Pith. sign in

REVIEW 3 major objections 5 minor 1 cited by

Xolver: Multi-Agent Reasoning with Holistic Experience Learning Just Like an Olympiad Team

T0 review · 3 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read Xolver claims a training-free memory loop lets small LLMs outscore frontier reasoning models on math and code.

desk verdict A well-assembled multi-agent scaffold with a serious evaluation hole: the retrieval corpus and in-test-set memory updates likely contaminate its headline SOTA numbers. read the letter →

arxiv 2506.14234 v1 pith:MDETPLRM submitted 2025-06-17 cs.CL cs.AI

Md Tanzib Hosain , Salman Rahman , Md Kishor Morol , Md Rizwan Parvez This is my paper
classification cs.CLcs.AI
keywords multi-agentreasoningholisticexperiencelearningepisodicmemoryretrieval-augmentedgenerationiterativerefinementmathcodeinference-time
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

Xolver is a training-free scaffold that wraps a single black-box LLM in a team of role-specialized agents and gives them two memories: an episodic store of past problems and solutions, and a shared store of the highest-scoring reasoning traces produced so far. The paper's central claim is that this holistic experience learning, not the strength of the backbone model, produces the reported gains. With a QWQ-32B backbone the framework often beats much larger models such as Qwen3-235B, Gemini 2.5 Pro, o3, and o4-mini-high, and with an o3-mini-high backbone it reports new best scores on GSM8K, AIME 2024, AIME 2025, Math-500, and LiveCodeBench. The authors position Xolver as evidence that isolated inference can be replaced by experience-aware agents that learn from related problems even during competition.

What carries the argument

The mechanism that carries the argument is the dual-memory experience loop. An episodic memory $D_E$ holds external corpora (the OpenMathReason dataset for math, a 9-million-token corpus of algorithmic problems and their C++ solutions for code) or falls back to the model's own parametric memory via self-retrieval; an intermediate shared memory $D_S$ keeps the top-$m$ agent trajectories by judge score, so each iteration builds on the best traces and evaluation feedback produced so far. A planner over-generates candidate expert roles and selects $m$ of them, and a judge agent scores each response while a verifier or debugger extracts the final answer. After each problem, UPDATEEPISODICMEMORY adds the top-ranked trace back into the episodic store, closing the loop that the paper calls holistic experience learning.

What would settle it

Measure the overlap between the OpenMathReason corpus and the GSM8K, AIME 2024, AIME 2025, and Math-500 test sets using near-duplicate detection (e.g., BM25 or embedding similarity), and rerun Xolver (+) with episodic updates disabled and with any overlapping problems removed from the corpus; a large accuracy drop in that condition would show that retrieval leakage, not holistic experience learning, explains part of the results.

Watch

Extended reading notes

Core claim

The central discovery claimed is that orchestrating a black-box LLM as a set of dynamic expert agents that retrieve similar solved problems, share their best intermediate reasoning in a fixed-size memory, receive judge feedback, refine iteratively, and then store the solved problem back into an episodic corpus yields consistent, backbone-agnostic improvements across mathematics and programming benchmarks. Concretely, the paper reports that Xolver (+) with o3-mini-medium improves AIME'24 from 75.8 to 93.8 and LiveCodeBench from 66.3 to 87.3, and that with o3-mini-high it achieves 98.1 on GSM8K, 94.4 on AIME'24, 93.7 on AIME'25, 99.8 on Math-500, and 91.6 on LiveCodeBench. The authors argue these results show that accumulating and reusing experience at inference time, rather than scaling the underlying model, is the key step toward expert-level generalist reasoning.

Load-bearing premise

The evaluation problems and their solutions do not appear in the retrieval corpus or in the episodic memory accumulated from earlier test problems, so the reported accuracy gains reflect reasoning transfer rather than recalling answers from memory.

Editorial extensions

If this is right

  • If the central claim is correct, training-free scaffolding can lift small open-weight models past much larger proprietary ones on competitive math and code benchmarks, reducing the compute and access barrier for expert-level reasoning.
  • The reported +3.5 point average gain of the cross-problem variant over the problem-specific variant implies that storing each solved test problem in episodic memory transfers useful experience to later problems in the same test set.
  • With linear wall-clock complexity in iterations (agents run in parallel) but higher token usage, the framework trades about 1.5x the tokens of a single-agent search baseline for substantial accuracy gains, making it a cost-performance alternative to self-consistency's 32-64 generations.
  • Because improvements hold across two different backbones (o3-mini-medium and QWQ-32B), the paper expects similar gains if the same scaffold is applied to other black-box LLMs.
  • Ablations showing the largest drops when multi-iteration, multi-agent, or judge components are removed indicate that collaborative iterative refinement, not tool use or retrieval alone, is the core of the claimed gain.

Reading between the lines

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

  • Inference: the headline numbers rest on the absence of overlap between the OpenMathReason retrieval corpus (and the in-test-set episodic updates) and the evaluation problems; a near-duplicate analysis would tell whether the gains are retrieval of memorized answers or transfer of reasoning.
  • Inference: since the judge is the same backbone LLM that generates the solutions, part of the reported gain may come from self-consistency-style selection among candidate answers; an equal-budget comparison against self-consistency with the same token usage would isolate the contribution of memory and collaboration.
  • Inference: the qualitative coding example shows the team abandoning a partial rule in favor of a general key-based factorization, suggesting the multi-agent loop explores a broader hypothesis space; measuring solution diversity and convergence could test whether this exploration is what improves hard problems.
  • Inference: the framework's dependence on the backbone means its error analysis (wrong derivations, uninvoked tools) points to a natural next test: giving agents stronger tool-use incentives or a verifier that is not the same model could further close the remaining gap.
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 / 5 minor

Summary. Xolver is a training-free, multi-agent inference framework that augments a black-box LLM with an episodic memory (external corpus or self-retrieval), an evolving intermediate shared memory, multi-agent role assignment, judge-based evaluation, tool use, iterative refinement, and a final verifier. The paper claims that this holistic experience learning yields large gains over the backbone alone and, with o3-mini-high, new state-of-the-art results on GSM8K (98.1%), AIME'24 (94.4%), AIME'25 (93.7%), Math-500 (99.8%), and LiveCodeBench-V5 (91.6%). The authors also report that even a QWQ-32B backbone wrapped in Xolver often surpasses much larger proprietary models. The evaluation includes ablations for each component, retrieval strategies, agent counts, iteration counts, token costs, and a limited order-shuffling analysis.

Significance. If the reported results are valid, the paper makes a useful contribution: it demonstrates a model-agnostic, training-free framework that combines retrieval, multi-agent collaboration, iterative refinement, and memory in one pipeline, with unusually detailed prompts and qualitative traces. The multi-run variance reporting and the component ablations are also positive features. However, the central claim is contingent on an unexamined evaluation protocol. The default Xolver(+) variant updates the episodic memory with each test problem during evaluation, and the math retrieval corpus is a public problem-and-solution dataset that plausibly contains AIME and related problems. No overlap analysis is provided. Because Figure 5 shows external retrieval is worth roughly 10-17 accuracy points, this is not a minor caveat: it directly bears on whether the headline SOTA numbers measure reasoning and experience learning or retrieval leakage. The comparison to non-adaptive baselines is also inequitable under this protocol. These concerns must be resolved before the paper's claims can be accepted.

major comments (3)
  1. The default protocol, Xolver(+), updates the episodic memory with each test problem and its top-ranked reasoning trace during evaluation (Algorithm 1, line 13; Section 2.2, UPDATEEPISODICMEMORY), and the math external memory is the OpenMathReason dataset (Section 3.1, Inference Details). OpenMathReason is a public dataset of math problems and solutions, and AIME'24, AIME'25, GSM8K, and Math-500 are public problem sets; exact or near-duplicate overlap between the retrieval corpus and the test problems is therefore plausible. The paper never reports an overlap analysis, and the shuffling ablation in Appendix D.3 only randomizes test order; it does not test corpus contamination. Figure 5 attributes roughly 10-17 accuracy points to external retrieval over no retrieval, so if retrieved exemplars contain the target problem and its solution, the abstract's SOTA claims are inflated by retrieval rather than by holistic experience learning. The comparison to non-adaptive baselines is also inequitable because Xolver(+) can use earlier test problems and their generated solutions, while the baseline o3/o4 runs cannot. Please report exact and near-duplicate overlap between OpenMathReason and each math benchmark, re-run all math experiments with test-set problems and near-duplicates removed from the retrieval corpus, and separately report both Xolver(+) and Xolver(-) under this cleaned protocol.
  2. [§3.1, §4, Figure 4] The main results use hand-selected hyperparameters--m=3 agents, I=2 iterations, k=5 retrieved examples, and temperature 0.2--and no validation split or hyperparameter selection procedure is described. Figure 4 shows that accuracy improves monotonically with more agents and more iterations, so the reported magnitude of gains could partly reflect tuning on the target benchmarks. Please provide a validation-based selection protocol or a sensitivity analysis showing that the qualitative conclusions are stable over a range of m, I, and k values, rather than only at the tuned operating point.
  3. [§D.3, Table 5] The data-shuffling ablation in Appendix D.3 reports means that are inconsistent with the main results in Table 1 (e.g., o3-mini-medium AIME'24: 92.2±0.4 vs. 93.8; AIME'25: 91.0±0.3 vs. 89.4). The table caption also mentions 'intermediate shared memory' rather than the episodic memory update that defines Xolver(+), making it unclear whether the shuffled experiment actually tests the protocol described in the main text. Please clarify the exact protocol used in D.3 and reconcile the numbers with Table 1.
minor comments (5)
  1. [§1] The sentence containing 'they standalone each problem instance' appears to have a verb error; it should read something like 'they treat each problem instance as standalone.'
  2. [§3.2] There is a missing closing parenthesis in 'Xolver (+outperforms o4-mini-high'.
  3. [Figure 8] The caption says 'Average Token Usage (Input, Think, Output) per Dataset' but the surrounding text describes a comparison with Search-o1; please clarify whether the figure shows Xolver only or Xolver alongside Search-o1, and define 'Think' tokens.
  4. [Appendix D.3] The title 'Does Data Shuffling Affect Xolver (+) Performance?' and the table caption 'Impact of using intermediate shared memory with shuffle of order in test set in Xolver' are inconsistent about which memory component is being updated; please align the terminology.
  5. [Table 1 and Section 3.1] For several baselines the paper reports official numbers from technical reports rather than reproducing them under the same evaluation settings; please state explicitly, per baseline, whether the comparison uses official numbers or local reproduction, and report standard deviations for the locally reproduced baselines.

Circularity Check

1 steps flagged · score 6.0 of 10

Xolver's headline SOTA is partly circular: Algorithm 1 writes every test problem into episodic memory during evaluation, so Xolver(+) gains measure in-test-set access; external-corpus contamination is also unexamined.

  1. fitted input called prediction [Section 2.2 (Episodic Memory DE) and Section 2.3 / Algorithm 1, lines 12-13; Section 3.1 (Inference Details)]
    "In the case of an external episodic memory, DE can also be updated with UPDATEEPISODICMEMORY by adding the top-ranked reasoning and response from DS, paired with the problem q, into the external corpus Dext E. That is, Dext E <- Dext E ∪ (q, T, R), where (T, R, S, a) is the top-ranked entry in DS. ... 13: UPDATEEPISODICMEMORY(DE, q, DS)"

    During evaluation, Algorithm 1 line 13 inserts the current benchmark query q and its top-ranked solution into the retrieval corpus that R(DE) searches for the next query. The first-iteration context is BUILDCONTEXT = {q} ∪ R(DE), so later test problems can retrieve earlier test problems with their model-generated solutions. The reported advantage of Xolver(+) over Xolver(–) — '+3.5 point improvement' on average and '+7.7 points' on LiveCodeBench — is therefore partly the benefit of exposing the model to the test set itself, not evidence of transferable holistic experience. The abstract's 'new best result' claim is not an independent prediction because the evaluation set is, by construction, an input to the episodic memory.

full rationale

The framework has substantial independent content: Xolver(–), which keeps episodic memory static, still improves over the backbone (e.g., o3-mini-medium AIME'24 87.2 vs 75.8 LongCoT), and the multi-agent/judge/verifier ablations isolate real components. However, the headline numbers in the abstract and Table 1 use Xolver(+), whose Algorithm 1 line 13 adds each test query and solution to DE during evaluation. This makes the 'cross-problem experience' gain partially circular: later benchmark items see earlier benchmark items and their solutions through BM25 retrieval, a transductive protocol not available to the listed baselines. No overlap analysis between the OpenMathReason retrieval corpus and AIME/GSM8K/Math-500 is reported, even though OpenMathReason's own model reaches 93.3 on AIME'24 in Table 1, so the large external-retrieval gains in Figure 5 (e.g., 91.6 vs 73.8 no-retrieval on AIME'25 with o3-mini-medium) could be inflated by near-duplicate retrieval rather than reasoning. Hyperparameters (m=3, I=2, k=5, temperature 0.2) are stated without a benchmark-independent selection rule, and Figure 4 shows monotonic gains with more agents/iterations, so the reported operating point is a tunable cap; this is a correctness/robustness concern rather than a demonstrated fit. Self-citations (CodeSim, MapCoder, Parvez et al.) are ordinary technique references and are not load-bearing for the derivation. Overall the derivation does not reduce to a fitted parameter in the classic sense, but the central SOTA claim is partly circular because the evaluation set is fed back into the retrieval memory by construction; score 6.

Assumptions & free parameters 6 free parameters · 5 assumptions · 0 invented entities

The framework introduces no new physical entities. The central claims rest on four unverified assumptions: no test-set leakage in retrieval corpora, faithful judge-simulated execution, reliable judge scoring, and accurate GPT-4o grading. The main free parameters are default hyperparameters tuned on the target benchmarks.

free parameters (6)
  • number_of_agents_m = 3
    Chosen as default; ablation in Figure 4 shows accuracy improves with more agents, so the reported results reflect a tuned value rather than a principled choice.
  • max_iterations_I = 2
    Chosen as default; Figure 4 shows accuracy improves with more iterations, so this value is tuned on the target benchmarks.
  • retrieval_k = 5
    Top-k retrieved examples; no sensitivity analysis is provided.
  • temperature = 0.2
    Sampling temperature for generation; no sensitivity analysis.
  • synthesized_test_cases_Ntest = 10
    Number of synthesized test cases per code problem; chosen ad hoc.
  • convergence_threshold = 1.0 / all tests passed
    Judge score required for early stopping; depends on judge reliability.
assumptions (5)
  • domain assumption The external retrieval corpora (OpenMathReason for math, cp-algorithms for code) contain no test-set problems or near-duplicates.
    Section 3.1 defines the corpora; no overlap or contamination analysis is provided, yet the SOTA claims depend on retrieved exemplars not including the benchmark answers.
  • domain assumption LLM-simulated execution of test cases (CodeSim protocol) is a faithful proxy for real execution.
    Section 2.1 (Judge Agent) states test outcomes are determined by LLM prompting to avoid compiler interaction latency; the reliability of this simulation is not validated.
  • domain assumption The judge LLM's correctness scores reliably rank candidate responses.
    Sections 2.1 and 2.2 use judge scores to select the top-m entries in shared memory and to declare convergence; self-judging bias is acknowledged in ablations, but the external judge is still the same backbone model family.
  • domain assumption GPT-4o grading of final math answers is accurate.
    Section 3.1 uses GPT-4o for accuracy; no human verification or error analysis of the grader is provided.
  • domain assumption OpenMathReason and cp-algorithms are representative enough to yield useful exemplars for the target benchmarks.
    Section 3.1; the choice of retrieval corpus is not a target result but a design assumption that affects all reported numbers.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Xolver: Multi-Agent Reasoning with Holistic Experience Learning Just Like an Olympiad Team." pith.science (2026). https://pith.science/paper/MDETPLRM

@misc{pith2026250614234,
  author       = {Pith},
  title        = {Pith review of: Xolver: Multi-Agent Reasoning with Holistic Experience Learning Just Like an Olympiad Team},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/MDETPLRM}},
  note         = {Machine review of arXiv:2506.14234}
}
read the original abstract

Despite impressive progress on complex reasoning, current large language models (LLMs) typically operate in isolation - treating each problem as an independent attempt, without accumulating or integrating experiential knowledge. In contrast, expert problem solvers - such as Olympiad or programming contest teams - leverage a rich tapestry of experiences: absorbing mentorship from coaches, developing intuition from past problems, leveraging knowledge of tool usage and library functionality, adapting strategies based on the expertise and experiences of peers, continuously refining their reasoning through trial and error, and learning from other related problems even during competition. We introduce Xolver, a training-free multi-agent reasoning framework that equips a black-box LLM with a persistent, evolving memory of holistic experience. Xolver integrates diverse experience modalities, including external and self-retrieval, tool use, collaborative interactions, agent-driven evaluation, and iterative refinement. By learning from relevant strategies, code fragments, and abstract reasoning patterns at inference time, Xolver avoids generating solutions from scratch - marking a transition from isolated inference toward experience-aware language agents. Built on both open-weight and proprietary models, Xolver consistently outperforms specialized reasoning agents. Even with lightweight backbones (e.g., QWQ-32B), it often surpasses advanced models including Qwen3-235B, Gemini 2.5 Pro, o3, and o4-mini-high. With o3-mini-high, it achieves new best results on GSM8K (98.1%), AIME'24 (94.4%), AIME'25 (93.7%), Math-500 (99.8%), and LiveCodeBench-V5 (91.6%) - highlighting holistic experience learning as a key step toward generalist agents capable of expert-level reasoning. Code and data are available at https://kagnlp.github.io/xolver.github.io/.

Figures

Figures reproduced from arXiv: 2506.14234 by the authors.

Figure 1
Figure 1. Results Summary on AIME ’24 (16 runs), AIME ’25 and LiveCodeBench (32 runs). Our framework Xolver, built on o3-mini-medium and o3-mini-high backbones (denoted (m) and (h)), achieves up to 30.9% gain over the baseline and often outperforms leading models on both tasks. approach problems in a vacuum. Instead, they draw upon a rich tapestry of cumulative experiences: absorbing mentorship from coaches, developing intuit… view at source ↗
Figure 2
Figure 2. Xolver Scaffold. At each iteration, agents receive their past reasoning history and top￾ranked exemplars to generate new thoughts and responses, using tools (e.g., code) as needed. A judge model ranks outputs, and an intermediate memory maintains the best responses over time. Exemplars are initialized via episodic retrieval and continually updated with high-quality solutions from the memory. Iteration stops when con… view at source ↗
Figure 3
Figure 3. Performance drop when removing each component [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figures from the paper (8 more)
Figure 4
Figure 4. Figure 4: Impact of iterations and agents in Xolver on AIME ’25 (QWQ-32B) and LIVECODEBENCH (o3-mini-medium). To probe deeper, we conduct a budget-controlled experiment on the AIME ’25 dataset, where the total reasoning budget (i.e., num￾ber of agents × number of iter￾ations) is…
Figure 5
Figure 5. Figure 5: Impact of different retrievals in Xolver. Effect of Retrieval Strategies on Xolver Performance. We evaluate the impact of different retrieval strategies on Xolver by comparing three settings: (1) External Re￾trieval, where the model retrieves the top-k (e.g., k = 5) mo…
Figure 6
Figure 6. Figure 6: Fine-grained performance comparison in MATH-500. [PITH_FULL_IMAGE:figures/full_fig_p009_6.png]
Figure 7
Figure 7. Figure 7: Performance comparison per difficulty levels in LiveCodeBench On LiveCodeBench, Xolver demonstrates even more pro￾nounced gains. The o3-mini-medium variant achieves 95.6%, 90.4%, and 85.8% accuracy on Easy, Medium, and Hard problems respectively, significantly outperfo…
Figure 8
Figure 8. Figure 8: Avg numbers of token usage across datasets in [PITH_FULL_IMAGE:figures/full_fig_p009_8.png]
Figure 9
Figure 9. Figure 9: Xolver Math and Code error distribution. More Error Analysis in Math and Code In [PITH_FULL_IMAGE:figures/full_fig_p010_9.png]
Figure 10
Figure 10. Figure 10: Agents Accuracy and Agreement over iterations. [PITH_FULL_IMAGE:figures/full_fig_p011_10.png]
Figure 11
Figure 11. Figure 11: Coding External Retrieval Corpus Count per Subject Category. [PITH_FULL_IMAGE:figures/full_fig_p034_11.png]

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. AutoTIR: Autonomous Tools Integrated Reasoning via Reinforcement Learning

    cs.CL 2025-07 conditional novelty 5.0 of 10

    AutoTIR applies GRPO with a hand-designed action reward so a 7B instruct model learns to mix search and code tools, beating tool-using baselines on ten benchmarks.

Reference graph

Works this paper leans on

75 extracted references · 24 canonical work pages · cited by 1 Pith paper

  1. [1]

    Opencodereasoning: Advancing data distillation for competitive coding, 2025

    Wasi Uddin Ahmad, Sean Narenthiran, Somshubra Majumdar, Aleksander Ficek, Siddhartha Jain, Jocelyn Huang, Vahid Noroozi, and Boris Ginsburg. Opencodereasoning: Advancing data distillation for competitive coding, 2025. URLhttps://arxiv.org/abs/2504.01943

    arXiv 2025

  2. [2]

    Claude 3.7 Sonnet, 2025

    Anthropic. Claude 3.7 Sonnet, 2025. URL https://www.anthropic.com/news/ claude-3-7-sonnet

    work page 2025

  3. [3]

    Program synthesis with large language models.arXiv preprint arXiv:2108.07732, 2021

    Jacob Austin, Augustus Odena, Maxwell Nye, Maarten Bosma, Henryk Michalewski, David Dohan, Ellen Jiang, Carrie Cai, Michael Terry, Quoc Le, et al. Program synthesis with large language models.arXiv preprint arXiv:2108.07732, 2021

    arXiv 2021

  4. [4]

    Mark Chen, Jerry Tworek, Heewoo Jun, Qiming Yuan, Henrique Ponde de Oliveira Pinto, Jared Kaplan, Harri Edwards, Yuri Burda, Nicholas Joseph, Greg Brockman, Alex Ray, Raul Puri, Gretchen Krueger, Michael Petrov, Heidy Khlaaf, Girish Sastry, Pamela Mishkin, Brooke Chan, Scott Gray, Nick Ryder, Mikhail Pavlov, Alethea Power, Lukasz Kaiser, Mohammad Bavarian...

    2021

  5. [5]

    Agentverse: Facilitating multi-agent collaboration and exploring emergent behaviors in agents.arXiv preprint arXiv:2308.10848, 2(4):6, 2023

    Weize Chen, Yusheng Su, Jingwei Zuo, Cheng Yang, Chenfei Yuan, Chen Qian, Chi-Min Chan, Yujia Qin, Yaxi Lu, Ruobing Xie, et al. Agentverse: Facilitating multi-agent collaboration and exploring emergent behaviors in agents.arXiv preprint arXiv:2308.10848, 2(4):6, 2023

    arXiv 2023

  6. [6]

    Training verifiers to solve math word problems.arXiv preprint arXiv:2110.14168, 2021

    Karl Cobbe, Vineet Kosaraju, Mohammad Bavarian, Mark Chen, Heewoo Jun, Lukasz Kaiser, Matthias Plappert, Jerry Tworek, Jacob Hilton, Reiichiro Nakano, et al. Training verifiers to solve math word problems.arXiv preprint arXiv:2110.14168, 2021

    arXiv

  7. [7]

    Gemini 2.5, 2025

    Google DeepMind. Gemini 2.5, 2025. URL https://blog.google/technology/ google-deepmind/gemini-model-thinking-updates-march-2025/

    work page 2025

  8. [8]

    Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning, 2025

    DeepSeek-AI. Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning, 2025

    2025

Show all 75 references
  1. [9]

    Im- proving factuality and reasoning in language models through multiagent debate.arXiv preprint arXiv:2305.14325, 2023

    Yilun Du, Shuang Li, Antonio Torralba, Joshua B Tenenbaum, and Igor Mordatch. Im- proving factuality and reasoning in language models through multiagent debate.arXiv preprint arXiv:2305.14325, 2023

    2023 arXiv

  2. [10]

    When one llm drools, multi-llm collaboration rules.ArXiv, abs/2502.04506, 2025

    Shangbin Feng, Wenxuan Ding, Alisa Liu, Zifeng Wang, Weijia Shi, Yike Wang, Zejiang Shen, Xiaochuang Han, Hunter Lang, Chen-Yu Lee, Tomas Pfister, Yejin Choi, and Yulia Tsvetkov. When one llm drools, multi-llm collaboration rules.ArXiv, abs/2502.04506, 2025. URL https: //api.s...

    2025 arXiv

  3. [11]

    Pal: Program-aided language models

    Luyu Gao, Aman Madaan, Shuyan Zhou, Uri Alon, Pengfei Liu, Yiming Yang, Jamie Callan, and Graham Neubig. Pal: Program-aided language models. InInternational Conference on Machine Learning, pages 10764–10799. PMLR, 2023

    2023

  4. [12]

    rstar-math: Small llms can master math reasoning with self-evolved deep thinking, 2025

    Xinyu Guan, Li Lyna Zhang, Yifei Liu, Ning Shang, Youran Sun, Yi Zhu, Fan Yang, and Mao Yang. rstar-math: Small llms can master math reasoning with self-evolved deep thinking, 2025. URL https://arxiv.org/abs/2501.04519

    2025 arXiv

  5. [13]

    Large language model based multi-agents: A survey of progress and challenges.arXiv preprint arXiv:2402.01680, 2024

    Taicheng Guo, Xiuying Chen, Yaqi Wang, Ruidi Chang, Shichao Pei, Nitesh V Chawla, Olaf Wiest, and Xiangliang Zhang. Large language model based multi-agents: A survey of progress and challenges.arXiv preprint arXiv:2402.01680, 2024

    2024 arXiv

  6. [14]

    REALM: Retrieval-augmented language model pre-training

    Kelvin Guu, Kenton Lee, Zora Tung, Panupong Pasupat, and Ming-Wei Chang. REALM: Retrieval-augmented language model pre-training. InInternational Conference on Machine Learning. JMLR.org, 2020. URLhttps://dl.acm.org/doi/abs/10.5555/3524938.3525306. 13

    2020

  7. [15]

    Measuring mathematical problem solving with the math dataset

    Dan Hendrycks, Collin Burns, Saurav Kadavath, Akul Arora, Steven Basart, Eric Tang, Dawn Song, and Jacob Steinhardt. Measuring mathematical problem solving with the math dataset. In J. Vanschoren and S. Yeung, editors,Proceedings of the Neural Information Processing Systems Tr...

    2021

  8. [16]

    Metagpt: Meta programming for multi-agent collaborative framework.arXiv preprint arXiv:2308.00352, 3(4):6, 2023

    Sirui Hong, Xiawu Zheng, Jonathan Chen, Yuheng Cheng, Jinlin Wang, Ceyao Zhang, Zili Wang, Steven Ka Shing Yau, Zijuan Lin, Liyang Zhou, et al. Metagpt: Meta programming for multi-agent collaborative framework.arXiv preprint arXiv:2308.00352, 3(4):6, 2023

    2023 arXiv

  9. [17]

    Ashraful Islam, Mohammed Eunus Ali, and Md Rizwan Parvez

    Md. Ashraful Islam, Mohammed Eunus Ali, and Md Rizwan Parvez. MapCoder: Multi-agent code generation for competitive problem solving. In Lun-Wei Ku, Andre Martins, and Vivek Srikumar, editors,Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistic...

    2024 doi

  10. [18]

    Ashraful Islam, Mohammed Eunus Ali, and Md Rizwan Parvez

    Md. Ashraful Islam, Mohammed Eunus Ali, and Md Rizwan Parvez. CodeSim: Multi-agent code generation and problem solving through simulation-driven planning and debugging. In Luis Chiruzzo, Alan Ritter, and Lu Wang, editors,Findings of the Association for Computational Lin- guist...

    2025 doi

  11. [19]

    Open-RAG: Enhanced retrieval augmented reasoning with open-source large language models

    Shayekh Bin Islam, Md Asib Rahman, K S M Tozammel Hossain, Enamul Hoque, Shafiq Joty, and Md Rizwan Parvez. Open-RAG: Enhanced retrieval augmented reasoning with open-source large language models. In Yaser Al-Onaizan, Mohit Bansal, and Yun-Nung Chen, editors,Findings of the As...

    2024 doi

  12. [20]

    Livecodebench: Holistic and contamination free evaluation of large language models for code.arXiv preprint arXiv:2403.07974, 2024

    Naman Jain, King Han, Alex Gu, Wen-Ding Li, Fanjia Yan, Tianjun Zhang, Sida Wang, Armando Solar-Lezama, Koushik Sen, and Ion Stoica. Livecodebench: Holistic and contamination free evaluation of large language models for code.arXiv preprint arXiv:2403.07974, 2024

    2024 arXiv

  13. [21]

    XCodeEval: An execution-based large scale multilingual multitask bench- mark for code understanding, generation, translation and retrieval

    Mohammad Abdullah Matin Khan, M Saiful Bari, Xuan Long Do, Weishi Wang, Md Rizwan Parvez, and Shafiq Joty. XCodeEval: An execution-based large scale multilingual multitask bench- mark for code understanding, generation, translation and retrieval. In Lun-Wei Ku, Andre Martins, ...

  14. [22]

    Kimi k1.5: Scaling reinforcement learning with llms, 2025

    Kimi Team. Kimi k1.5: Scaling reinforcement learning with llms, 2025

    2025

  15. [23]

    A systematic survey and critical review on evaluating large language models: Challenges, limitations, and recommendations

    Md Tahmid Rahman Laskar, Sawsan Alqahtani, M Saiful Bari, Mizanur Rahman, Mohammad Abdullah Matin Khan, Haidar Khan, Israt Jahan, Amran Bhuiyan, Chee Wei Tan, Md Rizwan Parvez, Enamul Hoque, Shafiq Joty, and Jimmy Huang. A systematic survey and critical review on evaluating la...

    2024

  16. [24]

    Techniquerag: Retrieval augmented generation for adversarial technique annotation in cyber threat intelligence text,

    Ahmed Lekssays, Utsav Shukla, Husrev Taha Sencar, and Md Rizwan Parvez. Techniquerag: Retrieval augmented generation for adversarial technique annotation in cyber threat intelligence text,

  17. [25]

    Retrieval-augmented generation for knowledge-intensive nlp tasks.Advances in Neural Information Processing Systems, 33:9459–9474, 2020

    Patrick Lewis, Ethan Perez, Aleksandra Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich Küttler, Mike Lewis, Wen-tau Yih, Tim Rocktäschel, et al. Retrieval-augmented generation for knowledge-intensive nlp tasks.Advances in Neural Information Processing Systems,...

    2020

  18. [26]

    Solving quantitative reasoning problems with language models.Advances in Neural Information Processing Systems, 35: 3843–3857, 2022

    Aitor Lewkowycz, Anders Andreassen, David Dohan, Ethan Dyer, Henryk Michalewski, Vinay Ramasesh, Ambrose Slone, Cem Anil, Imanol Schlag, Theo Gutman-Solo, et al. Solving quantitative reasoning problems with language models.Advances in Neural Information Processing Systems, 35:...

    2022

  19. [27]

    Camel: Communicative agents for" mind" exploration of large language model society.Advances in Neural Information Processing Systems, 36:51991–52008, 2023

    Guohao Li, Hasan Hammoud, Hani Itani, Dmitrii Khizbullin, and Bernard Ghanem. Camel: Communicative agents for" mind" exploration of large language model society.Advances in Neural Information Processing Systems, 36:51991–52008, 2023

    2023

  20. [28]

    Search-o1: Agentic search-enhanced large reasoning models.arXiv preprint arXiv:2501.05366, 2025

    Xiaoxi Li, Guanting Dong, Jiajie Jin, Yuyao Zhang, Yujia Zhou, Yutao Zhu, Peitian Zhang, and Zhicheng Dou. Search-o1: Agentic search-enhanced large reasoning models.arXiv preprint arXiv:2501.05366, 2025

    2025 arXiv

  21. [29]

    Deepseek-v3 technical report.arXiv preprint arXiv:2412.19437, 2024

    Aixin Liu, Bei Feng, Bing Xue, Bingxuan Wang, Bochao Wu, Chengda Lu, Chenggang Zhao, Chengqi Deng, Chenyu Zhang, Chong Ruan, et al. Deepseek-v3 technical report.arXiv preprint arXiv:2412.19437, 2024

    2024 arXiv

  22. [30]

    rstar-coder: Scaling competitive code reasoning with a large-scale verified dataset.arXiv preprint arXiv:2505.21297, 2025

    Yifei Liu, Li Lyna Zhang, Yi Zhu, Bingcheng Dong, Xudong Zhou, Ning Shang, Fan Yang, and Mao Yang. rstar-coder: Scaling competitive code reasoning with a large-scale verified dataset.arXiv preprint arXiv:2505.21297, 2025

    2025 arXiv

  23. [31]

    Dynamic llm-agent network: An llm-agent collaboration framework with agent team optimization.arXiv preprint arXiv:2310.02170, 2023

    Zijun Liu, Yanzhe Zhang, Peng Li, Yang Liu, and Diyi Yang. Dynamic llm-agent network: An llm-agent collaboration framework with agent team optimization.arXiv preprint arXiv:2310.02170, 2023

    2023 arXiv

  24. [32]

    Chameleon: Plug-and-play compositional reasoning with large language models.Advances in Neural Information Processing Systems, 36, 2024

    Pan Lu, Baolin Peng, Hao Cheng, Michel Galley, Kai-Wei Chang, Ying Nian Wu, Song-Chun Zhu, and Jianfeng Gao. Chameleon: Plug-and-play compositional reasoning with large language models.Advances in Neural Information Processing Systems, 36, 2024

    2024

  25. [33]

    Octotools: An agentic framework with extensible tools for complex reasoning.arXiv preprint arXiv:2502.11271, 2025

    Pan Lu, Bowen Chen, Sheng Liu, Rahul Thapa, Joseph Boen, and James Zou. Octotools: An agentic framework with extensible tools for complex reasoning.arXiv preprint arXiv:2502.11271, 2025

    2025 arXiv

  26. [34]

    American invitational mathematics examination - aime

    MAA. American invitational mathematics examination - aime. InAmerican Invitational Mathe- matics Examination - AIME 2024, February 2024. URL https://maa.org/math-competitions/ american-invitational-mathematics-examination-aime

    2024

  27. [35]

    American invitational mathematics examination - aime

    MAA. American invitational mathematics examination - aime. InAmerican Invitational Mathe- matics Examination - AIME 2025, February 2025. URL https://maa.org/math-competitions/ american-invitational-mathematics-examination-aime

    2025

  28. [36]

    Generative reward models.arXiv preprint arXiv:2410.12832, 2024

    Dakota Mahan, Duy Van Phung, Rafael Rafailov, Chase Blagden, Nathan Lile, Louis Castricato, Jan-Philipp Fränken, Chelsea Finn, and Alon Albalak. Generative reward models.arXiv preprint arXiv:2410.12832, 2024

    2024 arXiv

  29. [37]

    Aimo-2 winning solution: Building state-of-the-art mathe- matical reasoning models with openmathreasoning dataset, 2025

    Ivan Moshkov, Darragh Hanley, Ivan Sorokin, Shubham Toshniwal, Christof Henkel, Benedikt Schifferer, Wei Du, and Igor Gitman. Aimo-2 winning solution: Building state-of-the-art mathe- matical reasoning models with openmathreasoning dataset, 2025. URL https://arxiv.org/abs/ 2504.16891

    2025 arXiv

  30. [38]

    Webgpt: Browser-assisted question-answering with human feedback.arXiv preprint arXiv:2112.09332, 2021

    Reiichiro Nakano, Jacob Hilton, Suchir Balaji, Jeff Wu, Long Ouyang, Christina Kim, Christo- pher Hesse, Shantanu Jain, Vineet Kosaraju, William Saunders, et al. Webgpt: Browser-assisted question-answering with human feedback.arXiv preprint arXiv:2112.09332, 2021

    2021 arXiv

  31. [39]

    Show your work: Scratchpads for intermediate computation with language models.arXiv preprint arXiv:2112.00114, 2021

    Maxwell Nye, Anders Johan Andreassen, Guy Gur-Ari, Henryk Michalewski, Jacob Austin, David Bieber, David Dohan, Aitor Lewkowycz, Maarten Bosma, David Luan, et al. Show your work: Scratchpads for intermediate computation with language models.arXiv preprint arXiv:2112.00114, 2021

    2021 arXiv

  32. [40]

    Hello GPT-4o, 2024

    OpenAI. Hello GPT-4o, 2024. URLhttps://openai.com/index/hello-gpt-4o/

    2024

  33. [41]

    Learning to reason with llms, 2024

    OpenAI. Learning to reason with llms, 2024. URL https://openai.com/index/ learning-to-reason-with-llms/. 15

    2024

  34. [42]

    Introducing openai o3 and o4-mini, 2025

    OpenAI. Introducing openai o3 and o4-mini, 2025. URL https://openai.com/index/ introducing-o3-and-o4-mini/

    2025

  35. [43]

    Patil, Kevin Lin, Sarah Wooders, and Joseph Gonzalez

    Charles Packer, Vivian Fang, Shishir G. Patil, Kevin Lin, Sarah Wooders, and Joseph Gonzalez. Memgpt: Towards llms as operating systems.ArXiv, abs/2310.08560, 2023. URL https://api. semanticscholar.org/CorpusID:263909014

    2023 arXiv

  36. [44]

    Chain of evidences and evidence to generate: Prompting for context grounded and retrieval augmented reasoning

    Md Rizwan Parvez. Chain of evidences and evidence to generate: Prompting for context grounded and retrieval augmented reasoning. In Weijia Shi, Wenhao Yu, Akari Asai, Meng Jiang, Greg Durrett, Hannaneh Hajishirzi, and Luke Zettlemoyer, editors,Proceedings of the 4th Internatio...

    2025 doi

  37. [45]

    Evaluating the values of sources in transfer learning

    Md Rizwan Parvez and Kai-Wei Chang. Evaluating the values of sources in transfer learning. In Kristina Toutanova, Anna Rumshisky, Luke Zettlemoyer, Dilek Hakkani-Tur, Iz Beltagy, Steven Bethard, Ryan Cotterell, Tanmoy Chakraborty, and Yichao Zhou, editors,Proceedings of the 20...

    2021 doi

  38. [46]

    Retrieval augmented code generation and summarization

    Md Rizwan Parvez, Wasi Ahmad, Saikat Chakraborty, Baishakhi Ray, and Kai-Wei Chang. Retrieval augmented code generation and summarization. In Marie-Francine Moens, Xuanjing Huang, Lucia Specia, and Scott Wen-tau Yih, editors,Findings of the Association for Computational Lingui...

    2021 doi

  39. [47]

    Retrieval enhanced data augmentation for question answering on privacy policies

    Md Rizwan Parvez, Jianfeng Chi, Wasi Uddin Ahmad, Yuan Tian, and Kai-Wei Chang. Retrieval enhanced data augmentation for question answering on privacy policies. In Andreas Vlachos and Isabelle Augenstein, editors,Proceedings of the 17th Conference of the European Chapter of th...

    2023 doi

  40. [48]

    Qwq: Reflect deeply on the boundaries of the unknown, 2024

    Qwen Team. Qwq: Reflect deeply on the boundaries of the unknown, 2024. URL https: //qwenlm.github.io/blog/qwq-32b-preview/

    2024

  41. [49]

    QwQ-32B: Embracing the power of reinforcement learning, March 2025

    Qwen Team. QwQ-32B: Embracing the power of reinforcement learning, March 2025. URL https://qwenlm.github.io/blog/qwq-32b/

    2025

  42. [50]

    Ai debate aids assessment of controversial claims.arXiv preprint arXiv:2506.02175, 2025

    Salman Rahman, Sheriff Issaka, Ashima Suvarna, Genglin Liu, James Shiffer, Jaeyoung Lee, Md Rizwan Parvez, Hamid Palangi, Shi Feng, Nanyun Peng, et al. Ai debate aids assessment of controversial claims.arXiv preprint arXiv:2506.02175, 2025

    2025

  43. [51]

    DelucionQA: Detecting hallucinations in domain-specific question answering

    Mobashir Sadat, Zhengyu Zhou, Lukas Lange, Jun Araki, Arsalan Gundroo, Bingqing Wang, Rakesh Menon, Md Parvez, and Zhe Feng. DelucionQA: Detecting hallucinations in domain-specific question answering. In Houda Bouamor, Juan Pino, and Kalika Bali, editors,Findings of the Associ...

    2023

  44. [52]

    Reflexion: Language agents with verbal reinforcement learning

    Noah Shinn, Federico Cassano, Ashwin Gopinath, Karthik R Narasimhan, and Shunyu Yao. Reflexion: Language agents with verbal reinforcement learning. InThirty-seventh Conference on Neural Information Processing Systems, 2023

    2023

  45. [53]

    Reflexion: Language agents with verbal reinforcement learning.Advances in Neural Information Processing Systems, 36, 2024

    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, 2024. 16

    2024

  46. [54]

    Tenenbaum, Antonio Torralba, Shuang Li, and Igor Mordatch

    Vighnesh Subramaniam, Yilun Du, Joshua B. Tenenbaum, Antonio Torralba, Shuang Li, and Igor Mordatch. Multiagent finetuning: Self improvement with diverse reasoning chains. InThe Thirteenth International Conference on Learning Representations, 2025. URL https://openreview.net/ ...

    2025

  47. [55]

    Dynamic cheatsheet: Test-time learning with adaptive memory.arXiv preprint arXiv:2504.07952, 2025

    Mirac Suzgun, Mert Yuksekgonul, Federico Bianchi, Dan Jurafsky, and James Zou. Dynamic cheatsheet: Test-time learning with adaptive memory.arXiv preprint arXiv:2504.07952, 2025

    2025 arXiv

  48. [56]

    Qwen3, April 2025

    Qwen Team. Qwen3, April 2025. URLhttps://qwenlm.github.io/blog/qwen3/

    2025

  49. [57]

    Openmathinstruct-1: A 1.8 million math instruction tuning dataset.arXiv preprint arXiv:2402.10176, 2024

    Shubham Toshniwal, Ivan Moshkov, Sean Narenthiran, Daria Gitman, Fei Jia, and Igor Gitman. Openmathinstruct-1: A 1.8 million math instruction tuning dataset.arXiv preprint arXiv:2402.10176, 2024

    2024 arXiv

  50. [58]

    Self-consistency improves chain of thought reasoning in language models

    Xuezhi Wang, Jason Wei, Dale Schuurmans, Quoc V Le, Ed H Chi, Sharan Narang, Aakanksha Chowdhery, and Denny Zhou. Self-consistency improves chain of thought reasoning in language models. InThe Eleventh International Conference on Learning Representations

  51. [59]

    Self-consistency improves chain of thought reasoning in language models.ArXiv, abs/2203.11171, 2022

    Xuezhi Wang, Jason Wei, Dale Schuurmans, Quoc Le, Ed Huai hsin Chi, and Denny Zhou. Self-consistency improves chain of thought reasoning in language models.ArXiv, abs/2203.11171, 2022

    2022 arXiv

  52. [60]

    Learning to filter context for retrieval-augmented generation.arXiv preprint arXiv:2311.08377, 2023

    Zhiruo Wang, Jun Araki, Zhengbao Jiang, Md Rizwan Parvez, and Graham Neubig. Learning to filter context for retrieval-augmented generation.arXiv preprint arXiv:2311.08377, 2023

    2023 arXiv

  53. [61]

    Chain-of-thought prompting elicits reasoning in large language models.Advances in neural information processing systems, 35:24824–24837, 2022

    Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. Chain-of-thought prompting elicits reasoning in large language models.Advances in neural information processing systems, 35:24824–24837, 2022

    2022

  54. [62]

    Autogen: Enabling next-gen llm applications via multi-agent conversation framework.arXiv preprint arXiv:2308.08155, 3(4), 2023

    Qingyun Wu, Gagan Bansal, Jieyu Zhang, Yiran Wu, Shaokun Zhang, Erkang Zhu, Beibin Li, Li Jiang, Xiaoyun Zhang, and Chi Wang. Autogen: Enabling next-gen llm applications via multi-agent conversation framework.arXiv preprint arXiv:2308.08155, 3(4), 2023

    2023 arXiv

  55. [63]

    Grok 3 beta — the age of reasoning agents, 2025

    xAI. Grok 3 beta — the age of reasoning agents, 2025. URLhttps://x.ai/news/grok-3

    2025

  56. [64]

    Tree of thoughts: Deliberate problem solving with large language models.arXiv preprint arXiv:2305.10601, 2023

    Shunyu Yao, Dian Yu, Jeffrey Zhao, Izhak Shafran, Thomas L Griffiths, Yuan Cao, and Karthik Narasimhan. Tree of thoughts: Deliberate problem solving with large language models.arXiv preprint arXiv:2305.10601, 2023

    2023 arXiv

  57. [65]

    React: Synergizing reasoning and acting in language models

    Shunyu Yao, Jeffrey Zhao, Dian Yu, Nan Du, Izhak Shafran, Karthik Narasimhan, and Yuan Cao. React: Synergizing reasoning and acting in language models. InInternational Conference on Learning Representations (ICLR), 2023

    2023

  58. [66]

    Retroformer: Retrospective large language agents with policy gradient op- timization.ArXiv, abs/2308.02151, 2023

    Weiran Yao, Shelby Heinecke, Juan Carlos Niebles, Zhiwei Liu, Yihao Feng, Le Xue, Rithesh Murthy, Zeyuan Chen, Jianguo Zhang, Devansh Arpit, Ran Xu, Phil L Mui, Haiquan Wang, Caiming Xiong, and Silvio Savarese. Retroformer: Retrospective large language agents with policy gradi...

    2023 arXiv

  59. [67]

    Acecoder: Acing coder rl via automated test-case synthesis.arXiv preprint arXiv:2502.01718, 2025

    Huaye Zeng, Dongfu Jiang, Haozhe Wang, Ping Nie, Xiaotong Chen, and Wenhu Chen. Acecoder: Acing coder rl via automated test-case synthesis.arXiv preprint arXiv:2502.01718, 2025

    2025 arXiv

  60. [68]

    Alphaone: Reasoning models thinking slow and fast at test time

    Junyu Zhang, Runpei Dong, Han Wang, Xuying Ning, Haoran Geng, Peihao Li, Xialin He, Yutong Bai, Jitendra Malik, Saurabh Gupta, and Huan Zhang. Alphaone: Reasoning models thinking slow and fast at test time. 2025. URLhttps://arxiv.org/abs/2505.24863

    2025 arXiv

  61. [69]

    Liu, and Gao Huang

    Andrew Zhao, Daniel Huang, Quentin Xu, Matthieu Gaetan Lin, Y . Liu, and Gao Huang. Expel: Llm agents are experiential learners. InAAAI Conference on Artificial Intelligence, 2023. URL https://api.semanticscholar.org/CorpusID:261048772

    2023

  62. [70]

    Ldb: A large language model debugger via verifying runtime execution step-by-step.arXiv preprint arXiv:2402.16906, 2024

    Li Zhong, Zilong Wang, and Jingbo Shang. Ldb: A large language model debugger via verifying runtime execution step-by-step.arXiv preprint arXiv:2402.16906, 2024. 17 A Lists of Prompts 19 A.1 Planner Agent . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ...

    2024 arXiv

  63. [75]

    twice as many pillow cases as sheets

    of the frameworkXolver. A.1 Planner Agent Prompt for PLANNERAGENT You are a planner to solve a {coding/math} problem. Here is the problem for which you have to plan: {problem_dict[query[‘problem_id’]][‘description’]} First draft required strictly greater than {m} specialized r...

  64. [831]

    URLhttps://aclanthology.org/2024.findings-emnlp.831/

    2024

  65. [2023]

    doi: 10.18653/v1/2023.findings-emnlp.59

    Association for Computational Linguistics. doi: 10.18653/v1/2023.findings-emnlp.59. URL https://aclanthology.org/2023.findings-emnlp.59/

    2023 doi

  66. [2024]

    doi: 10.18653/v1/2024.acl-long.367

    Association for Computational Linguistics. doi: 10.18653/v1/2024.acl-long.367. URL https://aclanthology.org/2024.acl-long.367/

    2024 doi

  67. [2025]

    URLhttps://arxiv.org/abs/2505.11988

    arXiv

Pith tools

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