Pith. sign in

REVIEW 3 major objections 4 minor 64 references

Fewer Clarifications, Better Code: Benchmarking Cross-Session Personalized Ambiguity Adaptation in Coding Assistants

T0 review · 3 major / 4 minor · reviewed 2026-08-01 · deepseek-v4-flash

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

Pith's one-line read Coding assistants can learn a user's recurring ambiguity pattern from resolved sessions and apply it to new requests, reducing clarification and improving first-turn executable success.

desk verdict A genuinely new task and a well-controlled synthetic benchmark for cross-session ambiguity adaptation in coding assistants, but the generated histories are too clean to support the real-world transfer claim without release and further validation. read the letter →

arxiv 2607.26611 v1 pith:FBVDF5N5 submitted 2026-07-29 cs.AI cs.HC

Zijian Xu , Wenshuo Zhang , Zisen Qin , Rui Sheng , Yushi Sun , Huamin Qu , Chuhan Shi This is my paper
classification cs.AIcs.HC
keywords personalizedambiguityadaptationcodingassistantscross-sessionpersonalizationclarificationexecutablecodegenerationbenchmarkLLMevaluationlong-termmemory
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 argues that a coding assistant can learn a user's recurring ambiguity pattern from previously resolved coding sessions and use that pattern to interpret new ambiguous requests, producing intended executable code while asking fewer clarifying questions. To test this, it introduces CAPA, a benchmark of 600 coding sessions in which each synthetic user consistently obscures the same kind of implementation information and resolves it the same way, then measures 12 language models with and without that user's history. The central result is that same-user history improves first-turn executable success for all 12 models, by 15.6 percentage points on average, and shortens the average path to completion for every model. The paper also shows that correctly matched user identity matters more than generic history, and that a lightweight history-gating step can further improve first-turn success without hurting final success. If this transfers to real use, long-term coding assistants could become more accurate and less interruptive by remembering how a user typically leaves requirements underspecified.

What carries the argument

The central object is the personalized ambiguity-resolution assignment (ai, ri): for each user, a recurring mechanism ai specifies which required implementation information is omitted or obscured, paired with a resolution pattern ri that specifies the user's intended interpretation. CAPA instantiates this assignment across six ambiguity mechanisms and ten user profiles, generates 600 multi-turn coding sessions through a three-stage pipeline with an interactive simulation environment, and validates cross-session consistency before evaluation. This assignment is the load-bearing structure: it makes the target signal concrete, controllable, and testable, and it directly supports the proposed sa

What would settle it

Log real repeated coding requests from actual users across multiple sessions, then test whether an assistant given that user's history achieves higher first-turn executable success than one without it. If same-user history does not outperform no-history or shuffled history on real logs, or if user ambiguity patterns prove inconsistent across sessions, the core premise of CAPA is refuted.

Watch

Extended reading notes

Core claim

The paper claims that personalized ambiguity adaptation is a real and measurable capability: given a user's resolved session history and a new ambiguous request, an assistant can infer the user's stable ambiguity-resolution pattern and apply it to held-out tasks. On the CAPA benchmark, providing correctly matched same-user history improved executable success for 11 of 12 models, first-turn executable success for all 12, and reduced turns-to-completion for all 12, with an average first-turn gain of 15.6 percentage points. The authors further show this gain is not merely generic in-context learning: histories drawn from other users provide less first-turn benefit than correctly matched histori

Load-bearing premise

The benchmark's validity rests on synthetic user trajectories faithfully representing how real users express recurring ambiguity; if real users are less consistent or express ambiguity differently, the measured history gains may not transfer to deployment.

Editorial extensions

If this is right

  • Providing a user's resolved sessions lets assistants resolve recurring ambiguity earlier, lifting average first-turn executable success by 15.6 percentage points across the 12 tested models.
  • Same-user history reduces interaction cost: average turns-to-completion dropped by 0.81 turns, with every tested model finishing tasks faster.
  • Correct user identity matters: correctly matched histories improve first-turn success more than shuffled histories, though even generic history helps, showing both personalized and generic signals contribute.
  • General-purpose memory-management methods do not reliably beat raw history for this task; explicitly checking for ambiguity-resolution evidence yields larger first-turn gains.
  • CAPA is far from saturated: the strongest tested model still succeeds on only about 60% of first turns with history, leaving personalized ambiguity resolution an open challenge.

Reading between the lines

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

  • Beyond the paper, real-world gains may be smaller than benchmark gains because CAPA's synthetic personas maintain a fixed ambiguity pattern, whereas real users may be less consistent; a field study with genuine user logs is needed to confirm transfer.
  • Beyond the paper, the same pattern-recognition machinery—identifying a user's recurring underspecification and resolving it from history—could extend to other underspecified generation tasks, such as SQL queries, spreadsheet formulas, or data-analysis prompts.
  • Beyond the paper, the history-gating idea offers a cheap upgrade path for deployed assistants: add a pre-processing check that surfaces whether prior sessions contain consistent resolution evidence, without retraining the model.
  • Beyond the paper, the larger first-turn gain compared to final-success gain suggests history primarily helps assistants commit to the intended interpretation earlier, not rescue failed sessions; product design should therefore target early interpretation and confident first responses.
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

3 major / 4 minor

Summary. The paper proposes a new task, personalized ambiguity adaptation, in which a coding assistant uses a user's previously resolved coding sessions to infer a recurring ambiguity-resolution pattern and apply it to a new ambiguous request, ideally without asking clarifying questions. To support this task, the authors introduce CAPA, a benchmark constructed by starting from unambiguous HumanEval tasks and applying six mechanisms of personalized ambiguity through a three-stage pipeline. The dataset contains 600 sessions organized into 60 user–ambiguity cells, with five resolved history sessions and five held-out evaluation sessions per cell. The paper evaluates 12 LLMs, reporting Executable Success (ES), First-Turn Executable Success (FT-ES), and Turns-to-Completion (TTC) under no-history and same-user-history conditions. The central empirical finding is that same-user history improves ES for 11 of 12 models, FT-ES for all 12, and TTC for all models, with an average FT-ES gain of +15.6 percentage points. The paper additionally analyzes task difficulty, user identity, and memory methods, and proposes a same-user history gating method that improves FT-ES without hurting ES.

Significance. If the empirical claims are sustained, CAPA is a useful controlled benchmark for an underexplored problem: using cross-session user history to resolve recurring ambiguity in executable code generation. The paper has genuine methodological strengths: correctness is measured by hidden executable tests, the blank-control experiment shows that the underlying HumanEval tasks are nearly saturated under full specifications, and the shuffled-history control is a reasonable attempt to separate generic in-context learning from genuine user-specific adaptation. The proposed history gating is a lightweight, training-free method and is compared against two existing memory frameworks. However, the benchmark's external validity is the central open risk: the histories are produced by a ground-truth-aware simulation, and the paper itself limits the human validation to taxonomy labels rather than the final generated sessions. In addition, all headline comparisons are single-run point estimates without uncertainty quantification. These issues do not invalidate the contribution as a controlled benchmark, but they do require substantial revision before the paper's deployment-oriented claims can be acce

major comments (3)
  1. [Data Generation, Stage Two (AmbiSimu); Appendix 'Human validation of the taxonomy'] The central claim is that same-user history transfers to new tasks, but the histories are generated by a Clarification Planner that has access to ground truth and selects the single clarification that would 'most reduce the gap' to that ground truth. The user agent is simultaneously instructed to maintain a fixed, author-assigned ambiguity–resolution pattern. This produces near-optimal teaching sequences that real users are unlikely to provide: real users do not know a ground-truth specification, often give noisy or partial corrections, and may accept plausible-but-wrong code. The paper's own appendix states that the human annotation study validates only the taxonomy labels, not the final generated sessions. Because the magnitude of the reported history gains (+15.6 pp average FT-ES) may be an artifact of oracle-guided histories, the authors should either add a human naturalness/informat
  2. [Experiments, Table 1 (and Tables 3–5)] All head-to-head comparisons are single-run point estimates over N=300 held-out sessions, with no confidence intervals, standard errors, or significance tests. A 2.0 pp ES difference corresponds to only six sessions, and several reported gains are of this order (e.g., Claude Opus 4.8 +2.0 pp ES, DeepSeek V4 Pro +2.7 pp ES). The headline 'same-user history improves ES for 11 of 12 models' is therefore not established as a statistical claim. Likewise, the matched-vs-shuffled history gains in Table 3 are small for ES and moderately sized for FT-ES, but without variance estimates they cannot be separated from noise. Please report bootstrap confidence intervals or repeated evaluation runs, and consider effect sizes for the central comparisons.
  3. [Benchmark Task / Experiments (evaluation protocol)] The evaluation protocol does not specify how the user responds to the evaluated model's clarification questions. The task definition says only that 'the next user message u_i,k^(t) is added to the dialogue,' and the experiments section does not describe the user-side simulator used at evaluation time. If the same ground-truth-aware AmbiSimu user agent is reused online, the evaluated model receives oracle-aligned answers to its questions, which could inflate FT-ES and TTC gains. If the stored held-out traces are replayed instead, the assistant's clarification questions may not match the scripted user answers. This ambiguity affects every reported result. The authors must describe the exact interactive loop, identify which component produces user turns during evaluation, and demonstrate that it does not leak ground truth into the evaluation history.
minor comments (4)
  1. [Discussion, 'A lightweight same-user history gating method'] The method is described as 'parameter-free' while using a gate LLM. Please clarify that this means no additional training or fitted parameters, not that the workflow has no model parameters.
  2. [Data Generation, Stage Two] The 'Conversation LLM' used to generate the dialogue traces is never named. For reproducibility and to assess possible contamination with the evaluated models, please specify the exact model and version used for generation.
  3. [Appendix, 'Analysis by Personalized Ambiguity Mechanism'] Table 8 uses multi-label marginal aggregation because each session contains two jointly applied mechanisms. This should be flagged in the main text to prevent readers from interpreting mechanism-level numbers as causal effects of an isolated mechanism.
  4. [General] No data/code availability statement appears in the paper. Since CAPA is a benchmark, a release plan or link should be included.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: CAPA's headline gains are measured against external executable tests, with shuffled-history and blank controls; synthetic-history realism is an external-validity concern, not a derivation-cycle.

full rationale

CAPA is a benchmark-construction paper rather than an analytical derivation. The central empirical claim—that same-user history improves ES, FT-ES, and TTC relative to no history—is not fitted or defined into existence: it comes from running 12 LLMs on 300 held-out sessions, judged by hidden executable tests from HumanEval, under controlled no-history versus same-user-history interventions. The ground-truth-aware Clarification Planner in Stage Two certainly makes the synthetic histories unusually informative, and the human annotation study validates only the taxonomy labels rather than the generated sessions; these are legitimate concerns about construct validity and real-world transfer, but they do not make the measured history gains true by construction, since the LLMs could still fail to exploit the history. The paper even includes a shuffled-history control (RQ2) showing that matched same-user identity adds beyond generic context, and a blank-control experiment showing the underlying HumanEval tasks are solvable without ambiguity. The taxonomy is explicitly adapted from Li et al. (2024) and grounded in WildChat, not presented as a first-principles derivation. Self-citations in the related-work section are contextual and not load-bearing. The appendix's design-consideration caveat about marginal mechanism attribution is a statistical interpretation limitation, not a circular step. No equation, fitted parameter, or predicted quantity reduces to its own input.

Assumptions & free parameters 3 free parameters · 4 assumptions · 2 invented entities

The central experimental object is a synthetic benchmark. The main ledger entries are design choices and domain assumptions rather than fitted parameters. The benchmark depends on the realism of LLM-simulated users, the validity of the ambiguity taxonomy, and the treatment of HumanEval hidden tests as ground truth for the intended personal resolution.

free parameters (3)
  • Eight-turn interaction budget
    Hand-set for all models; defines the ES denominator and assigns unsolved sessions a TTC value of 8.
  • Difficulty partition thresholds (simple 1–2, medium 3–4, complex 5–8 assistant turns)
    Chosen from the DeepSeek V4 Pro no-history reference run and then reused as fixed labels for RQ1 comparisons on other models.
  • Five history sessions and five held-out sessions per cell
    Hand-selected split controls how much cross-session evidence a model receives; no ablation varying history size is reported.
assumptions (4)
  • domain assumption The six-mechanism taxonomy, adapted from Li et al. with human annotation, categorizes how real users leave coding intent underspecified.
    The taxonomy is human-validated with Fleiss' κ of 0.66, but the generated sessions themselves are not directly validated as natural or complete.
  • domain assumption LLM persona simulation faithfully reproduces a real user's recurring ambiguity style across sessions.
    Stage Two relies on a user agent repeatedly applying the same assigned ambiguity mechanism and resolution pattern; no real-user trajectory data is used.
  • domain assumption HumanEval hidden tests and the reference solution are sufficient to judge whether the intended ambiguous implementation was produced.
    The judge J combines test execution with reference-solution comparison, equating 'resolved ambiguity' with 'passes hidden tests'.
  • domain assumption Stage-Three LLM consistency validation catches any drift in the ambiguity pattern.
    A session is accepted if the LLM-based check finds the assigned personalized ambiguity preserved; no independent human check on final sessions is reported.
invented entities (2)
  • Ten synthetic user personas (P01–P10)
    purpose: Generate consistent expression styles and ambiguity-resolution behavior for simulated users.
    Personas are authored constructs; no empirical evidence that they correspond to real user distributions.
  • Six personalized ambiguity mechanisms (e.g., domain-cognitive polysemy, habitual context omission) independent evidence
    purpose: Taxonomy axes for injecting and measuring recurring ambiguity.
    Grounded in WildChat and human-annotated with Fleiss' κ=0.66; recognizable in requests, though not verified on final generated sessions.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Fewer Clarifications, Better Code: Benchmarking Cross-Session Personalized Ambiguity Adaptation in Coding Assistants." pith.science (2026). https://pith.science/paper/FBVDF5N5

@misc{pith2026260726611,
  author       = {Pith},
  title        = {Pith review of: Fewer Clarifications, Better Code: Benchmarking Cross-Session Personalized Ambiguity Adaptation in Coding Assistants},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/FBVDF5N5}},
  note         = {Machine review of arXiv:2607.26611}
}
read the original abstract

AI-assisted coding increasingly translates informal user intent into executable software, yet coding requests often contain ambiguities that recur in user-specific ways across tasks and sessions. Existing disambiguation methods typically address each ambiguous request in isolation within the current coding session, often through eliciting additional clarification. However, whether resolved session history from the same user can serve as memory for resolving recurring personalized ambiguity in a newly opened session remains underexplored. We formulate personalized ambiguity adaptation as a new task: given a user's previously resolved coding sessions and a new ambiguous request, an assistant should identify the recurring ambiguity pattern, produce the intended executable solution, and minimize clarification. To benchmark this task, we introduce CAPA, which characterizes personalized coding ambiguity through six mechanisms and injects these mechanisms into unambiguous executable tasks using a controlled three-stage generation pipeline. CAPA contains 600 coding sessions across 60 balanced user--ambiguity cells, including 300 held-out evaluation sessions. We evaluate 12 recent LLMs under no-history and same-user-history conditions using executable success, first-turn success, and turns-to-completion. Our analyses examine task difficulty, user identity, and memory-based history use, and we further propose same-user history gating as a lightweight inference-time method. CAPA provides a foundation for developing long-term coding assistants that better align generated code with user intent while reducing repeated clarification.

Figures

Figures reproduced from arXiv: 2607.26611 by the authors.

Figure 1
Figure 1. Overview of the history-conditioned personalized [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Overview of the CAPA data generation pipeline. Stage One combines task requirements with a user profile and an [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Overview of same-user history gating. Evaluation. Compared with raw same-user history, our gating method consistently improves FT-ES by 0.66–13.33 pp while keeping ES within ±1.0 pp across all three models. It also reduces TTC for GPT-5.5 and GLM-5.2, demonstrat￾ing that explicitly checking ambiguity-resolution evidence generally enables earlier completion without substantially affecting eventual success. Conclusion… view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Dataset balance and task difficulty. (a) CAPA comprises 600 sessions in 60 balanced cells defined by ten user profiles [PITH_FULL_IMAGE:figures/full_fig_p010_4.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

64 extracted references · 19 linked inside Pith

  1. [1]

    Proceedings of the 2020 conference on empirical methods in natural language processing (EMNLP) , pages=

    AmbigQA: Answering ambiguous open-domain questions , author=. Proceedings of the 2020 conference on empirical methods in natural language processing (EMNLP) , pages=

    2020

  2. [2]

    Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

    CLAMBER: A benchmark of identifying and clarifying ambiguous information needs in large language models , author=. Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

  3. [3]

    Advances in Neural Information Processing Systems , volume=

    Ambrosia: A benchmark for parsing ambiguous questions into database queries , author=. Advances in Neural Information Processing Systems , volume=

  4. [4]

    Findings of the Association for Computational Linguistics: NAACL 2025 , pages=

    Clarify when necessary: Resolving ambiguity through interaction with lms , author=. Findings of the Association for Computational Linguistics: NAACL 2025 , pages=

    2025

  5. [5]

    Luo, Sichun and Huang, Yi and Li, Mukai and Meng, Shichang and Liu, Fengyuan and Hu, Zefa and Feng, Junlan and Liu, Qi , journal =

  6. [6]

    and Liu, Alisa and Wu, Zhaofeng and Smith, Noah A

    Li, Margaret Y. and Liu, Alisa and Wu, Zhaofeng and Smith, Noah A. , journal =. A Taxonomy of Ambiguity Types for

  7. [7]

    Zhao, Wenting and Ren, Xiang and Hessel, Jack and Cardie, Claire and Choi, Yejin and Deng, Yuntian , journal =

  8. [8]

    Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

    Lamp: When large language models meet personalization , author=. Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

Show all 64 references
  1. [9]

    Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

    Evaluating very long-term conversational memory of llm agents , author=. Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

  2. [10]

    NeurIPS 2025 Workshop on Evaluating the Evolving LLM Lifecycle: Benchmarks, Emergent Abilities, and Scaling , year=

    Know me, respond to me: Benchmarking llms for dynamic user profiling and personalized responses at scale , author=. NeurIPS 2025 Workshop on Evaluating the Evolving LLM Lifecycle: Benchmarks, Emergent Abilities, and Scaling , year=

    2025

  3. [11]

    arXiv preprint arXiv:2604.20006 , year =

    From Recall to Forgetting: Benchmarking Long-Term Memory for Personalized Agents , author =. arXiv preprint arXiv:2604.20006 , year =

    arXiv

  4. [12]

    arXiv preprint arXiv:2410.10813 , year=

    Longmemeval: Benchmarking chat assistants on long-term interactive memory , author=. arXiv preprint arXiv:2410.10813 , year=

    arXiv

  5. [13]

    arXiv preprint arXiv:2107.03374 , year=

    Evaluating large language models trained on code , author=. arXiv preprint arXiv:2107.03374 , year=

    arXiv

  6. [14]

    arXiv preprint arXiv:2108.07732 , year=

    Program synthesis with large language models , author=. arXiv preprint arXiv:2108.07732 , year=

    arXiv

  7. [15]

    Advances in neural information processing systems , volume=

    Is your code generated by chatgpt really correct? rigorous evaluation of large language models for code generation , author=. Advances in neural information processing systems , volume=

  8. [16]

    ACM Transactions on Software Engineering and Methodology , volume=

    Humanevalcomm: Benchmarking the communication competence of code generation for llms and llm agents , author=. ACM Transactions on Software Engineering and Methodology , volume=. 2025 , publisher=

    2025

  9. [17]

    arXiv preprint arXiv:2604.21505 , year=

    Assessing the Impact of Requirement Ambiguity on LLM-based Function-Level Code Generation , author=. arXiv preprint arXiv:2604.21505 , year=

    arXiv

  10. [18]

    arXiv preprint arXiv:2607.00711 , year=

    ClarifyCodeBench: Evaluating LLMs on Clarifying Ambiguous Requirements for Code Generation , author=. arXiv preprint arXiv:2607.00711 , year=

    arXiv

  11. [19]

    International Conference on Learning Representations , volume=

    Swe-bench: Can language models resolve real-world github issues? , author=. International Conference on Learning Representations , volume=

  12. [20]

    The Thirty-eighth Annual Conference on Neural Information Processing Systems , year=

    Swe-agent: Agent-computer interfaces enable automated software engineering , author=. The Thirty-eighth Annual Conference on Neural Information Processing Systems , year=

  13. [21]

    International Conference on Learning Representations , volume=

    Convcodeworld: Benchmarking conversational code generation in reproducible feedback environments , author=. International Conference on Learning Representations , volume=

  14. [22]

    Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing , pages=

    Disambiguation in Conversational Question Answering in the Era of LLMs and Agents: A Survey , author=. Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing , pages=

    2025

  15. [23]

    American Journal of Computational Linguistics , volume=

    Coping with syntactic ambiguity or how to put the block in the box on the table , author=. American Journal of Computational Linguistics , volume=

  16. [24]

    ACM computing surveys (CSUR) , volume=

    Word sense disambiguation: A survey , author=. ACM computing surveys (CSUR) , volume=. 2009 , publisher=

    2009

  17. [25]

    Proceedings of the 5th SIGdial Workshop on Discourse and Dialogue at HLT-NAACL 2004 , pages=

    Causes and strategies for requesting clarification in dialogue , author=. Proceedings of the 5th SIGdial Workshop on Discourse and Dialogue at HLT-NAACL 2004 , pages=

    2004

  18. [26]

    Proceedings of the 2023 conference on empirical methods in natural language processing , pages=

    We’re afraid language models aren’t modeling ambiguity , author=. Proceedings of the 2023 conference on empirical methods in natural language processing , pages=

    2023

  19. [27]

    3rd Conference on Automated Knowledge Base Construction , year=

    Abg-coqa: Clarifying ambiguity in conversational question answering , author=. 3rd Conference on Automated Knowledge Base Construction , year=

  20. [28]

    European conference on information retrieval , pages=

    Identifying unclear questions in community question answering websites , author=. European conference on information retrieval , pages=. 2019 , organization=

    2019

  21. [29]

    arXiv preprint arXiv:2008.07559 , year=

    Resolving intent ambiguities by retrieving discriminative clarifying questions , author=. arXiv preprint arXiv:2008.07559 , year=

    2008 arXiv

  22. [30]

    arXiv preprint arXiv:2502.00537 , year=

    Detecting ambiguities to guide query rewrite for robust conversations in enterprise ai assistants , author=. arXiv preprint arXiv:2502.00537 , year=

    arXiv

  23. [31]

    arXiv preprint arXiv:2212.07769 , year=

    Clam: Selective clarification for ambiguous questions with generative language models , author=. arXiv preprint arXiv:2212.07769 , year=

    arXiv

  24. [32]

    Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing , pages=

    Aligning language models to explicitly handle ambiguity , author=. Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing , pages=

    2024

  25. [33]

    Can you unpack that? learning to rewrite questions-in-context , author=. Proceedings of the 2019 conference on empirical methods in natural language processing and the 9th international joint conference on natural language processing (EMNLP-IJCNLP) , pages=

    2019

  26. [34]

    Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing , pages=

    Query rewriting in retrieval-augmented large language models , author=. Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing , pages=

    2023

  27. [35]

    Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing , pages=

    ASQA: Factoid questions meet long-form answers , author=. Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing , pages=

    2022

  28. [36]

    Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing , pages=

    Tree of clarifications: Answering ambiguous questions with retrieval-augmented large language models , author=. Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing , pages=

    2023

  29. [37]

    Diversify-verify-adapt: Efficient and robust retrieval-augmented ambiguous question answering , author=. Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Pape...

    2025

  30. [38]

    Proceedings of the 42nd international acm sigir conference on research and development in information retrieval , pages=

    Asking clarifying questions in open-domain information-seeking conversations , author=. Proceedings of the 42nd international acm sigir conference on research and development in information retrieval , pages=

  31. [39]

    Asking clarification questions in knowledge-based question answering , author=. Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP) , pages=

    2019

  32. [40]

    Findings of the Association for Computational Linguistics: EMNLP 2023 , pages=

    Asking clarification questions to handle ambiguity in open-domain QA , author=. Findings of the Association for Computational Linguistics: EMNLP 2023 , pages=

    2023

  33. [41]

    Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies , pages=

    Open-domain question answering goes conversational via question rewriting , author=. Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies , pages=

    2021

  34. [42]

    Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

    Learning to ask good questions: Ranking clarification questions using neural expected value of perfect information , author=. Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

  35. [43]

    International Conference on Learning Representations , volume=

    Modeling future conversation turns to teach llms to ask clarifying questions , author=. International Conference on Learning Representations , volume=

  36. [44]

    arXiv preprint arXiv:2606.04547 , year=

    Beyond Retrieval: Learning Compact User Representations for Scalable LLM Personalization , author=. arXiv preprint arXiv:2606.04547 , year=

    arXiv

  37. [45]

    arXiv preprint arXiv:2606.20911 , year=

    Latent Personal Memory: Represent personal memory as dynamic soft prompts , author=. arXiv preprint arXiv:2606.20911 , year=

    arXiv

  38. [46]

    Findings of the Association for Computational Linguistics: ACL 2026 , pages=

    PersonaAgent: Bridging Memory and Action for Personalized LLM Agents , author=. Findings of the Association for Computational Linguistics: ACL 2026 , pages=

    2026

  39. [47]

    arXiv preprint arXiv:2606.04780 , year=

    PersonaTree: Structured Lifecycle Memory for Person Understanding in LLM Agents , author=. arXiv preprint arXiv:2606.04780 , year=

    arXiv

  40. [48]

    arXiv preprint arXiv:2512.13564 , year=

    Memory in the age of ai agents , author=. arXiv preprint arXiv:2512.13564 , year=

    arXiv

  41. [49]

    arXiv preprint arXiv:2504.19413 , year=

    Mem0: Building production-ready ai agents with scalable long-term memory , author=. arXiv preprint arXiv:2504.19413 , year=

    arXiv

  42. [50]

    Advances in Neural Information Processing Systems , volume=

    A-mem: Agentic memory for llm agents , author=. Advances in Neural Information Processing Systems , volume=

  43. [51]

    arXiv preprint arXiv:2605.25535 , year=

    Personalize-then-Store: Benchmarking and Learning Personalized Memory for Long-horizon Agents , author=. arXiv preprint arXiv:2605.25535 , year=

    arXiv

  44. [52]

    arXiv preprint arXiv:2607.00017 , year=

    Learning User-Aware Recall: Personalized Retrieval in Long-Term Conversational Memory , author=. arXiv preprint arXiv:2607.00017 , year=

    arXiv

  45. [53]

    arXiv preprint arXiv:2606.00832 , year=

    Momento: Evaluating Persistent Memory and Reasoning with Multi-Session Agentic Conversations , author=. arXiv preprint arXiv:2606.00832 , year=

    arXiv

  46. [54]

    arXiv preprint arXiv:2607.03162 , year=

    APeB: Benchmarking Personalization Ability of Large Language Model Agents , author=. arXiv preprint arXiv:2607.03162 , year=

    arXiv

  47. [55]

    arXiv preprint arXiv:2604.17886 , year=

    Latent Preference Modeling for Cross-Session Personalized Tool Calling , author=. arXiv preprint arXiv:2604.17886 , year=

    arXiv

  48. [56]

    2026 , month = apr, howpublished =

    2026

  49. [57]

    2026 , month = jul, howpublished =

    2026

  50. [58]

    2026 , month = may, howpublished =

    2026

  51. [59]

    2026 , month = feb, howpublished =

    2026

  52. [60]

    2026 , eprint=

    DeepSeek-V4: Towards Highly Efficient Million-Token Context Intelligence , author=. 2026 , eprint=

    2026

  53. [61]

    2026 , howpublished =

    2026

  54. [62]

    2026 , eprint=

    GLM-5: from Vibe Coding to Agentic Engineering , author=. 2026 , eprint=

    2026

  55. [63]

    2025 , eprint=

    Qwen3 Technical Report , author=. 2025 , eprint=

    2025

  56. [64]

    2024 , howpublished =

    2024

Pith tools

Reviewed August 1, 2026 · model on record in the stance chip above.