acceptodds
Under review as a conference paper at ICLR 2027

Investigation Capability versus Investigation Intelligence: Diagnosing Selective Evidence-Gathering in LLM Agents with Information-State Oracles

Abstract

LLM-controlled agents can gather costly diagnostic evidence before updating a belief; this capability alone is consequential, raising final-task success from approximately 0% to 90-100% in a controlled sequential-decision benchmark. However, capability does not imply investigation intelligence: an agent may have access to evidence without knowing when acquiring it is worth its cost. The original LLM-driven investigation trigger shows no significant advantage over simple heuristic policies in 10 of 12 paired comparisons; the two nominally significant differences do not survive Bonferroni correction. We introduce a validated agent-information oracle evaluating policies on only agent-available information, alongside a clairvoyant oracle giving an achievable ceiling. Using these oracles, we decompose policy regret into missed and unnecessary investigation and downstream action errors, revealing a systematic mismatch between investigation behavior and the oracle's value of information in a noisy, budget-scarce regime. Factorial experiments identify accumulated dismissed-failure count, rather than recency, as the primary behavioral factor associated with under-investigation in this setting. A targeted intervention further demonstrates that this behavior is manipulable, improving success from 91.7% to 94.0% in one regime (p = 0.020, n = 25), but not uniformly across budgets and noise levels. Because an explicit uncertainty scalar alone did not correct the remaining gap, we construct matched counterfactual histories from legitimate investigation-mediated evidence and localize, in the Hard, budget-2 regime, a narrow, repeatedly-observed belief-conditioned mismatch: behavior is consistent with an effective update threshold that is offset above the oracle's decision boundary. We further extend this diagnostic apparatus to an environment in which evidence must be attributed to the correct one of several tools, and find that simple heuristic baselines exhibit cross-tool evidence contamination that is absent in the tested LLM agents under the conditions evaluated - a result that is robust across models, a hard lexical confound, a vocabulary-domain transfer, and randomized phrasing, subject to an explicit, model-dependent boundary condition under simultaneous multi-source distractor pressure. These results suggest the central difficulty is not access to evidence or representation of uncertainty, but the selective, budget-aware conversion of uncertainty into appropriate investigation and update decisions.

Primary area:
foundation or frontier models, including LLMs
Keywords:
large language model agents, decision-theoretic evaluation, value of information, regret analysis, selective evidence gathering, sequential decision making, POMDP, belief updating

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_KTzFOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper purports to study information-seeking and evidence-gathering in LLMs. The writing is borderline unreadable and results are unclear. I find this work not worth my time as a reviewer to parse. The brief AI disclosure arguably fails to meet ICLR’s requirements. I move for desk rejection.

Critical strengths
  • The general question-asking / information-seeking setting is an interesting research problem.
Critical weaknesses
  • The writing strongly bears the main hallmarks of AI slop: prose is overly dense and unreadable. The abstract alone is almost an entire page and fails to distinguish between framing and irrelevant implementation details.
  • Related work coverage is wildly insufficient. The paper only includes 4 references.
  • Figures are low-quality Tikz or bar graph plots that appear AI generated (e.g., Fig. 1). The main results figures lack confidence intervals. In most cases there is no discernible difference between conditions.
Questions for authors
  1. What problem are you trying to solve with this paper?
  2. Would it be possible for you to rewrite this paper from scratch without the use of AI writing tools?
Reviewer_tt6XOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper studies whether LLM agents know when gathering diagnostic evidence is worth its cost. It introduces a synthetic API task with noisy diagnostics and limited budgets, together with two oracles for evaluating decisions and separating different investigation errors. Experiments examine history effects, a prompt intervention, belief-conditioned decisions, and evidence attribution across tools, with selected findings tested on three models. The paper reports gaps in investigation selectivity and better evidence scoping than the tested tool-blind heuristics, although the oracle and experimental issues weaken these conclusions.

Critical strengths
  1. The paper asks a useful question: can an agent gather evidence, and does it know when gathering evidence is worthwhile? The proposed oracle-based analysis separates investigation errors from final accuracy, which could help diagnose agent behavior more clearly (Sections 4–5).

  2. The authors provide substantial code and logs and report negative results and model-dependent effects. This makes the findings easier to check and helps show the limits of the proposed explanations (Sections 6–8 and the supplement).

Critical weaknesses
  1. The claimed exact oracle rounds beliefs to a 0.025 grid, which changes the central threshold result (Sections 4 and 7). Removing this rounding changes three of the five reference actions near 0.60 from UPDATE to INVESTIGATE and raises agreement for the unexposed condition from 20% to 80% in both main datasets. This substantially weakens the reported belief-conditioned mismatch.

  2. The generated histories always display the city parameter, even after the internal belief switches to location_id, so the posterior calculation and the model do not receive consistent historical evidence. Datasets 1 and 2 also reuse all five histories at the critical 0.60 region, weakening the claim of replication across independently constructed history sets.

Questions for authors
  1. Oracle validation: Does the threshold-offset finding survive removing belief rounding or using a demonstrably converged calculation?

  2. History construction: Can you clarify why historical calls always show city after the internal belief switches to location_id? Do the results hold with consistent serialization and separate history sets evaluated at the same step and budget?

  3. Evaluation objective: Is the oracle intended to maximize decision-label accuracy or actual task success? How do you justify labeling every permanent-change event UPDATE when repeated changes can leave the current belief correct?

Reviewer_W7SMOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
3
Read review
Paper summary

The paper studies whether LLMs can make the right information gathering steps. The paper constructs various toy experiments across a variety of settings and reports findings. They also propose a steering approach that could improve performance.

Critical strengths

The question studied seemed interesting and the experiments cover various domains

Critical weaknesses

The paper was very difficult to parse because of being almost entirely AI generated - i gave my best effort to understand the details but failed.

The experiments were on a toy setting which makes it more critical that the construction, inspiration behind the construction, takeaways are easy to read. Unfortunately, i couldn't get anything out of this paper.

Questions for authors

None

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.