acceptodds
Desk rejected at ICLR 2027

Robust Information-Gain Control Active Agentic Reasoning under Approximate Beliefs

Desk rejected at ICLR 2027, so left out of the paper lists and the map.

Abstract

Large language model (LLM) agents in multi-turn reasoning often fail to make epistemic progress under approximate belief tracking. We formalize this failure through Belief Trap Regions (BTRs), forward-invariant regions of belief space in which expected progress stagnates. We show that even information-gain maximization (IG-max) can induce such traps under belief approximation error: small perturbations in the agent’s internal belief may invert action rankings, causing repeated selection of uninformative actions despite the existence of informative alternatives. To mitigate this failure mode, we propose Distributionally Robust Information Gain (DR-IG), a decision-time objective that optimizes worst-case information gain over an ambiguity set around the current belief estimate. We show theoretically that robust worst-case control restores the negative drift of a progress potential under bounded belief error whenever uniformly informative actions exist. We further introduce a lightweight verbalized belief proxy that extracts approximate belief states from confidence-annotated model outputs, enabling belief-aware planning without additional training. Across interactive reasoning benchmarks and multiple LLM families, naive IG-based agents frequently collapse into deterministic self-locking loops with near-zero solve rates. In contrast, DR-IG substantially reduces trap incidence and partially restores epistemic progress under approximate beliefs. Together, our results identify a concrete mechanism underlying active-reasoning failure in LLM agents and demonstrate that robust belief-aware control can mitigate these pathologies in a training-free setting.

Primary area:
foundation or frontier models, including LLMs
Keywords:
agentic reasoning, LLM belief

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_y2XDOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper analyzes, from the perspective of approximate beliefs, why LLM agents in multi-turn active reasoning repeatedly issue uninformative queries and stop making progress. The authors formalize this as a Belief Trap Region, or BTR, a forward-invariant region of belief space in which the progress potential no longer decreases, and show that information-gain maximization, or IG-max, can also fall into such traps when the belief approximation error is large enough. To mitigate this, they propose DR-IG, which maximizes the worst-case information gain over an L1 ambiguity set around the current belief, together with a verbalized belief proxy that builds a top-M hypothesis distribution from the LLM's verbalized confidences. On CircuitDecoding and ConceptInduction, IG-max fails in most episodes by repeating the same actions, whereas DR-IG suppresses the repetition and recovers part of the solve rate. The controller is also run on a real browser task, MiniWoB++ guess-number.

Critical strengths

S1. The paper addresses a frequently observed failure. LLM agents repeating the same queries or getting stuck in uninformative actions is a representative failure of multi-turn agents. Defining BTRs through forward invariance and drift provides a clear vocabulary for discussing this phenomenon.

S2. The analysis of the failure mechanism is concrete. The analysis showing that, in trapped episodes, the belief freezes and the shortlist is regenerated identically at every turn illustrates the failure process well. The observation that a single-step ranking inversion is nearly harmless and that the damage arises from accumulation over multiple turns is also interesting.

S3. The ablation that separates escaping repetition from solving the task is useful. The result that a random top-K or an equal mix of the worst case and the mean breaks the repetition but does not solve the task shows well that action diversity alone is insufficient and that an informative direction is needed.

Critical weaknesses

W1. The theory does not support the proposed method, and the difference from prior work is unclear. The abstract and introduction claim to show theoretically that DR-IG restores progress under bounded belief error, but the paper contains no theorem about DR-IG. For Theorem 2.3 on traps, it is unclear how an actual ranking inversion is guaranteed when L is only an upper bound on the change in information gain, and where the conclusion that the same action keeps being selected comes from when the approximate belief changes over time. Moreover, since the BTR concept comes from prior work, the new contribution of this paper lies in DR-IG, yet its difference from existing robust approaches that optimize for the worst case under uncertain beliefs is not discussed.

W2. The observed traps appear to be an artifact of the evaluation pipeline, and the benefit of DR-IG is confined to a narrow setting. That all five LLMs fall into the same repetition pattern at the same step suggests that the loop is produced by the controller rather than by the LLMs, and the authors themselves call it a "framework-induced loop." Moreover, in both tasks an exact belief can be obtained simply by eliminating hypotheses that are inconsistent with the observations, and the Oracle indeed solves every episode, yet the belief is replaced with a handful of hypotheses verbalized by the LLM, which creates the problem artificially. Without a comparison to simple fixes such as excluding already-asked queries, it is impossible to judge whether DR-IG is necessary. The benefit of DR-IG also appears only at a specific ε with α = 0, and on the real browser task DR-IG does not outperform a simple LLM agent, 80.0% versus 85.0%.

W3. The computation of the verbalized proxy and DR-IG is not described in enough detail to be reproduced. For the LLM to name its top-M hypotheses among 500 circuits, the form in which the candidates are presented is crucial, but this is not explained. The predictive updater used to compute agent-side information gain, how particles are sampled within the ambiguity set, how ε was chosen, and the prompt that presents the shortlist to the LLM are also not described. Furthermore, α = 1 in the ablation is defined as the mean over particles, whereas in the original definition of DR-IG it corresponds to the nominal information gain.

Questions for authors

Q1. What are the theorem and assumptions behind the abstract's claim that DR-IG restores negative drift under bounded belief error? In Theorem 2.3, L is an upper bound on the change in information gain; how is an actual ranking inversion guaranteed, and under what assumptions does the same action keep being selected as the approximate belief changes over time? How does DR-IG differ from existing robust approaches such as robust Bayesian experimental design?

Q2. Does the advantage of DR-IG persist when compared with simple fixes such as excluding already-asked queries, smoothing the proxy, or removing hypotheses inconsistent with the observations?

Q3. How are the 500 hypotheses presented to the LLM? What are the predictive updater, the particle sampling scheme, and the procedure for selecting ε?

2 comments
AuthorsOpenReview ↗

Response to Reviewer y2XD (1/2)

Thank you for asking us to distinguish failures of the elicited belief from failures introduced by the controller. That distinction was decisive. Our audit found that the original CircuitDecoding, ConceptInduction, and LP-500 runs never successfully elicited a belief: the parser expected four-digit answers from another task. On CircuitDecoding, all 25,596 belief steps fell back to the uniform prior, and a replay without an LLM reproduces the loop signature (Appendix E). We withdraw the affected results and the original attribution of the identical cross-model loops to elicited beliefs.

W1 / Q1 — Theory and relationship to robust Bayesian experimental design

The submitted paper contained no theorem supporting the abstract’s negative-drift claim for DR-IG, and Theorem 2.3 was incorrect. As you point out, a Lipschitz upper bound cannot guarantee a ranking inversion. A one-step inversion also does not establish that the same action will be selected as the proxy changes. The revision removes the abstract claim and replaces the original theorem with separate statements. Proposition 2.2 gives δ≥η/(2L)\delta\geq\eta/(2L) as a necessary condition for a zero-progress IG-max choice. Proposition 2.3 requires the proxy to continue ranking the uninformative action first along the trajectory. Proposition 2.4 gives a guarantee for the exact robust objective when the ambiguity set contains the true belief and an available action has uniformly positive information gain within the set.

We now position DR-IG explicitly within robust Bayesian experimental design and credit prior work on worst-case expected information gain. Proposition 2.5 shows that, with deterministic binary feedback, exact DR-IG gives the same action ranking as nominal IG. The new results concern what can be checked when the belief is elicited. The mass assigned to hypotheses already refuted by feedback gives a lower bound on the belief error, which can be used to rule out coverage by the ambiguity set (Proposition 2.6). In the revised GuessNumber runs, this check rules out coverage in 84–100% of steps. Under the feedback assumptions of Proposition 2.7, directly testing a verified candidate guarantees positive progress without a bound on the elicited belief error.

W2 / Q2 — What survives the pipeline correction, and how simple fixes compare

With verified elicitation, parse failures account for at most 2.4% of steps in the main tables. Some degradation remains: within ten queries, IG-max underperforms a belief-free random shortlist on three of five GuessNumber models (78% vs. 95%, 52% vs. 91%, and 60% vs. 92%). On CircuitDecoding, one model reaches 7% under IG-max versus 33% with a random shortlist. The revised evidence supports a model-dependent degradation, with the original universal looping claim withdrawn.

We added the simple alternatives you asked for: repeat masking, smoothing, and consistency filtering of the named hypotheses. For the CircuitDecoding model that stalls, DR-IG improves identification from 7% to 36%, with the improvement over IG-max replicated on fresh secrets. Its performance is statistically indistinguishable from masking in the pooled comparison. With tuning on a separate development split, DR-IG reaches 51%, versus 56% for both smoothing and masking (Appendix D.3). We also retain the MiniWoB++ result in Appendix F with the direct comparison you highlighted: 80.0% for DR-IG versus 85.0% for the free-form agent.

On GuessNumber, filtering the named hypotheses and then scoring actions by IG had little effect (62.3% versus 59.7% within six queries). Directly testing candidates that passed the same consistency check reached 85.7% on the design secrets. On 300 fresh secrets, the prespecified comparison with free-form reasoning yielded 79.0% versus 78.3% (47 versus 45 discordant pairs, p=0.92p=0.92). This replication covers the free-form comparison, and the verified-candidate rule uses the known feedback function. Its distinct contribution is the progress guarantee: all 1,241 guesses drawn from the verified pool had positive information gain, as Proposition 2.7 predicts under its assumptions.

AuthorsOpenReview ↗

Response to Reviewer y2XD (2/2)

W3 / Q3 — Candidate presentation, scoring, and reproducibility

In the revised CircuitDecoding runs, the elicitation prompt presents each candidate’s outputs on the queried inputs under permuted IDs. Appendix C provides the elicitation and shortlist prompts, parser, fallback rule and logged rate, token limits, and ambiguity-set sampling procedure. Algorithm 1 collects the control step in one place. Appendix D.4 and Algorithm 2 document the verification experiments and their analysis plans.

We also corrected the definition of α=1\alpha=1 to use nominal information gain. The original description as a mean over particles was inaccurate. The original ε=0.5\varepsilon=0.5 was selected after inspecting the evaluation episodes, and we withdraw the single-peak conclusion from the inconsistently scored sweep. Appendix D.3 reports the revised selection on a separate development split.


Your review changed how we interpret the work. The corrected experiments establish where elicited beliefs can still impair action selection, how much simple fixes recover, and what can be guaranteed when a candidate passes an explicit feedback check.

Reviewer_s41QOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper observes that information-gain maximization, a common strategy for active reasoning, can induce belief trap regions under belief approximation error. To mitigate this failure mode, it proposes Distributionally Robust Information Gain (DR-IG), a decision-time objective that optimizes worst-case information gain over an ambiguity set around the current belief estimate. The paper presents both theoretical results and empirical studies across five LLM families to demonstrate the effectiveness of the proposed approach.

Critical strengths

S1. The paper shows theoretically that robust worst-case control restores the negative drift of a progress potential under bounded belief error whenever uniformly informative actions exist.

S2. Empirical studies over diverse LLMs demonstrate the effectiveness of the proposed approach, with in-depth and diverse analyses.

Critical weaknesses

W1. The main text presentation can be made more self-contained and clear. The notations in section 2 are not introduced in the main text and are only defined in the Appendix. Also, it will be better to include a pseudo code to have a central summary of the algorithm and training process.

Questions for authors

N.A.

3 comments
AuthorsOpenReview ↗

Response to Reviewer s41Q

Thank you for pointing out where the paper was difficult to follow and for suggesting a single pseudocode description of the pipeline. We revised the presentation and also want to correct the theoretical claim you highlighted in S1.

W1 — Self-contained notation and pseudocode

Section 2 now defines the belief, belief proxy, information-gain objective, and ambiguity set where they are used. Algorithm 1 brings together belief elicitation, parsing and fallback, action scoring, and shortlist selection. Appendix C supplies the prompts and implementation details. Algorithm 2 describes the verified-candidate procedure, including how candidates are checked against feedback and used as actions.

S1 — Scope of the theoretical guarantee

The submitted version did not establish the negative-drift result stated in the abstract, and the original Theorem 2.3 was incorrect. We removed that abstract claim and revised the theory in §2.2. Proposition 2.4 gives a conditional progress guarantee for the exact DR-IG objective: the ambiguity set must contain the true belief, and some available action must have uniformly positive information gain over the set. Proposition 2.6 gives a way to check when the coverage condition fails using the mass assigned to hypotheses already refuted by feedback. In the corrected GuessNumber runs, it fails in 84–100% of steps.

A separate guarantee applies when a candidate that passes the feedback check is tested directly. Under Proposition 2.7’s assumptions, that action has positive information gain without requiring a bound on the belief error. In the revised GuessNumber experiments, all 1,241 guesses drawn from the verified pool satisfied this prediction. On 300 fresh secrets, the verified rule and a free-form agent that reasons before guessing reached 79.0% and 78.3%, respectively (Table 10). This comparison supports a claim about certifying progress, with no measured accuracy advantage over the free-form agent.

S2 — Corrected scope of the empirical findings

Our audit also found that a parser error caused every CircuitDecoding belief step in the original runs to fall back to a uniform prior. We withdraw the affected experiments and document the error in Appendix E. In the corrected runs, the effects are model-dependent. DR-IG improves one CircuitDecoding result from 7% to 36%, with the improvement over IG-max replicated on fresh secrets. Repeat masking performs similarly in the reported comparison. Section 3 now presents these corrected experiments and the simple baselines alongside the theory.


Your suggestion to make the pipeline explicit was especially helpful here. Showing how the controller obtains and uses its belief makes both the experimental failure and the assumptions behind the revised guarantees easier to inspect.

Thank you for your honest reply. I'll keep my rating. Please make sure you correct and address all these issues in the next version of the manuscript.

Great thanks for your consistent support! We will definitely incorporate the revision into the next version of the manuscript.

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

The paper looks at LLM agents that pick actions by maximizing information gain (IG-max) with respect to a belief that the LLM verbalizes, rather than the true posterior. They argue that errors in this verbalized belief make the agent pick the same uninformative action over and over, and they formalize this as a Belief Trap Region (BTR). Thm 2.3 constructs a belief error under which IG-max gets trapped. The proposed fix, DR-IG, scores each action by its worst-case information gain over an L1 ball around the verbalized belief. On CircuitDecoding and ConceptInduction with a top-5 shortlist of actions, DR-IG removes the loops and solves some episodes where IG-max solves none, while an oracle with the exact posterior solves all of them.

Critical strengths
  1. The empirical diagnosis is careful. The oracle arm with the exact posterior solves everything with the same pipeline, so the paper convincingly isolates the verbalized belief as the cause of the looping, and the appendix analysis of trapped episodes (frozen belief, identical shortlist, zero oracle information gain) is quite thorough. This is why I think there is a real finding here even though I am recommending rejection.
  2. The ablations separate breaking the repetition from actually finding an informative action, since randomizing the shortlist also breaks the loop but still solves nothing. This matters because it is the only evidence that DR-IG does more than add noise.
Critical weaknesses
  1. The theory does not hold up. Unless I am gravely mistaken, Step 3 of the proof of Thm 2.3 uses Lipschitzness for an upper bound when it only gives a lower bound, uses IG(b_0, a_1) <= eta while assumption (iii) says >= eta, and justifies "-L delta >= eta - L delta" by eta >= 0 when it needs eta <= 0. A simple counterexample (IG(., a_1) constant above eta on the ball, IG(., a_0) zero) is Lipschitz and never inverts. On top of this, the abstract promises that worst-case control restores negative drift of the progress potential and I could not find this stated or proved anywhere. The natural Lipschitz version would only apply when delta < eta/(2L), which is exactly when IG-max cannot be inverted either, so it would not separate the two methods. The theory is advertised as a main contribution, so this is decisive for me.
  2. The paper's own Table 9 shows plain IG-max with K = 10 or 20 solving around half of the CD episodes while DR-IG at the same K drops to around 10%. The authors' remark that at K >= 10 the LLM's judgment can override the corrupted ranking reads to me as saying the trap is mostly a consequence of forcing a choice from IG's top 5. Combined with the missing consistency-filtered baseline (the likelihood is known in every main task, so one can simply discard hypotheses that contradict past observations, which is roughly what BED-LLM does), I am not sure what problem DR-IG is solving that a larger shortlist or a filter does not.
Questions for authors
  1. How do you repair Step 3 of Thm 2.3, and is the intended threshold eta/L or eta/(2L)? Under what relationship between eps, delta, eta and L does DR-IG provably avoid a trap that IG-max falls into?
  2. Given Table 9, why should one use DR-IG with K = 5 rather than IG-max with a larger shortlist? Could you compare the methods at the K that is best for each?
  3. How was eps = 0.5 chosen? In the sweep the solve rate is nonzero only at that value and there is no held-out split. Suggestions rather than questions: the novelty framing should be adjusted, since the BTR formalism and CircuitDecoding already appear in Zou et al. (2026b) and worst-case EIG over an ambiguity set has precedent in the Bayesian experimental design literature (Go and Isaac; Abdulsamad et al.). The DR-IG average in Table 1 does not match the per-model values.
2 comments
AuthorsOpenReview ↗

Response to Reviewer qckG

Thank you for identifying the errors in Step 3 and for pressing us on the large-shortlist results. We checked the proof, the saved trajectories, and the evaluation code in response to your review. This led us to withdraw the original theorem and the affected empirical claims. Your questions also helped us identify which conclusions survive after those corrections.

W1 / Q1 — The threshold and the guarantee for DR-IG

Your counterexample is valid. The Lipschitz assumption bounds the magnitude of a change in information gain and cannot guarantee a ranking inversion. Step 3 also used the wrong direction for the assumption involving η\eta and an invalid inequality. The original Theorem 2.3 has therefore been withdrawn.

The corrected threshold is η/(2L)\eta/(2L), and it is a necessary condition for a zero-progress choice. If an available action has true information gain at least η\eta, and the proxy is within distance δ\delta of the true belief, Proposition 2.2 bounds the true information gain of the nominal IG-max choice below by η−2Lδ\eta-2L\delta. Proposition 2.3 makes the additional condition for a persistent trap explicit: the proxy must continue to rank the uninformative action first along the trajectory.

Proposition 2.4 gives the guarantee we can establish for DR-IG. It applies to the exact robust objective when ε≥δ\varepsilon\geq\delta and some available action has uniformly positive information gain over the ambiguity set. As you anticipated, with a common Lipschitz constant the resulting bounds give no regime in which IG-max can make zero progress while DR-IG is guaranteed to progress. Appendix B.4 gives the action-dependent case. Proposition 2.5 further shows that under deterministic binary feedback, exact DR-IG and nominal IG give the same action ranking. We have removed the original abstract’s negative-drift claim.

W2 / Q2 — Larger shortlists and consistency filtering

Your reading of Table 9 was correct. Under a common solve criterion, the original trajectories give IG-max solve rates of 43.3% at K=10K=10 and 46.7% at K=20K=20, versus 10.0% for DR-IG at either size (Table 13). In those runs, the larger shortlist prevents queries from repeating. Masking previously asked queries also prevents repetition at K=5K=5.

We have not independently selected the best KK for each method and compared them under the corrected elicitation pipeline. The original trajectories cannot answer that question: the parser was configured for a different task, and all 25,596 CircuitDecoding belief steps fell back to a uniform prior. Our corrected runs instead compare DR-IG with repeat masking, smoothing, a random shortlist, and proxy-level consistency filtering. DR-IG improves the result for the CircuitDecoding model that stalls under IG-max from 7% to 36%, with the improvement over IG-max replicated on fresh secrets. Against repeat masking, the pooled CircuitDecoding comparison yields 30 versus 27 discordant pairs (Table 7). Taken together, these results give no empirical basis to prefer DR-IG at K=5K=5 over a larger IG-max shortlist or repeat masking.

Filtering the full hypothesis space with the known likelihood corresponds to our oracle baseline. We also tested your suggestion at the level of the hypotheses actually named by the model. On GuessNumber, filtering that proxy before IG scoring produced 62.3% identification within six queries, versus 59.7% for IG-max. Testing the verified candidates directly reached 85.7% on the same design secrets. On 300 fresh secrets, the prespecified comparison with a free-form agent that reasons before guessing yielded 79.0% versus 78.3% (47 versus 45 discordant pairs, p=0.92p=0.92). The replication covers the comparison with free-form, and the verified-candidate rule uses the known feedback function. Its contribution is a progress guarantee under Proposition 2.7’s assumptions, without enumerating the full hypothesis space.

Q3 — Choice of ε\varepsilon and other corrections

The original ε=0.5\varepsilon=0.5 was chosen after inspecting the evaluation episodes. The apparent single peak also depended on an inconsistent solve criterion. We withdraw that sweep conclusion. In the revised tuning experiment, we select ε\varepsilon and the smoothing weight on a separate development split. DR-IG then reaches 51%, versus 56% for both smoothing and masking (Appendix D.3).

The revision credits Zou et al. for the BTR formalism and CircuitDecoding, and Go and Isaac and Abdulsamad et al. for prior work on robust information-gain objectives. The corrected average for the original Table 1 is 21.2%. Under the common scoring criterion, the random shortlist reaches 33.3% and the free-form agent reaches 38.0%, so we also withdraw the original interpretation of those ablations.


We appreciate how precisely you identified the limits of the original argument. The revised paper’s contribution is narrower: it shows how feedback can reveal that a robust guarantee’s assumptions have failed, and when testing a verified hypothesis gives a progress guarantee of its own.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Meta-response to the Area Chair and Reviewers

Thank you for examining the proof, the identical loops across models, and the comparison with simpler controllers. Your reviews prompted us to audit the saved trajectories and evaluation code. The audit changed our conclusions. We have withdrawn the affected results and ask that the revision be evaluated on the corrected experiments and the narrower guarantees described below.

What we corrected

  • Belief elicitation failed in the original finite-hypothesis runs. The parser expected four-digit answers from another task. On CircuitDecoding, all 25,596 belief steps fell back to a uniform prior, and a replay without an LLM reproduces the original loop signature (Appendix E). We withdraw the attribution of those loops to the models’ elicited beliefs.
  • Several original comparisons used an inconsistent solve criterion. Rescoring under a common criterion gives 38.0% for the free-form agent, 33.3% for a random shortlist, and 21.2% for DR-IG (Table 13). We withdraw the affected ablation and sweep conclusions. The original ambiguity radius was selected after inspecting the evaluation episodes. The revised tuning experiment uses a separate development split.
  • Theorem 2.3 was incorrect. A Lipschitz bound cannot establish a ranking inversion. The revised bound makes η/(2L)\eta/(2L) a necessary threshold for a zero-progress IG-max choice, and the revised trap statement explicitly assumes that the proxy keeps ranking the uninformative action first. We removed the abstract’s negative-drift claim. The revision also credits prior work for the BTR formalism, CircuitDecoding, and worst-case expected information gain.

What survives after the corrections

With verified elicitation, parse failures account for at most 2.4% of steps in the main tables. IG-max still underperforms a belief-free random shortlist within ten queries on three of five GuessNumber models (78% vs. 95%, 52% vs. 91%, and 60% vs. 92%). On CircuitDecoding, one model reaches 7% with IG-max versus 33% with a random shortlist. The effect is model-dependent, and IG control helps on the strongest model (Tables 1–2).

DR-IG raises the stalled CircuitDecoding result from 7% to 36%, with the improvement over IG-max replicated on fresh secrets. Repeat masking performs similarly in the reported comparisons. With development-set tuning, DR-IG reaches 51%, versus 56% for both smoothing and masking (Appendix D.3). We have not compared independently selected shortlist sizes on held-out episodes under the corrected pipeline. The available evidence does not establish a reason to prefer DR-IG at K=5K=5 over a larger IG-max shortlist or repeat masking.

What the revision contributes

The revised theory gives a way to test an assumption behind robust control. Proposition 2.4 guarantees progress for the exact DR-IG objective when the ambiguity set contains the true belief and an available action has uniformly positive information gain over the set. Proposition 2.6 shows that belief error is at least twice the proxy mass assigned to hypotheses already refuted by feedback. This bound rules out coverage by the chosen ambiguity set in 84–100% of steps on GuessNumber. Proposition 2.5 also shows that exact DR-IG and nominal IG induce the same action ranking under deterministic binary feedback.

Feedback consistency also supports a different controller. Under Proposition 2.7’s assumptions, directly testing a verified candidate guarantees positive information gain without a bound on the elicited belief error. On the GuessNumber design secrets, filtering the proxy before applying IG yielded 62.3% identification within six queries versus 59.7% for IG-max. Directly testing verified candidates reached 85.7%. All 1,241 queries drawn from the verified pool had positive information gain, as the proposition predicts. On 300 fresh secrets, the prespecified comparison against a free-form agent that reasons before guessing yielded 79.0% versus 78.3% (47 versus 45 discordant pairs, p=0.92p=0.92). That replication covers the free-form comparison only. The verified-candidate experiments cover one task and one model family, and the controller uses the known feedback function, which the free-form agent was not given.


We appreciate the opportunity to correct the record. We ask the Area Chair to evaluate the revision on what the corrected evidence supports: model-dependent degradation from elicited beliefs, a runtime check for failure of the robust coverage assumption, and a progress guarantee for directly testing verified candidates under the stated feedback conditions.

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

Related papers

Discussion (0)

Sign in to comment.