acceptodds
Under review as a conference paper at ICLR 2027

From Prompts to Posteriors: Can LLM Agents Perform Bayesian Inference in Context?

Abstract

Large language models (LLMs) are increasingly applied across the scientific discovery pipeline, from synthesizing literature to generating hypotheses to informing experimental design. A unifying principle behind these tasks is the need to continuously update beliefs as evidence accumulates, for which Bayesian inference provides a natural framework. However, existing work assumes that LLMs cannot carry out principled Bayesian belief updating in context, instead relying on simplifications and heuristics. To challenge this assumption, we present a principled evaluation of Bayesian reasoning capabilities in LLMs, analyzing both the model's internal consistency as well as its ability to recover an externally determined posterior. Specifically, we build an evaluation framework spanning qualitative epistemological examples to probe probabilistic coherence, quantitative benchmarks of Bayesian inference capabilities, and analyses of the impact of non-Bayesian updates on downstream task performance. Across different scenarios, we find that LLMs generally violate Bayes' rule when updating their beliefs in context. However, when equipped with modern reasoning abilities and code execution capabilities, their performance drastically improves, enabling even the close approximation of analytically intractable posteriors. Our evaluation shows that current LLMs with access to appropriate tools could already be used to make complex belief updates in scientific discovery, enabling rapid acceleration of the efficiency and scale of AI-powered scientific discovery.

Primary area:
foundation or frontier models, including LLMs
Keywords:
LLMs, Bayesian inference, experimental design, scientific discovery

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_dKAxOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper tries to answer whether LLMs can perform approximate Bayesian network in context. It is shown that, in general, LLMs violate Bayes' rule when updating their beliefs in context. They propose evaluating Bayesian inference in context along three sequential axes. They compare against a recent baseline.

While the paper is interesting and easy to read, it was not made clear why these results are significant. The findings of this paper are not precise or agree with commonsense.

p7 Result 1. ... modern reasoning LLMs generally satisfy basic probability axioms ...

Agree with commonsense. Generally, but not guaranteed.

p8 Result 2. Without reasoning, LLMs fail to perform accurate Bayesian inference in context ...

Agrees with commonsense. LLMs do not perform accurate Bayesian network inference.

p8 Result 3. Reasoning substantially improves the ability of LLMs to perform Bayesian inference in context ...

Agrees with commonsense. Adding tools improves LLM ability.

p8 Result 3. ... closer to the optimium ...

Closer to the optimimum is not precise.

p9 Result 4. With access to external tools such as code environments, modern LLMs can select appropriate tools for Bayesian inference ...

Not precise. Can, not will. Can it be quantified? How often?

p9 Result 5. For modern reasoning LLMs, the structured harness employed by recent belief-updating frameworks such as BED-LLM may no longer be necessary ...

Not precise. Previous methods might be needed or they might not be needed. Can it be quantified when the previous method is needed?

In short, everyone knows that LLM answers cannot be trusted. When using LLMs, there is always a disclaimer shown that LLMs can make mistakes. In Bayesian inference, I can trust the answer from an exact Bayesian inference algorithm. I can trust that the answer from an approximate Bayesian inference algorithm lies between the provided lower and upper bounds. But the results put forth in this paper appear to come with no guarantees.

Critical strengths

Pursuing accurate Bayesian inference using LLMs is a worthy topic.

Critical weaknesses

The 5 Results reported in this paper are not precise or they agree with common sense.

Questions for authors

Why should the five stated results be considered significant?

3 comments

Thank you for your review! We appreciate that you found our paper interesting and clear and easy to follow, and that you have no substantive concerns about its soundness.

We are also glad that you consider pursuing accurate Bayesian inference with LLMs a worthy topic, and that from our work you learned something interesting and useful.

We address your concerns below.

Q6: Why should the five stated results be considered significant?

A6: Our contribution is an empirical evaluation framework, not a new inference algorithm with theoretical bounds. We believe such a framework is significant for three reasons:

  1. LLMs are already being used for belief updating in scientific discovery (Sec. 1), yet no theory characterizes when their belief updates can be trusted. Until such theory exists, systematic empirical evaluation is the only way to determine when LLM-based updates are reliable.
  2. Deep learning offers a clear precedent. Benchmarks guided its development and deployment long before theory could explain generalization, and the theory later caught up (for surveys on this topic see e.g. [2, 3]). We see our framework as playing an analogous role for LLM-based Bayesian reasoning.
  3. Our results directly change how LLM-based belief updating systems should be built. They show that the structural harnesses used by current frameworks provide no significant benefit for modern reasoning models (Res. 5).

The other reviewers likewise highlighted the significance of the framework, describing it as "genuinely integrative" (Reviewer 8SkZ), being useful to measure mutliple aspects of belief updating "without confounding factors" (Reviewer HBpS).

W5: Result 1 agrees with common sense.

R5: We respectfully note that this result runs counter to the prevailing view in the literature rather than confirming it, where empirical audits have repeatedly found probability judgments in LLMs overwhelmingly incoherent (e.g. [4, 5] - more in Sec. 2). Our new finding is that this incoherence largely disappears within a single dialogue. To our knowledge, this distinction has not been made before, and we are happy to be able to confirm common sense with our experimental results.

W6: Result 2 agrees with common sense.

R6: We would argue that this question is not settled in the literature. Previous work has shown that transformers are able to perform Bayesian inference [6], and that LLM's strong ICL capabilities can be framed as implicit Bayesian inference [7], with other sources showing systematic violations of Bayesian updates in different contexts. Our contribution is to provide an evaluation protocol that answers this for explicit, sequential belief updating, quantifies and qualifies the extent of the failure, and shows that it has measurable downstream consequences for experimental design, which prior work has not established.

W7: Results 3 and 4 agree with common sense: adding tools to LLMs improves their performance.

While it initially seems obvious that simply adding tools improves performance, even in our findings that is not always the case: On ProteinWorld-hard, reasoning leads to models reporting hallucinated Monte Carlo output. Our focus lies on an empricial evaluation, where we not only show which tools help in which regime, but also provide our qualitative analysis (App. B) to show how exactly these tools are applied, finding that LLMs use a fully human approach to perform Bayesian inference, rather than relying on some hidden internal shortcut.

W8: Results 3 and 4 are not precise enough.

Reading the results back, we agree that they can benefit from more precise phrasing, and have adjusted accordingly, see the changes highlighted in blue on page 9 of the revised paper.

W9: Result 5 is not precise. Can it be quantified when harnesses are needed?

R9: Thank you for pointing this out; we agree and updated the result accordingly, see again the changes in blue on page 9.

W10: Exact and approximate Bayesian inference algorithms come with guarantees, whereas the results in this paper appear to come with none.

R10: We fully agree that guarantees are essential, and we believe this is where our findings are most significant.

1. Given that there have not yet been found strong theoretical boundaries on LLM-based belief updates, we believe that our work is even more important to guide development and empirically verify current capabilities.

2. Our results, in particular the qualitative analysis in App. B, show that tool-augmented LLMs do not produce posteriors through opaque internal computation. One of our main findings is that they set up the probabilistic model explicitly, and either perform correct explicit analytical calculations or implement a standard Monte Carlo estimator. The resulting posterior is therefore the output of a classical inference algorithm and carries its usual guarantees. The question of trust therefore becomes significantly easier to verify by checking logs of the used tool, instead of having to treat the LLM's outputs as a blackbox.

In short: our findings show that LLM-based belief updates need not come without guarantees. LLMs are able to appropriately delegate computation to established inference algorithms, those algorithms' guarantees apply, and only the implemented procedure remains to be verified.


We hope these clarifications and the revised, quantified results address your concerns. Please let us know if any points remain unclear; we genuinely appreciate your time reviewing our work.

[2] Bartlett, Montanari and Rakhlin, “Deep Learning: A Statistical Viewpoint”, Acta Numerica, 2021.

[3] Fan, Ma and Zhong, "A selective overview of deep learning.", Statistical science: a review journal of the Institute of Mathematical Statistics 36.2, 2020.

[4] Freedman and Toni, "Exploring the potential for large language models to demonstrate rational probabilistic beliefs", In The International FLAIRS Conference, 2025.

[5] Li and Sreedhar, "Outcome-free arbitrage audits and coherence repairs for LLM forecasters", ICML 2026 Workshop on AI Forecasting, 2026.

[6] Mueller et al., "Transformers can do Bayesian inference", In International Conference on Learning Representations, 2022.

[7] Xie et al., "An explanation of in-context learning as implicit Bayesian inference", In International Conference on Learning Representations, 2022.

Dear Reviewer,

Thank you again for your review. We wanted to briefly follow up on our response, since the discussion period is coming to an end.

To summarize, we addressed your concerns as follows:

  • Significance of the results: Our contribution is an empirical evaluation framework. LLMs are already used for belief updating in scientific discovery, but no theory yet says when their updates can be trusted. Systematic empirical evaluation fills that gap, much as benchmarks guided deep learning before theory caught up.
  • Results vs. common sense: Several of our findings contradict or settle open questions in the literature. For example, prior audits found LLM probability judgments largely incoherent, whereas we show this incoherence mostly disappears within a single dialogue (Res. 1).
  • Precision of Results 3–5: We agreed with this point and rephrased these results more precisely in the revised paper (highlighted in blue, p. 9).
  • Guarantees: Our qualitative analysis (App. B) shows that tool-augmented LLMs explicitly set up the probabilistic model and either compute the posterior analytically or implement a standard Monte Carlo estimator. The resulting posteriors therefore inherit the guarantees of these classical algorithms. Only the implemented procedure needs to be verified, and that can be done by inspecting the tool logs.

We hope these clarifications and revisions address your concerns. If they do, we would be grateful if you would consider updating your score. If anything is still unclear, we are happy to discuss it further.

Best regards, The Authors

Reviewer_HBpSOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper investigates in detail the extent to which LLMs perform or approximate Bayesian inference in context. The authors answer the question using two environments: 20Q and ProteinWorld, where the former is a Q&A environment with two possible outcomes per experiment, infinite experiment and hypothesis space, and the latter is a more customizable environment, typically with a finite experiment and hypothesis space. The evaluation spans three different axes: probabilistic coherence, posterior fidelity and downstream task performance. Across the experiments with multiple LLMs, the authors conclude that within a single dialogue, LLMs elicit coherent distributions, access to reasoning substantially improves performance on all 3 axes and further access to advanced tools like code execution leads to improvement in setups with analytically intractable priors as well.

Critical strengths
  • Well-motivated and articulated problem statement:
  1. Clear definition of the axes of comparison: probabilistic coherence, posterior fidelity and downstream task performance.
  • Experimental questions are well-framed
  1. Questions progressively measure one thing and report the conclusions without confounding factors.
  2. Several LLMs have been tested for most experiments.
Critical weaknesses
  • The experimental setup is unclear and key details are missing from the main text
  1. How are the probabilities read from the LLM?
  2. For the ProteinWorld, how is the latent vector sampled for every run/experiment? Is it fixed?
  3. How are experiments selected?
  • Limited Evaluation for Code Execution Tool
  1. Results only with Gemini 2.5 Pro model
Questions for authors
  • Clarification Questions:
  1. App F2.2 reads that for reasoning models, probabilities are verbalized and read, whereas for non-reasoning models, token probabilities are used. Why does this differ across model types? Does this confound the results in Fig 5?
  2. It seems that the hypothesis space for 20Q environment can change throughout a single run as the LLM filters the set for animals that are compatible with all the experiments so far (App F2.1). This is contrary to traditional bayesian setup where a candidate might be pruned with negligible probability but no new candidate is added along the way. Is this the case with the 20Q environment? Why not sample a subset externally without an LLM and use that as the hypothesis set for the run?
  3. Figure 6 shows that the downstream task performance with reasoning and code surpasses the oracle at some experiment numbers. Is there an explanation behind this observation? Shouldn’t matching the oracle be the performance ceiling?
  • Suggested experiments:
  1. The experiments in the paper are over short horizons of upto 20 steps at max. However, the use in scientific discovery would require working with a few hundreds of experiments and belief update steps. Could you provide an experiment with a larger horizon and conclude if the same observations hold?
  2. Could you provide the results on the ProteinWorld-hard with at least 1 more LLM?
5 comments

Thank you for your review and for the detailed questions and suggestions! We appreciate that you found our problem statement well-motivated and clearly articulated, with a clear definition of the three evaluation axes: probabilistic coherence, posterior fidelity, and downstream task performance.

We are also grateful that you highlighted our well-framed experimental questions, which:

  • progressively measure one aspect at a time and report conclusions without confounding factors; and
  • are evaluated across several LLMs for most experiments.

Finally, we are glad that you found the paper interesting and useful, and that your soundness concerns are minor and fixable. We have addressed them below and believe the resulting changes substantially strengthen the paper.

We aim to address your concerns below:

W2: Key details about the experimental setup are missing from the main text.

R2: Given the technical nature of the setup, particularly for ProteinWorld, we chose to focus the main text on the conceptual design and to defer implementation details to the appendix. All three points you raise are discussed in detail there: see App. F2.2 and F2.4 for 20Q, and App. F3.1, F3.2, and F3.4 for ProteinWorld.

That said, after rereading the main text, we agree with you that brief explanations of these key details significantly improve readability. We have therefore added concise descriptions of each point to the main text (see the revised paper, highlighted in blue, pp. 4–6).

W3: The tool execution evaluation was done using only a single model.

R3: Please see R1 (under the review from 8SkZ), who raised the same point. In short: We repeated the experiments on two frontier models and observed the same outcome.

Q3: Do the different probability-extraction methods confound the results in Fig. 5?

A3: This is an important point, and one we considered carefully. The short answer is no. If anything, using the same extraction method for both model types would make the contrast even stronger.

In our setup, token log-probabilities are highly concentrated on one answer: typical outcomes in 20Q are either p(Yes)≈1, p(No)≈0p(\text{Yes}) \approx 1,\ p(\text{No}) \approx 0 or vice versa. This holds for both the individual experiment likelihoods and the aggregate belief elicitation. As a result, belief updating effectively reduces to a binary decision, which makes Bayesian inference considerably easier. Verbalized probabilities, by contrast, regularly retain small residual uncertainties (e.g., p(Yes)≈0.95, p(No)≈0.05p(\text{Yes}) \approx 0.95,\ p(\text{No}) \approx 0.05), which makes correct aggregation substantially harder.

In other words, reasoning models face the harder elicitation setting and still outperform non-reasoning models, which makes the finding sharper, not weaker. We used token log-probabilities for non-reasoning models purely for computational efficiency: 20Q requires one LLM call per belief and experiment at every iteration, so even small savings in output tokens add up significantly.

Finally, note that in ProteinWorld, all probabilities are elicited verbally for all models, so this potential confounder does not arise there.

Q4: Does the changing hypothesis space in 20Q violate Bayesian updating?

A4: As you point out, the natural approach to exact Bayesian updating would be to fix a set of nn animals externally and run the entire evaluation on that set. However, combined with the unconstrained experiment space Ξ\Xi of 20Q, this makes the task trivial: the LLM reaches perfect accuracy after ⌈log⁡2n⌉\lceil\log_2 n\rceil experiments simply by asking whether the animal lies in the first half of the remaining hypotheses and discarding the appropriate half based on the answer. This would render the downstream task performance analysis uninformative.

The informative alternative is hypothesis resampling, which is a practical approximation over open-ended hypothesis spaces established by prior work [1]; we follow this approach for consistency with the literature. For more details, see App. F2.1.

Q5: Why can the LLM outperform the oracle in the tool-use setting of ProteinWorld?

A5: In ProteinWorld-Hard, the posterior no longer has a closed-form solution, so the best achievable results rely on numerical integration. Both the oracle and, as shown in App. B.4, the LLM obtain their solutions via MC sampling. Both are therefore stochastic estimates of the same posterior

While the oracle is the ceiling in expectation, individual runs can fall on either side of it due to Monte Carlo variance. The observed difference between them lies well within the standard deviation of this sampling process, and we note that this establishes no statistical significance in the difference.

E1: Provide an experiment with a longer horizon, to better reflect realistic scientific discovery processes.

O1: Thank you for this suggestion. We found it very compelling and have implemented it. Specifically, we extended ProteinWorld so that each experiment ξ\xi can be repeated independently, with outcomes drawn from p(y∣θ⋆,ξ)p(y \mid \theta^\star, \xi). By increasing the observation noise, each individual outcome becomes less informative, which allows us to extend the horizon arbitrarily while keeping the inference problem non-trivial.

The full setup and results are provided in the newly added App. H. In short, we have confirmed that our findings hold for horizons of as long as 200 experiments.

E2: Use more than one model for the findings in ProteinWorld - hard.

O2: Same as in W3, we kindly ask to refer to R1 (under the review from 8SkZ). In short: We repeated the experiments on two frontier models and observed the same outcome.

[1] Choudhury et al., "BED-LLM: Intelligent information gathering with LLMs and Bayesian experimental design", In International Conference on Learning Representations, 2026.

Dear Reviewer,

Thank you again for your detailed review and suggestions. We wanted to briefly follow up on our response, since the discussion period is coming to an end.

To summarize, we addressed your concerns as follows:

  • Experimental setup details: We added short descriptions of how probabilities are elicited, how ProteinWorld latents are sampled, and how experiments are selected to the main text (highlighted in blue, pp. 4–6).
  • Code-execution evaluation: We extended the tool-use analysis to two more frontier models (Sonnet 5.5 and Opus 5.5; updated Fig. 6). The results are consistent with the original Gemini-2.5-Pro findings.
  • Longer horizons: Following your suggestion, we extended ProteinWorld to support repeated noisy experiments (App. H). Our findings hold for horizons of up to 200 experiments.
  • Clarification questions: We explained why the different probability-extraction methods do not confound Fig. 5 (if anything, they make the contrast sharper). We also motivated the hypothesis resampling in 20Q, and showed that the apparent gains over the oracle in ProteinWorld-Hard are within Monte Carlo variance.

We hope these additions address your concerns. If they do, we would be grateful if you would consider updating your score. If anything is still unclear, we are happy to discuss it further.

Best regards, The Authors

Reviewer_8SkZOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper investigates whether Large Language Models (LLMs) can perform in-context Bayesian belief updating—a foundational requirement for autonomous "AI Scientists" navigating sequential experimental workflows. The authors evaluate LLMs across three progressive axes: probabilistic coherence, posterior fidelity, and downstream task performance in adaptive experimental design. To validate their framework across 20 Questions and a novel synthetic environment called ProteinWorld, they evaluate diverse model families across non-reasoning, reasoning, and tool-augmented modes. Their findings demonstrate that while standard non-reasoning LLMs systematically violate Bayes' rule, native test-time reasoning and Python code execution enable models to compute exact posteriors or autonomously approximate analytically intractable posteriors using Monte Carlo estimation, ultimately rendering strict external belief-updating harnesses redundant for modern reasoning LLMs.

Critical strengths

A Genuinely Integrative Evaluation Framework: Most prior work evaluates probabilistic coherence, posterior fidelity to an oracle, or downstream utility in isolation. This paper ties all three together on the same models and task instances, enabling a clear, causal narrative—demonstrating that test-time reasoning and tools resolve coherence and fidelity failures, which directly translates into downstream experimental design accuracy.

Transparent, Mechanistic Failure-Mode Analysis: The paper provides verbatim reasoning traces that differentiate specific, actionable failure modes. Specifically, it cleanly separates "hallucinated simulation" (where models describe a Monte Carlo procedure they never executed and report fabricated output) from "numerical over-simplification" (where crude deterministic approximations incorrectly assign zero mass to valid hypotheses).

Critical weaknesses

Headline Tool-Use Result Rests on a Single Model and Instance: The paper's most striking claim—that code execution resolves the intractable-posterior gap—is demonstrated using only one model (Gemini-2.5-Pro), one tool environment, and a single worked instance in ProteinWorld Hard. In contrast to the broader multi-model evaluations used elsewhere in the paper, no other tool-augmented reasoning models are tested in this regime. This asymmetry leaves the result as a promising existence proof rather than a general capability, making the paper’s sweeping claim that "current LLMs with access to appropriate tools" can solve intractable belief updates over-generalized relative to the supporting evidence.

Questions for authors

Scaling Hypothesis Spaces: How do the reasoning and code-execution capabilities hold up when the hypothesis space Θ\Theta cannot be exhaustively enumerated in the context prompt (e.g., infinite continuous hypothesis spaces or non-parametric models)? Does the model still write functional sampling scripts (e.g., MCMC or SMC) when forced to define hypotheses programmatically?

Computational Budget & Verification Loops: The qualitative analysis notes that reasoning models engage in heavy, redundant re-verification for thousands of thinking tokens even after deriving the exact posterior. Have you explored early-stopping mechanisms or prompt interventions to mitigate this inference-cost overhead during extended experimental design loops?

2 comments

Thank you for your review! We appreciate that you found our paper clear and easy to follow, that it gave you a substantially deeper understanding of the problem, and that you have no substantive concerns about its soundness.

We are also grateful that you highlighted:

  • how our framework evaluates coherence, fidelity, and downstream experimental design together on the same models and tasks, rather than in isolation; and
  • our qualitative analysis of reasoning traces, which separates distinct failure modes, such as models reporting simulations they never ran or oversimplifying numerically to the point of ruling out valid hypotheses.

We aim to address your concerns below:

W1: The tool-use result rests on a single model and instance.

R1: We would first like to clarify that the tool-use evaluation is not based on a single instance: it spans 5 separate ProteinWorld instances with fully independent parameters. Given the structure of ProteinWorld, these can be viewed as 5 distinct environments. Additionally, we now extended our analysis to a continuous setting of ProteinWorld, see App. G, supporting our claims with yet another even more different environment.

Regarding the choice of model, we deliberately restricted the original analysis to Gemini-2.5-Pro because it is comparatively weak relative to current frontier models. We therefore expected the demonstrated behavior to carry over to stronger models. Nonetheless, for the sake of completeness, we decided to verify this explicitly. We have therefore extended the analysis to two additional frontier models, Sonnet 5.5 and Opus 5.5 (see the updated Fig. 6). The new results consistently reinforce our claim, which now rests on multiple models across multiple independent environments rather than a single model on a single instance.

Q1: How do the ProteinWorld findings transfer to non-enumerable hypothesis spaces?

A1: We agree that this is a natural next step after introducing ProteinWorld. To address it, we have implemented a continuous version of ProteinWorld, in which the goal is to estimate a latent continuous parameter. As in the discrete case, we consider two variants: one with an analytically closed-form posterior, and one that requires numerical integration. The full setup and results are provided in App. G.

In short, our findings transfer directly to the continuous setting:

  • Without reasoning, LLMs fail to recover even simple continuous posteriors.
  • With reasoning, they accurately compute closed-form posteriors.
  • With code execution, LLM agents accurately estimate complex posteriors through numerical simulation.

Q2: Have you explored early-stopping mechanisms to address redundant re-verification in LLM reasoning?

A2: We agree that this is an important point, and we think it involves two separate questions: Can LLM reasoning be stopped early, and what are the consequences of doing so?

Can reasoning be stopped early? We explored two approaches. First, as you suggested, we tried prompt interventions, for example by adding instructions such as "Once you have arrived at a final solution, return it immediately without further verification" to the system prompt. Across a variety of prompts, this did not yield consistent improvements across models. Second, we applied a hard-coded intervention: inserting the end-of-thinking token once a fixed token budget was reached. This reliably stops further reasoning. However, it is not a practical solution, since the number of reasoning tokens needed for a given inference problem is not known in advance.

What are the consequences? We observed a critical cutoff point. If the model has already begun re-verifying, truncating its reasoning has no measurable impact on performance: the model simply outputs the correct solution it has already computed. If reasoning is stopped before the first correct calculation is complete, however, performance drops sharply, to a level barely above that of non-reasoning models. This mirrors the strikingly human-like way in which models approach Bayesian inference: without working through the calculation "by hand," there is no shortcut to oracle-quality answers.

Dear Reviewer,

Thank you again for your thoughtful review. We wanted to briefly follow up on our response, since the discussion period is coming to an end.

To summarize, we addressed your main concern by:

  • Extending the tool-use analysis to two more frontier models (Sonnet 5.5 and Opus 5.5; updated Fig. 6). The results are consistent with the original Gemini-2.5-Pro findings.
  • Clarifying that the evaluation already covered 5 independent ProteinWorld instances, not one.
  • Adding a continuous version of ProteinWorld (App. G). It answers your question about non-enumerable hypothesis spaces and shows the same pattern across non-reasoning, reasoning and tool-augmented models.

We also reported our early-stopping experiments, which use both prompt interventions and hard token-budget truncation.

We hope these additions address your concerns. If they do, we would be grateful if you would consider updating your score. If anything is still unclear, we are happy to discuss it further.

Best regards, The Authors

Comments on the paper

1 comment

We thank the reviewers for raising concerns, providing feedback, and suggesting new experiments to improve our work. We appreciate the acknowledgement that:

  • The paper is clear and easy to follow (8SkZ, dKAx);
  • The methodology is sound, with no substantive concerns (8SkZ, dKAx);
  • The problem is well-motivated and worth pursuing (HBpS, dKAx);
  • The evaluation framework is well-designed, jointly assessing probabilistic coherence, posterior fidelity, and downstream task performance (8SkZ, HBpS);
  • The work provides interesting and useful insights (all reviewers).

Following the reviewers' suggestions, we have added new experiments and revisions to the paper (highlighted in blue). In detail:

  • Two additional frontier models (Sonnet 5.5 and Opus 5.5) for the tool-use evaluation on ProteinWorld-Hard (updated Fig. 6; W1, 8SkZ; W3 and E2, HBpS);
  • A continuous version of ProteinWorld, extending our findings to non-enumerable hypothesis spaces (App. G; Q1, 8SkZ);
  • A long-horizon version of ProteinWorld, confirming our findings over up to 200 experiments (App. H; E1, HBpS);
  • More precise formulations of our main results (Results 3–5, p. 9; W8 and W9, dKAx).

We address the individual questions and concerns in a comment following each reviewer's feedback.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.