acceptodds
Under review as a conference paper at ICLR 2027

When Failure Banks Kill Good Ideas: Calibrated Negative Memory for Autonomous Research

Abstract

Shared failure banks turn noisy experimental outcomes into reusable negative knowledge. We study when an irreversible binary failure record becomes a wrongful conviction: a genuinely improving idea is rejected and never retried. Under success-gated proposal supply we show theoretically that this local Type-II error can extinguish an entire research lineage—a deference phase transition and, at finite budget, an inverted-U relationship between deference and discovery; under exchangeable supply, sharing remains beneficial. Across two live ML loops and 122,817 benchmark convictions, wrongful convictions concentrate on marginal improvements, at a rate set by boundary density rather than effect variance; verification-stage near misses are 35–50% wrong and do not vanish as measurement noise shrinks. In controlled choice experiments, twelve deployed LLMs from nine families treat DEAD as a directive: they revisit 0/94 marked ideas and suppress an a-priori-best option despite identical evidence. Replacing binary records with evidence-graded, retryable statuses yields 30–32% more truth-validated discoveries in a controlled success-gated loop (8 replicates), while the advantage reverses in a low-noise, short-horizon cell. These results identify a conditional systems risk and a practical fix: make irreversible locking depend on recorded evidence and continuation value, with an explicit reopening path. The intervention is a schema change, not a model change.

Primary area:
foundation or frontier models, including LLMs
Keywords:
autonomous research agents, negative knowledge, shared memory, false negatives, wrongful convictions, branching processes, LLM agents, evidence-graded memory, AI scientist, agent memory

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.60 over 5 reviews, above 10% of the others.
Reviewer_DDaiOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper studies how shared failure banks in autonomous research systems may prematurely eliminate promising research directions. The authors propose evidence-graded negative memory that allows uncertain failures to be revisited.

Critical strengths

The problem is relevant, and the paper includes extensive theoretical and empirical analyses.

Critical weaknesses
  1. I find both the problem formulation and the proposed method to be incremental. Noisy failures can lead to premature rejection, and retaining evidence and allowing reconsideration are natural remedies. Although the paper formalizes this risk, I am not convinced that it establishes a sufficiently substantial advance over existing uncertainty-aware memory and sequential decision principles.

  2. The connection between the theory and the evaluated method is incomplete. The theory emphasizes the value of future research opportunities, but the experiments mainly use a fixed posterior threshold. They do not directly establish whether adapting the threshold to this future value provides additional benefit. It is unclear what new algorithmic insight is provided beyond standard uncertainty-aware decision-making.

  3. The strongest gains come from controlled experiments with an algorithmic proposer and enforced locks. Results with LLM proposers are mixed: graded memory improves the enforced DeepSeek setting, but not the soft-lock setting, and performs worse on average than binary memory in the enforced GPT-4o setting. These results support a conditional effect, but leave unclear how to identify systems where the intervention will help.

  4. The presentation makes the contribution difficult to assess. The organization is confusing, important findings are buried in excessive details, and the presentation of experiments and figures could be substantially improved.

Questions for authors
  1. Which theoretical or empirical finding provides the main advance over existing uncertainty-aware memory and sequential decision principles? Please clarify what new knowledge this finding establishes beyond the intuitive risk of premature rejection.

  2. Is future research value explicitly used in the evaluated method? If so, which analysis or experiment demonstrates its contribution? If not, how does this theoretical principle inform the fixed-threshold method used in the experiments?

  3. Do the current experiments identify conditions under which graded memory should improve discovery outcomes in a new research system? How do these conditions explain the different results for DeepSeek and GPT-4o?

Reviewer_C43UOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper argues that failure banks in autonomous research systems can permanently remove good ideas, as one noisy negative result marks an idea DEAD, and it is never retried. This claim is supported through a simple statistical model, small training loops and a study showcasing that LLMs avoid DEAD-labeled ideas regardless of evidence. The proposed remedy is evidence-graded, retryable records, which give about 30% more discoveries in one controlled loop.

Critical strengths
  1. Across twelve LLMs from nine families and two domains, a DEAD level suppresses selection of an idea even when the evidence is identical to that of an idea labelled inconclusive.

  2. The paper reports negative and reversed results alongside positive ones.

Critical weaknesses
  1. It is quite well established that noisy test results reject small improvements and that elimination needs confidence bounds. The proposed fix is a schema change. Clarifying what the paper adds beyond these results, for example by centring the label-deference finding, would strengthen the contribution.

  2. The gain appears in a constructed algorithmic loop. It doesn't carry over to the LLM proposer settings, the low-noise setting or the deployed bank of [1] (App. B). Showing the benefit in a realistic success-gated system or predicting in advance which settings will benefit would substantially strengthen the claim.

Questions for authors
  1. Beyond the established results on noisy selection and confidence-based elimination, what do you see as the paper's main new insight?

  2. Is there a deployed success-gated system where evidence-graded records improve outcomes?

  3. Can a loop's position relative to the phase boundary be predicted from measured quantities before it is run?

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

This paper investigates a failure mode of shared negative memory in autonomous research agents: experimental failures caused by intrinsic noise may be recorded as "dead ends", thereby preventing genuinely useful ideas from being further explored. The authors use theory, simulations, LLM behavioral studies, and closed-loop experiments to show that such misidentified failure labels disproportionately affect marginal improvements and that LLMs tend to follow categorical failure labels rather than the underlying evidence. They propose evidence-graded, retryable failure records to mitigate this issue, which improves discovery in certain settings.

Critical strengths
  1. This paper touches an interesting topic of what if the so-called "shared negative memory" of autonomous research agents were wrongly labelled (e.g., due to experiment noise). Personally I think the impact of this problem extends beyond the autonomous research scenario: any agentic search framework (e.g., [1]) has to decide whether an already explored direction worth further exploitation given existing results.

References

[1] An AI system to help scientists write expert-level empirical software, Nature (2026).

Critical weaknesses
  1. This paper itself reads like an artifact of end-to-end autonomous research. I'm not against AI-assisted writing, but in its current form the paper is full of obscure expressions in almost every individual sentence, making it very hard to follow. At the high level, the paper is poorly organized. In particular, it is unclear how those different sets of experiments were connected to support the main claim, and what their individual roles are for serving that purpose. Overall, the manuscript needs considerable revision to meet the writing standard of prestigious venues like ICLR.

  2. Those dense maths in Section 4 are unnecessary, as most ideas delivered there can be explained in one intuitive sentence.

  3. The only end-to-end result is obtained in a highly constructed setting where success-gated proposal generation and irreversible locking are both explicitly enforced. These are exactly the assumptions required by the proposed failure mode, so the result feels somewhat circular. More importantly, the conclusion is not even consistent across LLM proposers: graded memory outperforms binary memory with DeepSeek, while the ordering reverses with GPT-4o.

  4. The only realistic experiment does not actually support the proposed intervention. When the authors modify the original negative-knowledge system with OPEN/CLOSED calibration, the calibrated version achieves 12/15 successes versus 13/15 for the original negative-only condition. The paper itself states that the primary hypothesis is not supported in this experiment.

  5. The empirical validation is also very limited in scope. Most actual experiments are conducted on small ML training or hyperparameter-optimization loops, such as tiny-Shakespeare/charGPT and tabular MLPs. It is unclear whether the proposed failure mode generalizes to broader autonomous scientific research settings.

Questions for authors

See above.

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

This paper explores the potential risks associated with the shared failure memory of autonomous scientific research agents. By analyzing this issue, the author proposes a negative memory mechanism based on evidence calibration and allowing for revalidation, which enhances the efficiency of scientific discovery under specific conditions.

Critical strengths
  1. This paper reveals the issue that shared error and failure memories among multiple agents may inhibit innovation, which has practical significance.

  2. The author combines theoretical analysis, simulation, and large-scale model experiments to verify the influence of negative memory from multiple perspectives.

Critical weaknesses
  1. However, the degree to which their assumptions, such as success-driven proposal generation and irreversible rejection, align with real scientific research systems remains unclear.

  2. The returns in real large model agent experiments are limited, and in some scenarios, traditional methods even outperform.

Questions for authors
  1. Is there a widespread problem of failed memory of research hypotheses in real scientific research systems?

  2. Why hasn't the 30%-32% improvement been consistently replicated in real-world large model agent experiments?

  3. Compared to existing confidence decision-making and sequential decision-making methods, what is the core innovation?

  4. How to estimate key conditions such as noise level, exploration behavior, and budget in practical applications?

  5. Can the method be extended to open-ended scientific research where experimental costs are high and results are difficult to quantify?

Reviewer_x7raOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates a potential risk of shared failure memory in autonomous research: genuinely beneficial ideas may be permanently discarded following noisy experimental outcomes. When new candidates can only descend from confirmed successes, such errors may also eliminate entire research lineages. The authors characterize this mechanism through theoretical analysis, replays of real training experiments, behavioral studies across multiple LLMs, and controlled closed-loop search. They propose evidence-graded negative memory that permits reassessment. The approach improves discovery counts in some success-gated settings, although its benefits depend on noise, budget, and the proposer.

Critical strengths
  1. A clear and practically relevant problem. The paper connects experimental noise, permanent failure records, and downstream candidate generation, showing how a local evaluation error can have consequences beyond the discarded candidate.
  2. Converging evidence across several levels. Theoretical analysis, real-experiment replays, label interventions, and closed-loop search together provide converging evidence for the proposed mechanism, although the end-to-end gains are established mainly in the algorithmic loop. The experiments showing different choices under different status labels with identical evidence are particularly informative.
  3. Controls help identify the source of improvement. The volume-matched parole control indicates that, in the tested controlled setting, gains come from selecting retries based on evidence rather than simply increasing their frequency.
  4. Transparent reporting of limitations and negative results. The authors report reversals under low noise, differences across proposers, and an in-situ intervention that did not improve task outcomes, helping readers assess the scope of the conclusions.
Critical weaknesses
  1. Limited evidence of practical impact. The full pathology requires irreversible negative memory combined with success-gated supply. Only 1 of 8 surveyed systems has an explicit shared negative bank, and the in-situ intervention shows no outcome improvement (12/15 vs. 13/15), albeit near ceiling. Since the remedy draws on standard experimental design and sequential testing, demonstrating consequential failures in an existing research-agent system would strengthen the contribution.
  2. End-to-end gains rely mainly on a purpose-built loop. The 30–32% improvement comes from an algorithmic proposer with enforced gating and locking. Results with LLM proposers are mixed: discoveries remain flat under free supply, soft binary locks match or outperform graded memory, and under enforced locks graded memory wins for DeepSeek but loses for GPT-4o. The abstract and introduction should acknowledge this limited transfer.
  3. Missing simple baselines and component ablations. The paper lacks budget-matched multi-seed initial evaluations and a policy that retries each failure once before locking. Graded memory also changes evidence retention, evidence pooling, and reopening simultaneously. Comparisons with identical evidence but different locking policies would clarify the source of improvement.
  4. Narrow empirical scope and uneven statistical support. All executed loops involve small ML tasks; drug discovery is tested only through vignettes. Initial behavioral samples are small, although later replications strengthen some findings. Given the many comparisons and extensions, primary versus exploratory tests, effect-size intervals, and stopping plans should be clearer.
Questions for authors
  1. Beyond the reported parameter-misspecification tests, how should σ be estimated in practice, and how robust is the graded rule to configuration-dependent noise?
  2. Do real pipelines satisfy the assumed bounded false-positive window w? Can the authors measure how quickly false positives are identified and actually retired?
  3. Since the locking threshold depends on c/V, how should practitioners estimate the continuation value V of an exonerated idea?
  4. Does the label effect generalize beyond DEAD to deployed wordings such as “FAILED” or “ruled out”?
  5. As OPEN records accumulate, how do memory size, prompt cost, and retry allocation scale over long horizons?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.