to-stop' literature in planning would help readers situate the contribution.

Simulated Author's Rebuttal

3 responses · 0 unresolved

Thank you for the detailed referee report. We address each major comment point by point below, indicating where revisions will be made to the manuscript.

read point-by-point responses

  1. Referee: [§3, §4.1] §3 (Problem Definition) and §4.1 (Task Construction): the criteria used to label the 28k tasks as requiring abstention (e.g., 'no valid result matches the instruction' in WebShop) are described at a high level; without an explicit decision procedure, inter-annotator agreement, or pre-registration of the ambiguity-injection process, it is impossible to rule out that the baseline gap and the CONVOLVE improvement are partly artifacts of how the evaluation tasks were constructed.

    Authors: We agree that more explicit detail on labeling is warranted. Task labels are derived from deterministic environment feedback (e.g., WebShop search returning zero matching products, or terminal execution producing no valid output). We will revise §4.1 to include pseudocode and an explicit decision procedure for each environment. Because labeling relies on programmatic environment signals rather than subjective human judgment, traditional inter-annotator agreement does not apply; we will nevertheless document the manual verification steps used for quality assurance. These additions should reduce concerns that results are artifacts of task construction. revision: yes

  2. Referee: [§4.2] §4.2 (Evaluation Metrics): the 'timely recall rate' is introduced as the headline metric yet no closed-form definition, weighting between abstention accuracy and interaction length, or statistical controls for task difficulty are supplied; the reported jump from 26.7 to 57.4 therefore cannot be assessed for sensitivity to the precise operationalization of 'timely'.

    Authors: We will insert a closed-form definition of timely recall rate in the revised §4.2, explicitly stating the weighting between abstention correctness and interaction length. We will also add sensitivity analyses and statistical controls, including performance stratified by task difficulty (e.g., number of required steps or ambiguity level). These changes will allow readers to evaluate robustness of the reported improvement from 26.7 to 57.4. revision: yes

  3. Referee: [§5] §5 (CONVOLVE Experiments): the claim that CONVOLVE improves abstention 'without updating model parameters' is load-bearing, but the paper does not report whether the distilled stopping rules were tuned on a held-out portion of the same 28k tasks or whether the improvement holds under a strict train/test split of the newly constructed environments.

    Authors: CONVOLVE generates stopping rules from interaction trajectories without any gradient updates to LLM parameters; that is the intended meaning of the claim. The current experiments distill rules from the full set of trajectories. We will revise §5 to state this explicitly and treat the absence of a held-out split as a limitation. We will also release the distilled rules alongside the dataset so that independent verification on held-out tasks is possible, and we commit to reporting results under a strict split in an appendix or follow-up if feasible. revision: partial

Circularity Check

0 steps flagged

No circularity: empirical evaluation on constructed tasks with direct measurements

full rationale

The paper defines Agentic Abstention as a sequential decision problem, constructs 28k tasks across environments, evaluates 13 LLM agents plus scaffolds, and introduces CONVOLVE as a context-engineering method that distills trajectories into stopping rules. All reported results (e.g., timely recall rates, comparisons across model scales) are direct empirical measurements on these tasks rather than predictions derived from equations, fitted parameters renamed as forecasts, or load-bearing self-citations. No derivation chain exists that reduces outputs to inputs by construction; the work is self-contained as an experimental study whose validity rests on task design and metric definitions, not on internal reduction.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 1 invented entities

The work is an empirical study that defines a new problem and evaluates existing systems plus one new method; the abstract mentions no explicit free parameters, mathematical axioms, or newly postulated physical entities.

invented entities (1)
  • Agentic Abstention no independent evidence
    purpose: To name and frame the sequential abstention decision problem for LLM agents
    Newly defined in the paper as distinct from standard single-turn abstention.

pith-pipeline@v0.9.1-grok · 5874 in / 1210 out tokens · 37531 ms · 2026-06-30T09:51:30.657686+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

115 extracted references · 23 canonical work pages · 13 internal anchors

  1. [1]

    Mind2web: Towards a generalist agent for the web.Advances in Neural Information Processing Systems, 36:28091–28114, 2023

    Xiang Deng, Yu Gu, Boyuan Zheng, Shijie Chen, Sam Stevens, Boshi Wang, Huan Sun, and Yu Su. Mind2web: Towards a generalist agent for the web.Advances in Neural Information Processing Systems, 36:28091–28114, 2023

    2023

  2. [2]

    Webshop: Towards scalable real-world web interaction with grounded language agents.Advances in Neural Information Processing Systems, 35:20744–20757, 2022

    Shunyu Yao, Howard Chen, John Yang, and Karthik Narasimhan. Webshop: Towards scalable real-world web interaction with grounded language agents.Advances in Neural Information Processing Systems, 35:20744–20757, 2022

    2022

  3. [3]

    Susbench: An online benchmark for evaluating dark pattern susceptibility of computer-use agents

    Longjie Guo, Chenjie Yuan, Mingyuan Zhong, Robert Wolfe, Ruican Zhong, Yue Xu, Bingbing Wen, Hua Shen, Lucy Lu Wang, and Alexis Hiniker. Susbench: An online benchmark for evaluating dark pattern susceptibility of computer-use agents. InProceedings of the 31st International Conference on Intelligent User Interfaces, pages 1917–1937, 2026

    1917

  4. [4]

    Terminal-Bench: Benchmarking Agents on Hard, Realistic Tasks in Command Line Interfaces

    Mike A Merrill, Alexander G Shaw, Nicholas Carlini, Boxuan Li, Harsh Raj, Ivan Bercovich, Lin Shi, Jeong Yeon Shin, Thomas Walshe, E Kelly Buchanan, et al. Terminal-bench: Benchmarking agents on hard, realistic tasks in command line interfaces.arXiv preprint arXiv:2601.11868, 2026

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  5. [5]

    Patil, Ion Stoica, and Joseph E

    Charles Packer, Sarah Wooders, Kevin Lin, Vivian Fang, Shishir G. Patil, Ion Stoica, and Joseph E. Gonzalez. Memgpt: Towards llms as operating systems, 2024

    2024

  6. [6]

    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. InThe eleventh international conference on learning representations, 2022

    2022

  7. [7]

    Voyager: An Open-Ended Embodied Agent with Large Language Models

    Guanzhi Wang, Yuqi Xie, Yunfan Jiang, Ajay Mandlekar, Chaowei Xiao, Yuke Zhu, Linxi Fan, and Anima Anandkumar. V oyager: An open-ended embodied agent with large language models. arXiv preprint arXiv:2305.16291, 2023

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  8. [8]

    Uncertainty Quantification in LLM Agents: Foundations, Emerging Challenges, and Opportunities

    Changdae Oh, Seongheon Park, To Eun Kim, Jiatong Li, Wendi Li, Samuel Yeh, Xuefeng Du, Hamed Hassani, Paul Bogdan, Dawn Song, et al. Uncertainty quantification in llm agents: Foundations, emerging challenges, and opportunities.arXiv preprint arXiv:2602.05073, 2026

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  9. [9]

    Position: Uncertainty quantification needs reassessment for large-language model agents, 2025

    Michael Kirchhof, Gjergji Kasneci, and Enkelejda Kasneci. Position: Uncertainty quantification needs reassessment for large-language model agents.arXiv preprint arXiv:2505.22655, 2025

    work page arXiv 2025

  10. [10]

    Smart: Self-aware agent for tool overuse mitigation

    Cheng Qian, Emre Can Acikgoz, Hongru Wang, Xiusi Chen, Avirup Sil, Dilek Hakkani-Tur, Gokhan Tur, and Heng Ji. Smart: Self-aware agent for tool overuse mitigation. InFindings of the Association for Computational Linguistics: ACL 2025, pages 4604–4621, 2025

    2025

  11. [11]

    Over-searching in search-augmented large language models, 2026

    Roy Xie, Deepak Gopinath, David Qiu, Dong Lin, Haitian Sun, Saloni Potdar, and Bhuwan Dhingra. Over-searching in search-augmented large language models, 2026

    2026

  12. [12]

    Check Yourself Before You Wreck Yourself: Selectively Quitting Improves LLM Agent Safety

    Vamshi Krishna Bonagiri, Ponnurangam Kumaragurum, Khanh Nguyen, and Benjamin Plaut. Check yourself before you wreck yourself: Selectively quitting improves llm agent safety.arXiv preprint arXiv:2510.16492, 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  13. [13]

    Clarify when necessary: Resolving ambiguity through interaction with lms

    Michael JQ Zhang and Eunsol Choi. Clarify when necessary: Resolving ambiguity through interaction with lms. InFindings of the Association for Computational Linguistics: NAACL 2025, pages 5526–5543, 2025

    2025

  14. [14]

    Clamber: A benchmark of identifying and clarifying ambiguous information needs in large language models

    Tong Zhang, Peixin Qin, Yang Deng, Chen Huang, Wenqiang Lei, Junhong Liu, Dingnan Jin, Hongru Liang, and Tat-Seng Chua. Clamber: A benchmark of identifying and clarifying ambiguous information needs in large language models. InProceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 10746–10766, 2024

    2024

  15. [15]

    Active task disambiguation with llms.arXiv preprint arXiv:2502.04485, 2025

    Katarzyna Kobalczyk, Nicolas Astorga, Tennison Liu, and Mihaela van der Schaar. Active task disambiguation with llms.arXiv preprint arXiv:2502.04485, 2025

    work page arXiv 2025

  16. [16]

    Clarify or answer: Reinforcement learning for agentic vqa with context under-specification.arXiv preprint arXiv:2601.16400, 2026

    Zongwan Cao, Bingbing Wen, and Lucy Lu Wang. Clarify or answer: Reinforcement learning for agentic vqa with context under-specification.arXiv preprint arXiv:2601.16400, 2026. 10

    work page arXiv 2026

  17. [17]

    Know your limits: A survey of abstention in large language models.Transactions of the Association for Computational Linguistics, 13:529–556, 2025

    Bingbing Wen, Jihan Yao, Shangbin Feng, Chenjun Xu, Yulia Tsvetkov, Bill Howe, and Lucy Lu Wang. Know your limits: A survey of abstention in large language models.Transactions of the Association for Computational Linguistics, 13:529–556, 2025

    2025

  18. [18]

    Don’t hallucinate, abstain: Identifying llm knowledge gaps via multi-llm collaboration

    Shangbin Feng, Weijia Shi, Yike Wang, Wenxuan Ding, Vidhisha Balachandran, and Yulia Tsvetkov. Don’t hallucinate, abstain: Identifying llm knowledge gaps via multi-llm collaboration. InProceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 14664–14690, 2024

    2024

  19. [19]

    Abstentionbench: Reasoning llms fail on unanswerable questions.arXiv preprint arXiv:2506.09038, 2025

    Polina Kirichenko, Mark Ibrahim, Kamalika Chaudhuri, and Samuel J Bell. Abstentionbench: Reasoning llms fail on unanswerable questions.arXiv preprint arXiv:2506.09038, 2025

    work page arXiv 2025

  20. [20]

    The art of saying no: Contextual noncompliance in language models.Advances in Neural Information Processing Systems, 37:49706–49748, 2024

    Faeze Brahman, Sachin Kumar, Vidhisha Balachandran, Pradeep Dasigi, Valentina Pyatkin, Abhilasha Ravichander, Sarah Wiegreffe, Nouha Dziri, Khyathi Chandu, Jack Hessel, et al. The art of saying no: Contextual noncompliance in language models.Advances in Neural Information Processing Systems, 37:49706–49748, 2024

    2024

  21. [21]

    AgentBench: Evaluating LLMs as Agents

    Xiao Liu, Hao Yu, Hanchen Zhang, Yifan Xu, Xuanyu Lei, Hanyu Lai, Yu Gu, Hangliang Ding, Kaiwen Men, Kejuan Yang, et al. Agentbench: Evaluating llms as agents.arXiv preprint arXiv:2308.03688, 2023

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  22. [22]

    Gpt-5 mini.https://platform.openai.com, 2025

    OpenAI. Gpt-5 mini.https://platform.openai.com, 2025. Accessed: 2026-03-25

    2025

  23. [23]

    Grok 4.1 model card

    xAI. Grok 4.1 model card. Technical report, xAI, 2025

    2025

  24. [24]

    Llama 3.3 70b instruct, 2024

    Meta. Llama 3.3 70b instruct, 2024

    2024

  25. [25]

    gpt-oss-120b & gpt-oss-20b Model Card

    Sandhini Agarwal, Lama Ahmad, Jason Ai, Sam Altman, Andy Applebaum, Edwin Arbus, Rahul K Arora, Yu Bai, Bowen Baker, Haiming Bao, et al. gpt-oss-120b & gpt-oss-20b model card.arXiv preprint arXiv:2508.10925, 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  26. [26]

    Minimax m2.5

    MiniMax. Minimax m2.5. https://www.minimax.io/models/text, 2026. Official model page

    2026

  27. [27]

    Qwen3 technical report, 2025

    Qwen Team. Qwen3 technical report, 2025

    2025

  28. [28]

    Gemma 4 31b it

    Google DeepMind. Gemma 4 31b it. https://huggingface.co/google/ gemma-4-31B-it , 2026. Hugging Face model card for the instruction-tuned Gemma 4 31B checkpoint. Accessed: 2026-04-20

    2026

  29. [29]

    Glm-5: from vibe coding to agentic engineering, 2026

    GLM-5-Team, :, Aohan Zeng, Xin Lv, Zhenyu Hou, Zhengxiao Du, Qinkai Zheng, Bin Chen, Da Yin, Chendi Ge, Chenghua Huang, Chengxing Xie, Chenzheng Zhu, Congfeng Yin, Cunx- iang Wang, Gengzheng Pan, Hao Zeng, Haoke Zhang, Haoran Wang, Huilong Chen, Jiajie Zhang, Jian Jiao, Jiaqi Guo, Jingsen Wang, Jingzhao Du, Jinzhu Wu, Kedong Wang, Lei Li, Lin Fan, Lucen Z...

    2026

  30. [30]

    On Evaluation of Embodied Navigation Agents

    Peter Anderson, Angel Chang, Devendra Singh Chaplot, Alexey Dosovitskiy, Saurabh Gupta, Vladlen Koltun, Jana Kosecka, Jitendra Malik, Roozbeh Mottaghi, Manolis Savva, et al. On evaluation of embodied navigation agents.arXiv preprint arXiv:1807.06757, 2018

    work page internal anchor Pith review Pith/arXiv arXiv 2018

  31. [31]

    Agentic Context Engineering: Evolving Contexts for Self-Improving Language Models

    Qizheng Zhang, Changran Hu, Shubhangi Upasani, Boyuan Ma, Fenglu Hong, Vamsidhar Ka- manuru, Jay Rainton, Chen Wu, Mengmeng Ji, Hanchen Li, et al. Agentic context engineering: Evolving contexts for self-improving language models.arXiv preprint arXiv:2510.04618, 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  32. [32]

    Marvel: Modular abstention for reliable and versatile expert llms

    Bingbing Wen, Faeze Brahman, Zhan Su, Shangbin Feng, Yulia Tsvetkov, Lucy Lu Wang, and Bill Howe. Marvel: Modular abstention for reliable and versatile expert llms. InICML 2025 Workshop on Reliable and Responsible Foundation Models

    2025

  33. [33]

    Characterizing llm abstention behavior in science qa with context perturbations

    Bingbing Wen, Bill Howe, and Lucy Lu Wang. Characterizing llm abstention behavior in science qa with context perturbations. InFindings of the Association for Computational Linguistics: EMNLP 2024, pages 3437–3450, 2024

    2024

  34. [34]

    Natural questions: a benchmark for question answering research.Transactions of the Association for Computational Linguistics, 7:453–466, 2019

    Tom Kwiatkowski, Jennimaria Palomaki, Olivia Redfield, Michael Collins, Ankur Parikh, Chris Alberti, Danielle Epstein, Illia Polosukhin, Jacob Devlin, Kenton Lee, et al. Natural questions: a benchmark for question answering research.Transactions of the Association for Computational Linguistics, 7:453–466, 2019

    2019

  35. [35]

    Musique: Multihop questions via single-hop question composition.Transactions of the Association for Computational Linguistics, 10:539–554, 2022

    Harsh Trivedi, Niranjan Balasubramanian, Tushar Khot, and Ashish Sabharwal. Musique: Multihop questions via single-hop question composition.Transactions of the Association for Computational Linguistics, 10:539–554, 2022

    2022

  36. [36]

    Coqa: A conversational question answering challenge.Transactions of the Association for Computational Linguistics, 7:249–266, 2019

    Siva Reddy, Danqi Chen, and Christopher D Manning. Coqa: A conversational question answering challenge.Transactions of the Association for Computational Linguistics, 7:249–266, 2019

    2019

  37. [37]

    Quac: Question answering in context

    Eunsol Choi, He He, Mohit Iyyer, Mark Yatskar, Wen-tau Yih, Yejin Choi, Percy Liang, and Luke Zettlemoyer. Quac: Question answering in context. InProceedings of the 2018 conference on empirical methods in natural language processing, pages 2174–2184, 2018

    2018

  38. [38]

    Ambigqa: Answering ambiguous open-domain questions

    Sewon Min, Julian Michael, Hannaneh Hajishirzi, and Luke Zettlemoyer. Ambigqa: Answering ambiguous open-domain questions. InProceedings of the 2020 conference on empirical methods in natural language processing (EMNLP), pages 5783–5797, 2020

    2020

  39. [39]

    Situatedqa: Incorporating extra-linguistic contexts into qa

    Michael Zhang and Eunsol Choi. Situatedqa: Incorporating extra-linguistic contexts into qa. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 7371–7387, 2021

    2021

  40. [40]

    Do large language models know what they don’t know? InFindings of the association for Computational Linguistics: ACL 2023, pages 8653–8665, 2023

    Zhangyue Yin, Qiushi Sun, Qipeng Guo, Jiawen Wu, Xipeng Qiu, and Xuan-Jing Huang. Do large language models know what they don’t know? InFindings of the association for Computational Linguistics: ACL 2023, pages 8653–8665, 2023

    2023

  41. [41]

    Knowledge of knowledge: Exploring known-unknowns uncertainty with large language models

    Alfonso Amayuelas, Kyle Wong, Liangming Pan, Wenhu Chen, and William Yang Wang. Knowledge of knowledge: Exploring known-unknowns uncertainty with large language models. InFindings of the Association for Computational Linguistics: ACL 2024, pages 6416–6432, 2024

    2024

  42. [42]

    Pubmedqa: A dataset for biomedical research question answering

    Qiao Jin, Bhuwan Dhingra, Zhengping Liu, William Cohen, and Xinghua Lu. Pubmedqa: A dataset for biomedical research question answering. InProceedings of the 2019 conference on empirical methods in natural language processing and the 9th international joint conference on natural language processing (EMNLP-IJCNLP), pages 2567–2577, 2019

    2019

  43. [43]

    A dataset of information-seeking questions and answers anchored in research papers

    Pradeep Dasigi, Kyle Lo, Iz Beltagy, Arman Cohan, Noah A Smith, and Matt Gardner. A dataset of information-seeking questions and answers anchored in research papers. InProceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 4599–4610, 2021. 12

    2021

  44. [44]

    Realtime qa: What’s the answer right now? Advances in neural information processing systems, 36:49025–49043, 2023

    Jungo Kasai, Keisuke Sakaguchi, Ronan Le Bras, Akari Asai, Xinyan Yu, Dragomir Radev, Noah A Smith, Yejin Choi, Kentaro Inui, et al. Realtime qa: What’s the answer right now? Advances in neural information processing systems, 36:49025–49043, 2023

    2023

  45. [45]

    Alignment for honesty.Advances in Neural Information Processing Systems, 37:63565–63598, 2024

    Yuqing Yang, Ethan Chern, Xipeng Qiu, Graham Neubig, and Pengfei Liu. Alignment for honesty.Advances in Neural Information Processing Systems, 37:63565–63598, 2024

    2024

  46. [46]

    When not to trust language models: Investigating effectiveness of parametric and non-parametric memories

    Alex Mallen, Akari Asai, Victor Zhong, Rajarshi Das, Daniel Khashabi, and Hannaneh Ha- jishirzi. When not to trust language models: Investigating effectiveness of parametric and non-parametric memories. InProceedings of the 61st annual meeting of the association for computational linguistics (volume 1: Long papers), pages 9802–9822, 2023

    2023

  47. [47]

    Simple entity-centric questions challenge dense retrievers

    Christopher Sciavolino, Zexuan Zhong, Jinhyuk Lee, and Danqi Chen. Simple entity-centric questions challenge dense retrievers. InProceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 6138–6148, 2021

    2021

  48. [48]

    Wildguard: Open one-stop moderation tools for safety risks, jailbreaks, and refusals of llms.Advances in neural information processing systems, 37:8093–8131, 2024

    Seungju Han, Kavel Rao, Allyson Ettinger, Liwei Jiang, Bill Yuchen Lin, Nathan Lambert, Yejin Choi, and Nouha Dziri. Wildguard: Open one-stop moderation tools for safety risks, jailbreaks, and refusals of llms.Advances in neural information processing systems, 37:8093–8131, 2024

    2024

  49. [49]

    Realtox- icityprompts: Evaluating neural toxic degeneration in language models

    Samuel Gehman, Suchin Gururangan, Maarten Sap, Yejin Choi, and Noah A Smith. Realtox- icityprompts: Evaluating neural toxic degeneration in language models. InFindings of the association for computational linguistics: EMNLP 2020, pages 3356–3369, 2020

    2020

  50. [50]

    Toxigen: A large-scale machine-generated dataset for adversarial and implicit hate speech detection

    Thomas Hartvigsen, Saadia Gabriel, Hamid Palangi, Maarten Sap, Dipankar Ray, and Ece Kamar. Toxigen: A large-scale machine-generated dataset for adversarial and implicit hate speech detection. InProceedings of the 60th annual meeting of the association for computational linguistics (volume 1: Long papers), pages 3309–3326, 2022

    2022

  51. [51]

    Latent hatred: A benchmark for understanding implicit hate speech

    Mai ElSherief, Caleb Ziems, David Muchlinski, Vaishnavi Anupindi, Jordyn Seybolt, Munmun De Choudhury, and Diyi Yang. Latent hatred: A benchmark for understanding implicit hate speech. InProceedings of the 2021 conference on empirical methods in natural language processing, pages 345–363, 2021

    2021

  52. [52]

    Toxicchat: Unveiling hidden challenges of toxicity detection in real-world user-ai conversation

    Zi Lin, Zihan Wang, Yongqi Tong, Yangkun Wang, Yuxin Guo, Yujia Wang, and Jingbo Shang. Toxicchat: Unveiling hidden challenges of toxicity detection in real-world user-ai conversation. InFindings of the Association for Computational Linguistics: EMNLP 2023, pages 4694–4702, 2023

    2023

  53. [53]

    Beavertails: Towards improved safety alignment of llm via a human-preference dataset.Advances in Neural Information Processing Systems, 36:24678–24704, 2023

    Jiaming Ji, Mickel Liu, Josef Dai, Xuehai Pan, Chi Zhang, Ce Bian, Boyuan Chen, Ruiyang Sun, Yizhou Wang, and Yaodong Yang. Beavertails: Towards improved safety alignment of llm via a human-preference dataset.Advances in Neural Information Processing Systems, 36:24678–24704, 2023

    2023

  54. [54]

    Cvalues: Measuring the values of chinese large language models from safety to responsibility.arXiv preprint arXiv:2307.09705, 2023

    Guohai Xu, Jiayi Liu, Ming Yan, Haotian Xu, Jinghui Si, Zhuoran Zhou, Peng Yi, Xing Gao, Jitao Sang, Rong Zhang, et al. Cvalues: Measuring the values of chinese large language models from safety to responsibility.arXiv preprint arXiv:2307.09705, 2023

    work page arXiv 2023

  55. [55]

    Xstest: A test suite for identifying exaggerated safety behaviours in large language models

    Paul Röttger, Hannah Kirk, Bertie Vidgen, Giuseppe Attanasio, Federico Bianchi, and Dirk Hovy. Xstest: A test suite for identifying exaggerated safety behaviours in large language models. InProceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers),...

    2024

  56. [56]

    Latent jailbreak: A benchmark for evaluating text safety and output robustness of large language models.arXiv preprint arXiv:2307.08487, 2023

    Huachuan Qiu, Shuai Zhang, Anqi Li, Hongliang He, and Zhenzhong Lan. Latent jailbreak: A benchmark for evaluating text safety and output robustness of large language models.arXiv preprint arXiv:2307.08487, 2023

    work page arXiv 2023

  57. [57]

    do anything now

    Xinyue Shen, Zeyuan Chen, Michael Backes, Yun Shen, and Yang Zhang. " do anything now": Characterizing and evaluating in-the-wild jailbreak prompts on large language models. InProceedings of the 2024 on ACM SIGSAC Conference on Computer and Communications Security, pages 1671–1685, 2024. 13

    2024

  58. [58]

    Do-not-answer: Evaluating safeguards in llms

    Yuxia Wang, Haonan Li, Xudong Han, Preslav Nakov, and Timothy Baldwin. Do-not-answer: Evaluating safeguards in llms. InFindings of the Association for Computational Linguistics: EACL 2024, pages 896–911, 2024

    2024

  59. [59]

    All languages matter: On the multilingual safety of large language models

    Wenxuan Wang, Zhaopeng Tu, Chang Chen, Youliang Yuan, Jen-tse Huang, Wenxiang Jiao, and Michael R Lyu. All languages matter: On the multilingual safety of large language models. arXiv preprint arXiv:2310.00905, 2023

    work page arXiv 2023

  60. [60]

    Salad-bench: A hierarchical and comprehensive safety benchmark for large language models

    Lijun Li, Bowen Dong, Ruohui Wang, Xuhao Hu, Wangmeng Zuo, Dahua Lin, Yu Qiao, and Jing Shao. Salad-bench: A hierarchical and comprehensive safety benchmark for large language models. InFindings of the Association for Computational Linguistics: ACL 2024, pages 3923–3954, 2024

    2024

  61. [61]

    Sorry-bench: Systematically evaluating large language model safety refusal.arXiv preprint arXiv:2406.14598, 2024

    Tinghao Xie, Xiangyu Qi, Yi Zeng, Yangsibo Huang, Udari Madhushani Sehwag, Kaixuan Huang, Luxi He, Boyi Wei, Dacheng Li, Ying Sheng, et al. Sorry-bench: Systematically evaluating large language model safety refusal.arXiv preprint arXiv:2406.14598, 2024

    work page arXiv 2024

  62. [62]

    Position: Agent Should Invoke External Tools ONLY When Epistemically Necessary

    Hongru Wang, Cheng Qian, Manling Li, Jiahao Qiu, Boyang Xue, Mengdi Wang, Heng Ji, and Kam-Fai Wong. Toward a theory of agents as tool-use decision-makers.arXiv preprint arXiv:2506.00886, 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  63. [63]

    Reflexion: Language agents with verbal reinforcement learning.Advances in neural information processing systems, 36:8634–8652, 2023

    Noah Shinn, Federico Cassano, Ashwin Gopinath, Karthik Narasimhan, and Shunyu Yao. Reflexion: Language agents with verbal reinforcement learning.Advances in neural information processing systems, 36:8634–8652, 2023

    2023

  64. [64]

    Optimizing generative ai by backpropagating language model feedback.Nature, 639(8055):609–616, 2025

    Mert Yuksekgonul, Federico Bianchi, Joseph Boen, Sheng Liu, Pan Lu, Zhi Huang, Carlos Guestrin, and James Zou. Optimizing generative ai by backpropagating language model feedback.Nature, 639(8055):609–616, 2025

    2025

  65. [65]

    GEPA: Reflective Prompt Evolution Can Outperform Reinforcement Learning

    Lakshya A Agrawal, Shangyin Tan, Dilara Soylu, Noah Ziems, Rishi Khare, Krista Opsahl-Ong, Arnav Singhvi, Herumb Shandilya, Michael J Ryan, Meng Jiang, et al. Gepa: Reflective prompt evolution can outperform reinforcement learning.arXiv preprint arXiv:2507.19457, 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  66. [66]

    Dynamic cheatsheet: Test-time learning with adaptive memory

    Mirac Suzgun, Mert Yuksekgonul, Federico Bianchi, Dan Jurafsky, and James Zou. Dynamic cheatsheet: Test-time learning with adaptive memory. InProceedings of the 19th Conference of the European Chapter of the Association for Computational Linguistics (Volume 1: Long Papers), pages 7080–7106, 2026

    2026

  67. [67]

    The power of scale for parameter-efficient prompt tuning

    Brian Lester, Rami Al-Rfou, and Noah Constant. The power of scale for parameter-efficient prompt tuning. InProceedings of the 2021 conference on empirical methods in natural language processing, pages 3045–3059, 2021

    2021

  68. [68]

    Demystifying prompts in language models via perplexity estimation

    Hila Gonen, Srini Iyer, Terra Blevins, Noah A Smith, and Luke Zettlemoyer. Demystifying prompts in language models via perplexity estimation. InFindings of the Association for Computational Linguistics: EMNLP 2023, pages 10136–10148, 2023

    2023

  69. [69]

    Unnatural instructions: Tuning language models with (almost) no human labor

    Or Honovich, Thomas Scialom, Omer Levy, and Timo Schick. Unnatural instructions: Tuning language models with (almost) no human labor. InProceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 14409–14428, 2023

    2023

  70. [70]

    Dialogguard: Multi-agent psychosocial safety evaluation of sensitive llm responses.arXiv preprint arXiv:2512.02282, 2025

    Han Luo and Guy Laban. Dialogguard: Multi-agent psychosocial safety evaluation of sensitive llm responses.arXiv preprint arXiv:2512.02282, 2025

    work page arXiv 2025

  71. [71]

    Alcuna: Large language models meet new knowledge

    Xunjian Yin, Baizhou Huang, and Xiaojun Wan. Alcuna: Large language models meet new knowledge. InProceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 1397–1414, 2023

    2023

  72. [72]

    Bbq: A hand-built bias benchmark for question answering

    Alicia Parrish, Angelica Chen, Nikita Nangia, Vishakh Padmakumar, Jason Phang, Jana Thomp- son, Phu Mon Htut, and Samuel Bowman. Bbq: A hand-built bias benchmark for question answering. InFindings of the Association for Computational Linguistics: ACL 2022, pages 2086–2105, 2022. 14

    2022

  73. [73]

    Beyond the imitation game: Quantifying and extrapolating the capabilities of language models

    Aarohi Srivastava, Abhinav Rastogi, Abhishek Rao, Abu Awal Md Shoeb, Abubakar Abid, Adam Fisch, Adam R Brown, Adam Santoro, Aditya Gupta, Adrià Garriga-Alonso, et al. Beyond the imitation game: Quantifying and extrapolating the capabilities of language models. Transactions on machine learning research, 2023

    2023

  74. [74]

    Won’t get fooled again: Answering questions with false premises

    Shengding Hu, Yifan Luo, Huadong Wang, Xingyi Cheng, Zhiyuan Liu, and Maosong Sun. Won’t get fooled again: Answering questions with false premises. InProceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 5626–5643, 2023

    2023

  75. [75]

    Simcse: Simple contrastive learning of sentence embeddings

    Tianyu Gao, Xingcheng Yao, and Danqi Chen. Simcse: Simple contrastive learning of sentence embeddings. InProceedings of the 2021 conference on empirical methods in natural language processing, pages 6894–6910, 2021

    2021

  76. [76]

    Mediq: Question-asking llms and a benchmark for reliable interactive clinical reasoning.Advances in Neural Information Processing Systems, 37:28858–28888, 2024

    Shuyue S Li, Vidhisha Balachandran, Shangbin Feng, Jonathan S Ilgen, Emma Pierson, Pang W Koh, and Yulia Tsvetkov. Mediq: Question-asking llms and a benchmark for reliable interactive clinical reasoning.Advances in Neural Information Processing Systems, 37:28858–28888, 2024

    2024

  77. [77]

    Evaluating the moral beliefs encoded in llms.Advances in Neural Information Processing Systems, 36:51778–51809, 2023

    Nino Scherrer, Claudia Shi, Amir Feder, and David Blei. Evaluating the moral beliefs encoded in llms.Advances in Neural Information Processing Systems, 36:51778–51809, 2023

    2023

  78. [78]

    2: Question answering with questionable assumptions

    Najoung Kim, Phu Mon Htut, Samuel Bowman, and Jackson Petty. 2: Question answering with questionable assumptions. InProceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 8466–8487, 2023

    2023

  79. [79]

    Worldsense: A synthetic benchmark for grounded reasoning in large language models.arXiv preprint arXiv:2311.15930, 2023

    Youssef Benchekroun, Megi Dervishi, Mark Ibrahim, Jean-Baptiste Gaya, Xavier Martinet, Grégoire Mialon, Thomas Scialom, Emmanuel Dupoux, Dieuwke Hupkes, and Pascal Vincent. Worldsense: A synthetic benchmark for grounded reasoning in large language models.arXiv preprint arXiv:2311.15930, 2023

    work page arXiv 2023

  80. [80]

    Measuring Massive Multitask Language Understanding

    Dan Hendrycks, Collin Burns, Steven Basart, Andy Zou, Mantas Mazeika, Dawn Song, and Jacob Steinhardt. Measuring massive multitask language understanding.arXiv preprint arXiv:2009.03300, 2020

    work page internal anchor Pith review Pith/arXiv arXiv 2009

Showing first 80 references.