acceptodds
Under review as a conference paper at ICLR 2027

ColdSplash: A Borrowed Thought for Agent Self-Reflection

Abstract

As language agents perform increasingly long and complex multi-turn tasks, ro- bust performance requires not only correct initial actions but also the ability to re- cover from errors through re-planning. Self-reflection is a common mechanism for recovery, but agents can remain anchored to the interpretation that caused the fail- ure, leading them to reproduce the same erroneous reasoning despite having access to relevant evidence in the trajectory. We introduce COLDSPLASH, a lightweight cross-model intervention that mitigates this anchoring effect by transferring only the opening of a reflection from a heterogeneous model. The original agent then completes the reflection and resumes the task using the same trajectory and ev- idence. Thus, COLDSPLASH changes the initialization of the reflection process without delegating the task or modifying the agent’s subsequent reasoning pro- cess. We evaluate COLDSPLASH on the self-reflection failure sets of τ 2-bench across Airline, Retail, and Telecom. On the pooled Airline/Retail cohort, a 16- token opening improves pass@any(4), the fraction of tasks solved in at least one of four attempts, from 5.2% to 9.4%, an 80% relative improvement, with gains observed across actors and domains. Diagnostic audits, trajectory-level analyses, and reflection-channel analyses further show that the intervention can change sub- sequent behavior rather than merely the linguistic form of reflection. These results demonstrate that a small, well-placed external nudge can unlock recovery capacity that agents already latently possess, pointing to lightweight cross-model reflection as a practical complement to scaling a single model’s introspection. Code and data is https://anonymous.4open.science/r/coldflash-1D84

Primary area:
generative models
Keywords:
Agent Recovery, Self-Reflection, Reflection Initialization, Cross-Model Collaboration, Tool-Use Reliability

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

This paper proposes ColdSplash, which uses another model to write a short opening for an agent’s reflection, while the original agent completes the reflection and continues the task. It evaluates the method on selected self-reflection failure cases from three τ²-bench domains, using several actor and helper models. The authors report higher four-attempt recovery rates than fresh self-reflection in most tested settings, with additional helper-model, random-seed, latency, and private-note evaluations.

Critical strengths

The method is simple and potentially useful: another model provides only a short reflection opening, while the original actor completes the reflection and continues the task without training. This makes it a practical idea for improving agent recovery with limited external assistance.

The evaluation covers several actors and three domains, with helper-model and random-seed comparisons. The repeated direction of improvement is encouraging, although the size and uncertainty of the gains need further validation.

Critical weaknesses
  1. The main experiment does not isolate better actor reflection. The reflection is visible to the user simulator, which can respond before the actor continues. The private-note experiment changes several other settings and gains only one additional recovered task, so it does not resolve this concern.

  2. The evidence for a reliable recovery gain is limited. Most improvements in Table 1 involve only one to four additional recovered tasks, and paired uncertainty estimates are not reported. The consistent direction across several settings is encouraging, but the current analysis is not enough to establish a robust benefit.

Questions for authors
  1. Clarification: Can you provide paired task-level confidence intervals and significance tests for the gains in Tables 1 and 4, accounting for repeated tasks across seeds? These would help establish how reliable the reported improvements are.

  2. Evidence needed for the current claim: Does the improvement remain when the reflection is private, while keeping the checkpoint, cohort, sampling settings, and budgets matched to the public-note experiment? This comparison would help separate better actor reflection from changes in the user simulator’s response.

Reviewer_63iiOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper identifies a limitation of simple prompted self reflection in agentic deployments of LLMs where an agent hallucinates problems when asked to reflect or the self reflection does not actually result in improved continuation. To tackle this a simple method leveraging an auxiliary second model is introduced. The second model generates the first 16 tokens of the reflection, then that beginning is used to prompt the original model's self reflection which continues from that point onwards. The method is evaluated on 3 open source models of sizes 3B-14B and on a subsets of the airline, retail, and telecom domains of tau^2-bench.

Critical strengths

Analysing failure modes of self reflection and addressing those is an interesting and relevant problem. The proposed method is simple and inexpensive.

Critical weaknesses

I have strong concerns regarding this paper. It is unclear whether the results are statistically significant and the entire motivation relies on a judge that is told a mistake exists introducing a bias and violating the paper's own evaluation principles of not leaking future outcome. Here are my main concerns:

  • The task subset is selected based on SR failing, and then SR is rerun and used as the baseline. This selection biases the entire evaluation in favour of any other method not used for the selection.
  • The above point paired together with the small changes of unclear statistically significance between SR and Coldsplash are serious cause for concern. The paper in fact acknowledges that no statistical significance is established in the caption of Figure 2.
  • The judge used to establish the motivation is told the trajectory contains at least one mistake which biases it towards raising issues even when it cannot find any. This undermines the motivation of the paper.
  • Important baselines are missing and would be useful to add: What happens if B generates the entire review, what happens if no self reflection is performed at all, what happens if a fixed generic opening is provided (I wonder how much the first 16 tokens actually vary). Also, at least some examples of the openings should be shown.
  • A comparison between the A's openings and B's borrowed openings would be helpful.
  • An ablation on k would be important (and in the extreme of full response also partly address the above point).
  • A justification for the choice of checkpoint would be useful.
  • Why does the private note setting differ such that no clean comparison can be made? This is a counterintuitive design choice.
  • The paper spends just under 2 pages explaining a quite straight forward method in my opinion overcomplicating it with mathematical notation. I think a simple method is a good thing and the space would be better spent otherwise.
  • The code link leads to an empty repo with a blank readme.
  • The paper mentions "reflection-channel analyses" and "trajectory-level analyses", could the authors point me to what sections exactly these refer to?

Beyond these hard issues, several additional concerns remain:

  • the writing is vague and hard to follow throughout the paper.
  • often hard to parse terminology is used e.g. check the caption of Figure 2, it was not clear to what any of it meant when I arrived at the top of page 3. Also, the caption of Figure 1 mentioned an archived fork experiment and the private note method which are not shown in the figure and confused me.
  • there are also some other presentational issues e.g. in Figure 2 with labels overlapping.
  • k and K denote unrelated quantities, this could be improved.
  • Why are the proposed prompts reported if they were not used to produce any results in the paper?
Questions for authors

See above.

Reviewer_6NYHOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

The paper proposes a simple technique to improve reasoning of language models in agentic tasks. In this method, the first kk (set to 16 by default) tokens of self-reflection parts of the reasoning traces is written by a secondary model. They show that this is superior to using the same model for the whole reasoning.

Critical strengths

The proposed method is very simple. I can see this something that can be tested in real setups with minimal cost and used if proven to effective. It is an inference-only modification that requires no fine-tuning or complex implementations.

The type of tasks this paper focus on, i.e. agentic workflows, is a significant and important application.

Critical weaknesses

Bad writing. This is one of my main concerns with the current draft of the paper. Reading the paper is not very pleasant. At a high level, the story is not well-thought and at a low level, many vague and meaningless terms appear in the text. Here are some examples:

  • The introduction is very vague and full of terminology that is not clear. There are paragraphs that feel displaced and do not add anything to the story. One example are lines 082 - 086, which are one of the last paragraphs of the introduction. The method is already motivated and introduced. Yet, these lines read like an opening with some philosophical viewpoint. The phrases "Our cognitive description", "This is an operational hypothesis", "not a measurement of internal belief" are not meaningful to me. It is clarifying that the work is beyond producing diverse text, but it is never mentioned that diversity is part of what the method achieves.

  • Vague or undefined terminology (e.g. "reflection checkpoint" (line 156), "hallucinated" in figure 2, ) is constantly introduced. Even though these terms have general meanings, the exact implementation for such categorizations should be explained.

  • At a high level, I cannot understand how either of the two failure modes in in Section 2 - Motivation are a motivation for the paper.

  • Comparison of novelty with related work. From the related work section, it appears that the same idea has been tested in reasoning traces. A better discussion on the exact novelty of this work would improve the draft.

Evaluations and investigation depth. The empirical evaluation and investigations can be improved a lot.

  • The technique is only tested on a single benchmark (τ2-bench). To establish the generality of the technique it is necessary to see it perform well on multiple settings.
  • Statistical signifcance of results. I see that some of the reported results are on subsets with small number of tasks. The improvement in some cases is down to solving one additional task out of 20. It is necessary to make sure this is beyond the variation caused by the noise. I only see Table 4 that is only for only one experiment configurations.
  • Robustness on the choice of models A and B. This is a very important investigation for the method. The paper does this in Table 3. I find the combinations that are tested to be limited. I wonder if the method remains beneficial once the role of two models are swapped, which is not tested. Also the actor is always a Qwen family model and the reflection source is always a Hermes model. These fixed choices raise questions on whether specific terminology Hermes models use may be the driving factor for the improvement.
  • Choice of the number opening tokens. The paper fixes the number of tokens generated by model B to 16. It would very appreciated if they showed a study on how this choice affects the results. Furthermore, I am very curious to see what these 16 tokens usually are. Are they just general english words to start an analysis of the path until that moment, or do they contain a trajectory-specific findings of the past mistakes?
Questions for authors

Please see weaknesses.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.