acceptodds
Withdrawn from ICLR 2027

An Information-Theoretic Approach to Agentic Scientific Exploration

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Answering open scientific questions requires gathering evidence scattered across literature, databases, and search engines whose contents are too vast to model in advance. Retrieval-augmented generation selects actions by semantic similarity, treating each step independently and ignoring what the agent has already learned or still needs to resolve. We propose an information-theoretic framework that computes expected information gain over a set of competing hypotheses rather than over the unbounded space of possible observations, enabling principled action selection in open systems where observation-space information gain is intractable. The agent maintains an evolving memory graph, generates hypotheses, and selects actions by a multi-objective utility combining hypothesis-space information gain, novelty, coherence, and challenger-driven falsification pressure. Beliefs are updated by Bayesian scoring over accumulated evidence. On cancer-type prediction from gene coexpression networks, our agent is the only method to achieve balanced sensitivity and specificity and positive Matthews correlation, whereas strong agentic baselines (ReAct, Tree-of-Thoughts, MCTSr, and AlphaApollo) each collapse to a class bias, defaulting to affirmation or to rejection and achieving no better than chance-level discrimination. Ablations show that each utility term contributes to this balanced discrimination.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
LLM agents, agentic scientific exploration, expected information gain, hypothesis-space reasoning, Bayesian belief updating, biomedical discovery

Reviews

Among:
Shape:
1234this paper, Bottom 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.00 over 3 reviews, above 0% of the others.
Reviewer_Q83qOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

The authors proposed an information theoretic framework for guiding LLM agents in scientific exploration. While traditional RAG methods have Markovian memory their proposed graph memory mechanism permits a sequential accumulation of evidence. To evaluate their approach the authors construct a custom TCGA benchmark for binary classification and compare their methods against several other popular agentic frameworks as well ablations of their own method.

Critical strengths

The problem of studying scientific exploration with information theoretic approaches for agents is timely and well-motivated.

Critical weaknesses

The chosen benchmark seems very atypical for evaluating information retrieval and scientific exploration.

The method seems like a straightforward application of standard information theoretic approaches.

The reporting of the results is very confusing and I am not convinced the method achieves SOTA, as well as a poor overall performance of all methods on the benchmark.

The authors only evaluate their approach using a single LLM backbone on a single benchmark.

Questions for authors

Could the authors please run additional benchmarks for the approach including more common benchmarks for information retrieval/scientific exploration such as SciGym explain why their setting is more suitable.

Why were 15 tumour types excluded and why is the number of samples capped at 200?

What were the criteria for manual selection of the negatives per type?

Can the authors please report AUROC and AUPRC.

I am very confused about the authors intention when showing Table 2 as under the shown numbers AlphaApollo is the best method but the text argues extensively why the metric is flawed so why include it at all?

Can the authors please run a more sensitivity analysis of the hyperparameters in the ablations as opposed to deactivated a certain component fully. Since the results if any of the terms are removed are random there is a concern that the current system is just overfitted on the benchmark. Furthermore, is the development set excluded from the metrics being reported?

Can the authors please test a general classifier e.g. XGBoost for the task with cross validation to establish how much predictive information is available in the dataset as a potential upper bound.

Section 2.4 the authors state that traditional EIG requires a generative model but then proceed to use a LLM for that task. Could you please elaborate the novelty in this section.

All the experiments use bulk RNA settings, which inevitably mixes tumout types and surrounding tissues and as such is ill suited for the task at hand. Can the authors please run experiments with scRNA sequencing methods and discuss the limitation of their current experiments more candidly.

Minor Comments: Figure 1 is very small please make it full width

Reviewer_54sxOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper addresses a challenge within the context of open scientific inquiry: how to select better actions to resolve a given problem. The authors propose an approach to calculate the expected information gain for competing hypotheses. The core of this method is a memory graph, with beliefs about hypotheses updated via evidence-weighted scoring with prior anchoring. The authors evaluate their method on a manually built task over 18 cancer types.

Critical strengths
  • Well-motivated paper, the information-theoretical formulation is sound.
  • Comparing with one strong baseline AlphaApollo and statistical significance testing.
Critical weaknesses
  • The authors claim that no existing benchmarks are available, a assertion that is highly questionable. To my knowledge, benchmarks such as ActiveSciBench, OpenHypBench, and Autonomous Science Bench may overlap with the tasks addressed in this paper. If these existing benchmarks are deemed unsuitable, the authors should provide a well-reasoned justification. Alternatively, if no appropriate benchmarks truly exist, the authors may need to, at least expand the data collection scale (e.g., within the TCGA dataset) to construct a compelling benchmark. Given the current form of the paper, that includes significant noise and cases where certain metrics fall below those of the baselines (Table 1), the evidence does not support the validity of the proposed method.
Questions for authors

My primary concern lies in the experimental evaluation, please see Weaknesses.

Reviewer_2zP9OpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes an agent that maintains competing hypotheses and selects tool actions using hypothesis-partition entropy, combined with challenger, novelty, coherence, and exploration terms. Evaluation on 72 binary queries derived from 18 cancer coexpression networks shows more balanced sensitivity and specificity than the tested baselines.

Critical strengths

The focus on discriminative evidence rather than merely relevant retrieval is useful. The challenger mechanism addresses premature commitment, and the component ablations provide some evidence that the combined design helps.

Critical weaknesses

The methodological contribution is limited. The approach combines established information-gain ideas with several manually weighted heuristics. Partition entropy has an information-theoretic interpretation under an idealized observation model, but the implemented belief update is a composite score rather than Bayesian conditioning. The paper does not establish that the proxy reliably predicts useful actions. A particularly important missing comparison is a simpler agent explicitly instructed to investigate competing hypotheses and seek counter-evidence during exploration.

The evidence is too limited for the claims. Evaluation uses one domain, one backbone, and only 18 underlying networks. Balanced accuracy improves from 54.7% for ReAct to 58.0%, while runtime increases by 3.2×; the reported McNemar tests do not establish significance of the MCC improvement. There are also direct inconsistencies: the text claims all baselines have nonpositive MCC, whereas Table 1 reports positive means for ReAct and ToT. Likewise, removing coherence leaves positive mean MCC in Table 3, contrary to the ablation discussion. These overstatements weaken confidence in the conclusions.

Questions for authors

Do you have a comparison with exploration-time counter-evidence prompting under matched inference cost and the same final evidence audit?

Can you reconcile the MCC claims with Tables 1 and 3, and provide uncertainty estimates for the MCC difference that account for queries sharing the same network?

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

Related papers

Discussion (0)

Sign in to comment.