acceptodds
Under review as a conference paper at ICLR 2027

Does Recursive Knowledge Construction Help Agents Verify Consensus Claims in Science?

Abstract

Recent scientific agents build knowledge recursively by retrieving evidence, constructing an intermediate state, and judging whether that state contains sufficient information. These agents are evaluated end-to-end, usually without ground truth, so the effect of each design choice is unknown. We examine two design axes: (1) what the recursion builds (the state of evidence) and (2) what controls it (the judge that decides when to stop). To measure them, we define consensus scientific claim verification and measure these effects on this task, where a system must establish a claim's consensus stance from the open literature. We build a new consensus claim dataset, ProClaim-eval, including 718 claims with SUPPORT and REFUTE gold labels from five expert-curated databases in molecular biology, nuclear physics, and astronomy. Our instrument, ProClaim, persists an evidence state outside the model's context, so each axis can be studied with the other held fixed. Against verification systems and general-purpose agent harnesses, ProClaim ranks among the top three, with a significant advantage over the harnesses. We find that the evidence state improves performance over in-context accumulation, and it also matters what is stored in it. In addition, a trained judge that reads the evidence state performs better than fixed-round stopping. A judge trained separately and never exposed to ProClaim-eval achieves the highest agreement with the gold labels on three of five subsets, where it matches the off-the-shelf Claude Sonnet 4.6 judge and outperforms Claude Haiku 4.5. Our results on ProClaim-eval support recursively maintaining a structured evidence state outside the model's context as an effective design for verifying claims in scientific agents.

Primary area:
foundation or frontier models, including LLMs
Keywords:
LLM Agents, Scientific Claim Verification

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_eFDSOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper asks whether the loop that most scientific agents now follow (retrieve, update some intermediate record, decide whether to stop) actually helps, and if so, which part helps. The authors split the loop into two knobs: what the recursion builds (the evidence state) and what controls it (the stopping judge). They need a task where the loop matters, so they define "consensus claim verification": given a single-sentence claim, decide its stance with respect to the open literature, not a single paired passage. To evaluate this, they build ProClaim-eval, 718 claims drawn from five curated databases. Most REFUTE labels come from curator decisions such as removed edges, rejected entries, refuted planets and disproved isomeric states.

Critical strengths

The benchmark is the real contribution, and I think it will outlast the method. Getting claim-level labels whose REFUTE side comes from experts is hard, and the authors found a sensible way to do it: track what curators removed or rejected between releases, and use archives that explicitly record refutations. It spans three fields, and the Figure 3 experiment makes a fair case that reading one reference abstract is not enough. The serialized release to keep it out of pretraining crawls is a nice touch. Even if every method claim in the paper had to be softened, I would still want this dataset to exist, and that carries a lot of weight in my assessment.

Critical weaknesses
  1. The comparisons don't isolate the evidence state, which is the variable the paper is about. The central claim is that keeping state outside the context window beats accumulating evidence in context. But Appendix D.1 says the full execution log is sent to the model at every step, and C.4 confirms ProClaim re-sends its accumulated history. Table 7 shows ProClaim's main agent consuming roughly 610K to 1.16M input tokens per claim, against 82K to 261K for ReAct. So ProClaim's context grows more than ReAct's, not less. The real difference is what sits in the context: printed state summaries and stance-labelled facts in one case, raw paper text in the other. That is a perfectly interesting result, but it is the "what it stores" half of Finding 2, not the "outside the context" half. As far as I can tell, the paper has no experiment that tests the second half on its own.
  2. The labels don't quite mean "consensus," and the ablation evidence is statistically fragile. I like the dataset, but I'm not convinced its labels measure consensus in the sense the title suggests. On top of that, the verdict definition in B.1 counts "a thorough search finds nothing" as REFUTE, while SUPPORT often requires two to five papers. That rewards a "can't find it, so it's false" policy, and it makes the task partly a test of retrieval recall. It also weakens the Figure 3 argument. For a REFUTE claim, the reference abstract is the very paper the database later decided was wrong, so of course reading it alone produces poor agreement.
Questions for authors
  1. Is the final verdict produced from the state file or from the accumulated log? Could you run a variant where the agent sees only the current state snapshot and not the history? This would directly test the "outside the context" claim.
  2. Which tools and access did the ReAct controls and the harnesses have? Specifically, did they get full-text APIs, S2 recommendations, gap search, the three-paper minimum, and identical label definitions? And which backbone did Codex use? §4 says every agentic system runs on Sonnet 4.6, which is unusual for Codex.
  3. Are the 40-claim results from the same run as Table 2, or from a rerun?
5 comments
AuthorsOpenReview ↗

Clarification and results on the evidence state

Comment eFDS-C1 (Clarification and results on the evidence state):

The comparisons don't isolate the evidence state … The real difference is what sits in the context: printed state summaries and stance-labelled facts in one case, raw paper text in the other. That is a perfectly interesting result, but it is the "what it stores" half of Finding 2, not the "outside the context" half. Is the final verdict produced from the state file or from the accumulated log? Could you run a variant where the agent sees only the current state snapshot and not the history? This would directly test the "outside the context" claim.

We thank the reviewer for raising these comments and questions. Since weakness 1 and question 1 are related, we explain them together.

Verdicts are produced from the log. ProClaim writes the verdict from its execution log. As a safeguard, the verdict is derived by the same LLM from state only when the agent fails to give a verdict from the log (e.g., when it is hard stopped by allocated budget and has to be reprompted).

Storing evidence "outside the context" gives higher agreement. We agree with the reviewer that “outside the context” half of our claim deserves to be tested. In Section 4.2, we ran the suggested experiment variant on the same 200 claims (same 40-claim samples from each dataset subset), where each call sees the current evidence state JSON and only the previous step, with no earlier history. The previous step is needed to avoid tool calling loops. We also removed full text from the JSON state to avoid overflowing the context.

Pooled over the 200 claims, ProClaim with the state in-context (0.69) has lower agreement than ProClaim (0.77, p = 0.049), with paired exact McNemar test. This shows that the evidence state should be kept outside the context.

MethodSIGNORConnectomeDBExTRINUBASENASA
ProClaim with state in-context (Claude Sonnet 4.6)0.800.930.750.350.63
ProClaim (Claude Sonnet 4.6)0.830.950.850.500.73

Are the 40-claim results from the same run as Table 2, or from a rerun?

Yes, Table 3 rows come from the Table 2 runs, restricted to the 40 claims randomly sampled from each domain subset. In Table 4, the trained-judge row comes from the Table 2 runs, and the other stopping rules are separate runs on the same 40 claims, because ablating the judge requires new ProClaim runs.

We clarified this point in the “setting” paragraph of Section 4.2.

AuthorsOpenReview ↗

Clarification on the consensus claim dataset

Comment eFDS-C2 (Clarification on the consensus claim dataset):

I like the dataset, but I'm not convinced its labels measure consensus in the sense the title suggests.

We appreciate the reviewer's thoughtful comments.

We want to clarify how each database represents domain consensus. For example, in the SIGNOR subset, our re-curation guideline explicitly instructed curators to examine multiple publications, and the final verdict was decided by majority voting among curators. Curators in the ConnectomeDB 2025 release examined more than one paper and documented a consensual justification per entry. The ExTRI corpus was read by at least two domain experts, with disagreements resolved to consensus. In NUBASE, evaluators weighed published measurements and flagged states that were later shown not to hold. The NASA Exoplanet Archive maintains curated dispositions of entries, which are periodically examined by domain experts from the peer-reviewed literature. We have explained how the ProClaim-eval subsets relate to domain consensus in Appendix B. 7.

Therefore, we consider the databases the best proxy for consensus these fields maintain, which others already rely on for knowledge-intensive work, for example, as prior knowledge for downstream bioinformatics analysis [1,2]. This is highlighted in Section 1.

On top of that, the verdict definition in B.1 counts "a thorough search finds nothing" as REFUTE, while SUPPORT often requires two to five papers. That rewards a "can't find it, so it's false" policy, and it makes the task partly a test of retrieval recall.

We agree that the "can't find it, so it's false" policy lets poor retrieval get correct REFUTE answers. However, this policy would be penalised by the SUPPORT class. A label states whether the claim is supported as consensus, and we design it to carry the burden of proof (Appendix. B.1). In Figure 3, the naive top-5 retrieval does not guarantee finding the right papers and is a proxy for this policy. It reaches 92% on gold REFUTE but only 19% on Gold SUPPORT. Therefore, a naive policy would not perform well on our dataset.

It also weakens the Figure 3 argument. For a REFUTE claim, the reference abstract is the very paper the database later decided was wrong, so of course reading it alone produces poor agreement.

We further clarify this in our explanation of Figure 3 in Section 2. Having the reference abstract gives high agreement (0.82) in the REFUTE class, which is expected because the abstract is wrong. The low agreement (0.43) actually comes from the SUPPORT class, where the abstract is supposed to be correct. Therefore, a single-reference-guided verification does not predict consensus, which is what Figure 3 aims to show.

AuthorsOpenReview ↗

Clarification on the shared tooling

Comment eFDS-C3 (Clarification on the shared tooling):

Which tools and access did the ReAct controls and the harnesses have? Specifically, did they get full-text APIs, S2 recommendations, gap search, the three-paper minimum, and identical label definitions? And which backbone did Codex use? §4 says every agentic system runs on Sonnet 4.6, which is unusual for Codex.

We thank the reviewer for these questions. We clarified tools and access per system in Appendix C.4.

Shared components. Every system searched PubMed and Semantic Scholar (S2 only on NUBASE / NASA), with web search off (to avoid database leakage), and received the same label definitions. Static retrieval only used Semantic Scholar. The three-paper minimum rule also applied to the LLM judges in Table 4.

ProClaim-specific components.

  • Full text. Giving it to the agents is not viable in Table 2. For example, ReAct re-sends its whole conversation each step (its total input grows quadratically) with full-text-sized documents would raise input from 127K to 0.8–1.2M tokens per claim, and 22% of claims would exceed a 200K window. In ProClaim, a local extractor reads full text and only facts enter the main context.
  • Gap-targeted retrieval and S2 recommendations are part of ProClaim's implementation, not the subject of study, and so the baselines did not get them. Our separate analysis suggests they do not affect ProClaim’s performance: runs with gap-targeted retrieval do not perform better than runs without; S2 recommendations only supply < 1% of final-state facts. Both have negligible performance effects.

Codex backbone runs on Sonnet 4.6. In Table 2, the intention is to let Claude Code, Codex and ProClaim share one backbone, with only the harness varying. Codex uses only OpenAI's API, so a small local bridge forwards its requests to Sonnet 4.6 (Vertex AI, via LiteLLM) and returns OpenAI-format replies.

Reviewer_xFR8OpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies two design choices in scientific claim verification: how an agent stores accumulated evidence and how it decides when to stop searching. It introduces ProClaim-eval, 718 claims from five curated scientific databases, and ProClaim, an agent with an external evidence state and a replaceable sufficiency judge. Experiments compare evidence representations and stopping rules, reporting gains from extracted facts and promising transfer of a small trained judge.

Critical strengths
  • Useful Resource. Benchmarks for scientific claim verification have become increasingly important with the emergence of agentic AI for scientific tasks.
  • Substantial Gains. The Proclaim agent achieves substantial gains on proclaim-eval.
Critical weaknesses
  • Limited Benchmark Novelty and Complexity. The proposed benchmarks lack sufficient novelty and complexity. They rely on a biased selection of three disciplines—molecular biology, nuclear physics, and astronomy—and primarily feature claims requiring only single-hop reasoning rather than more complex, multi-hop reasoning.
  • Outdated Experimental Evaluation. The experiments use models released in mid-to-late 2025, such as Gemini 2.5 Pro. It is unclear whether the findings would hold when evaluated using more recent LLMs.
  • Weak Presentation. The formatting in Table 2, particularly the notation for relative score changes, is unconventional and difficult to interpret. Similarly, the rationale behind the visualization in Figure 3 is unclear.
Questions for authors

N/A

4 comments
AuthorsOpenReview ↗

Clarification on benchmark novelty and complexity

Comment xFR8-C1 (Clarification on benchmark novelty and complexity):

The proposed benchmarks lack sufficient novelty and complexity. They rely on a biased selection of three disciplines—molecular biology, nuclear physics, and astronomy—and primarily feature claims requiring only single-hop reasoning rather than more complex, multi-hop reasoning.

We appreciate the reviewer for recognising the usefulness of our benchmark dataset. However, we want to clarify the reviewer’s points on the novelty and complexity of our benchmark:

ProClaim-eval targets a gap in the literature. As discussed in Section 5, prior claim verification datasets provide claim-level labels for open-literature verification.

The benchmark is challenging. Sonnet 4.6 and Opus 5.5 are struggling on our ProClaim-eval dataset. In Table 2, without retrieval, Sonnet 4.6 reaches only 0.21 / 0.05 / 0.07 agreement with labels on ExTRI / NUBASE / NASA. Opus 5.5 achieved only 0.46 and 0.44 on ExTRI and NASA.

ProClaim-eval requires multi-hop. In Figure 3, our motivational analysis shows that single-hop reading of the single reference abstract and five retrieved abstracts gives 0.60 and 0.52 agreement, respectively, while ProClaim's multi-hop verification reaches 0.79 on the same set of claims.

Criteria for selecting databases. The database selection follows requirements that are implicit in Section 2 and Appendix B, and we have stated them explicitly in Section 2. Our proposed consensus claim verification task requires a source to be (i) actively curated by domain experts, (ii) record entry removals, rejections, retractions, and not just new additions. The five databases (SIGNOR, ConnectomeDB, ExTRI, NUBASE, NASA) satisfy these criteria.

Further, we would argue that benchmark datasets covering biology, physics, and astronomy can be considered broad-reaching.

AuthorsOpenReview ↗

Experimental results with newer model

Comment xFR8-C2 (Experimental results with newer model):

The experiments use models released in mid-to-late 2025, such as Gemini 2.5 Pro. It is unclear whether the findings would hold when evaluated using more recent LLMs.

We appreciate the reviewer’s comment on the LLM models used in the paper. We used mainly Claude Sonnet 4.6, which was released in February 2026.

In addition, we added Claude Opus 5.5 as a second backbone and reran the experiments behind Findings 1–3 (Tables 2–4, new rows in blue). Opus 5.5 runs with its default adaptive thinking, because its API does not allow disabling thinking or setting a temperature (Appendix C).

A stronger backbone does not saturate ProClaim-eval. Without retrieval, Opus 5.5 reaches below 0.5 agreement with labels on ExTRI (0.46) and NASA (0.44) subsets, while achieving the best performance on SIGNOR (0.86) and NUBASE (0.62). The difference in domain knowledge that the model holds is larger for Opus 5.5 than in Sonnet 4.6.

Finding 1 holds with Opus 5.5. ProClaim has the highest agreement on ConnectomeDB, ExTRI and NASA, and the highest pooled agreement with labels (0.82 vs. 0.78, p = 0.043).

Method (Claude Opus 5.5 backbone)SIGNORConnectomeDBExTRINUBASENASA
LLM only0.860.830.460.620.44
Static retrieval, S2 top 50.630.500.620.120.38
FIRE0.790.650.790.140.76
Claude Code0.800.830.700.610.75
ProClaim0.790.840.800.560.77

Finding 2 holds in part. The evidence state still improves performance over in-context accumulation. However, extract facts only benefit the weaker Sonnet 4.6 backbone. We revised Finding 2 to state that the facts are a better input for a weaker model.

Evidence & methods (Claude Opus 5.5 backbone)SIGNORConnectomeDBExTRINUBASENASA
No retrieval0.850.930.500.600.38
Abstracts0.830.900.930.350.78
Full text0.830.900.950.400.80
Facts0.800.850.930.450.88
Facts + abstracts0.830.930.930.430.75
ProClaim0.780.930.950.630.80

Finding 3 holds with Opus 5.5. With Opus 5.5 as the backbone in ProClaim, the trained judge, fitted only on SciFact-Open, has higher agreement than Sonnet 4.6 (0.82 vs. 0.70, p < 0.001) and Opus 5.5 (0.82 vs. 0.75, p = 0.0072) (Table 4).

Judge (Claude Opus 5.5 backbone)SIGNORConnectomeDBExTRINUBASENASA
Claude Sonnet 4.6 judge0.700.950.830.200.83
Claude Opus 5.5 judge0.780.930.830.400.83
Trained judge0.780.930.950.630.80

We updated the shared setting (Section 4), Tables 2–5 and 7, and Findings 1–3 accordingly.

AuthorsOpenReview ↗

Table and figure presentation

Comment xFR8-C3 (Table and figure presentation):

The formatting in Table 2, particularly the notation for relative score changes, is unconventional and difficult to interpret. Similarly, the rationale behind the visualization in Figure 3 is unclear.

We thank the reviewer for suggestions on improving the presentation of the table and Figure 3.

Removed relative scores in Table 2. We removed the sharded relative scores in Table 2 that showed the performance gap between ProClaim and the compared methods.

Figure 3 shows the task novelty and the need for recursive knowledge construction. Figure 3 puts document guide verification with expert reference abstract in the context of consensus verification, with 60.3% agreement between verdict predictions and labels. Figure 3 also illustrates that consensus verification cannot be reduced to naive retrieval, and recursive knowledge construction instead produces stronger alignment with the SUPPORT labels and raises overall agreement to 78.8%.

Clarified Figure 3 and descriptions. We re-organised the explanation paragraph of figure 3 with blue highlights and updated the Table 2 layout to improve presentation.

Dear Reviewer xFR8,

We would like to thank you for your constructive comments again. With the discussion period ending on 18th Nov, we want to confirm whether our responses have addressed your three points: the benchmark’s novelty and complexity (xFR8-C1), the additional results with a newer model (xFR8-C2), and the revised presentation of Table 2 and Figure 3 (xFR8-C3).

We are happy to provide any further clarification in the remaining time, and would appreciate it if you could revisit your score in light of the responses.

Best regards,

The Authors

Reviewer_AeRXOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper defines a new dataset ProClaim-eval for claim verification. The dataset, is intended to cover consensus, as opposed to local, claim support, but remains limited to easily decontextualised claims that it aggregates from a set of existing claims corpora.

The system built on this dataset aggregates evidence in a json records, read by per-evaluation python code. This enables ready separation of the evidence access and decision components. This setup enables ProClaim to achieve generally better performance than unaided models. Given the evidence storage and access means, a simple classifier suffices to make verification judgements.

Consensus claim evaluation does not have an existing baseline, so unaided general agent models (codex, claude code), and static and adaptive IR during looped classification attempts are compared with ProClaim.

Critical strengths

The paper introduces an initial approach to consensus claim verification, and generates a small corpus that can be used to drive future work.

The paper show that evidence assembly is effective in supporting good performance agains this corpus.

Critical weaknesses

The corpus assembled only covers claims that are almost decontextualised on creation (protein bindings, nuclide status) and not complex scientific claims for which consensus is at all difficult to obtain. This is, of course, a broader problem - identifying exactly what constitutes a scientific claim. The limitation on the complexity of the claims being analysed weakens the downstream likely impact of the success of ProClaim, since it's not clear whether its straightforward evidence assembly or effectiveness of simple classification findings would extend to even slightly broader domains.

It's not entirely clear why the term "recursive" is used here instead of "iterative".

Questions for authors

Suppose that you wanted to extend the ProClaim-eval corpus and your classifier to more complex claims, like those in your paper (e.g. "A lightweight trained judge can transfer and replace off-the-shelf LLM judges". How would you sufficiently decontextualise such claims, and, after decontextualisation, would they be subject to consensus?

Given the very low level of contextual ambiguity in the corpora you have aggregated, how to you see your paper fitting into a more ambitious research agenda that, for example, could be applied to AI research effectively?

4 comments
AuthorsOpenReview ↗

Limitation of claim complexity

Comment AeRX-C1 (Limitation of claim complexity)

The limitation on the complexity of the claims being analysed weakens the downstream likely impact of the success of ProClaim, since it's not clear whether its straightforward evidence assembly or effectiveness of simple classification findings would extend to even slightly broader domains.

We agree with the reviewer that the format of ProClaim-eval claims may appear simple. However, these claims came from recognised and widely used databases. For example, SIGNOR and ExTRI are standard sources of prior knowledge for downstream bioinformatics analyses [1,2]. We clarified the impact of our claim dataset and these databases in Section 1 and 6.

We want to highlight that, to our knowledge, no prior work has studied consensus scientific claim verification, even for decontextualised claims. ProClaim-eval by design follows the atomic factual claims of the SciFact line of work [3], and expert curation has already removed most of the context. We agree that what constitutes a scientific claim is a broader problem, noted in our Section 6 and Appendix A revision.

Therefore, our work lays the foundation for two contextualised extensions we see for future work, each requiring its own dataset beyond the scope of this paper. The first should transfer more directly, and the second requires refining the harness:

  • Compound claims require subclaim decomposition. ProClaim already decomposes claims, but often too shallowly given the atomic claims in ProClaim-eval. Prior work that studied claim decomposition specifically outside of a scientific setting was [4], which we mentioned in Section 5.
  • Context-bound claims require aggregation of results across studies in related contexts, such as datasets and metrics. For these claims, their context is part of what they assert, such as claims about AI research empirical findings. ProClaim already has the evidence state as a storage for additional context such as datasets or metrics, but the harness needs to be refined for this new purpose.

We clarified these extensions in Section 6.

AuthorsOpenReview ↗

Claim contextualisation in AI research

Comment AeRX-C2 (Claim contextualisation in AI research)

Suppose that you wanted to extend the ProClaim-eval corpus and your classifier to more complex claims, like those in your paper (e.g. "A lightweight trained judge can transfer and replace off-the-shelf LLM judges". How would you sufficiently decontextualise such claims, and, after decontextualisation, would they be subject to consensus?

We thank the reviewer for raising this question. As mentioned above, decontextualising claims like the ones in our paper may not give general consensus. The research agenda we foresee from our work is to explore claims that are situated in papers’ own datasets and metrics, which we call context-bound claims in the above response. However, contextualised claims are outside the scope of our current work.

Given the very low level of contextual ambiguity in the corpora you have aggregated, how to you see your paper fitting into a more ambitious research agenda that, for example, could be applied to AI research effectively?

We appreciate this comment on how we see consensus verification can progress. As we discussed previously, claims from AI research may need to retain their context. For claim datasets, prior work such as ClaimCheck [5] considered annotated weaknesses in AI research scientific claims from rejected NeurIPS papers. Other efforts such as [6] and the Hugging Face ICML paper reproduction hackathon aimed to reproduce claims of individual papers. However, the gap remains the problem that ProClaim-eval aims to address - there is no claim-level consensus. We see this gap as an even more interesting challenge since claim datasets for AI research do not supply consensus labels. We discussed this in Section 6 for future work.

AuthorsOpenReview ↗

Clarification on the recursive knowledge construction

Comment AeRX-C3 (Clarification on the recursive knowledge construction):

It's not entirely clear why the term "recursive" is used here instead of "iterative".

We thank the reviewer for raising this. “Iterative" only says that a step repeats, and in Section 3.1, we use "recursive" in the sense of a recurrence, where the state the pipeline outputs at each step is the input that defines the next step. Writing one step as st+1=U(s,R(c,s))s_{t+1} = U(s, R(c, s)), the final state is sT=U(sT−1,R(c,sT−1))=U(U(sT−2,R(c,sT−2)),R(c,U(sT−2,R(c,sT−2))))=⋯s_T = U(s_{T-1}, R(c, s_{T-1})) = U(U(s_{T-2}, R(c, s_{T-2})), R(c, U(s_{T-2}, R(c, s_{T-2})))) = \cdots, where cc is a claim sts_{t} is the evidence state at step tt, U and R are retrieval and update operations.

This is the same conceptual framing from the Recursive Language Model work [7], and we differ in that the recursion operates on the constructed knowledge rather than calling the model itself.

We clarified the recursion notation and Recursive Language Model inspiration in Section 3.1.

References:

[1] Müller-Dott et al. Expanding the coverage of regulons from high-confidence prior knowledge for accurate estimation of transcription factor activities. Nucleic Acids Res., 2023.

[2] Türei et al. Integrated intra- and intercellular signaling knowledge for multicellular omics analysis. Mol. Syst. Biol., 2021.

[3] Wadden et al. Fact or Fiction: Verifying Scientific Claims. EMNLP 2020.

[4] Wei et al. Long-form factuality in large language models. NeurIPS 2024.

[5] Ou et al. CLAIMCHECK: How Grounded are LLM Critiques of Scientific Papers? Findings of EMNLP 2025.

[6] Miao et al. Reimagining research papers as interactive and reliable AI agents. Nature, 2026.

[7] Zhang et al. Recursive Language Models. arXiv:2512.24601, 2025.

Dear Reviewer AeRX,

Thank you again for your review. As the discussion period will close soon, we wanted to check whether our responses addressed your concerns, in particular, on the complexity of ProClaim-eval claims (AeRX-C1) and how we consider contextualisation within a broader agenda in AI research (AeRX-C2).

We have also committed to the manuscript changes listed in our response, including the explicit future work statement and the footnote on “recursive”. If anything remains unclear, we are happy to answer further questions before the deadline.

If our responses have resolved your concerns, we would be grateful if you would consider updating your assessment.

Best regards, The Authors

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.