- Rating
- 3
- Confidence
- 2
- Informativeness
- 3
- Presentation
- 3
- Soundness
- 3
Read review
Paper summary
- The authors study the effect of correcting mistakes an agent has made in the past and to what effect it has on future actions.
- They describe a simple state-space based examples and show that even if the source becomes inaccessible (i.e. falls out of 'context' window, for example), it can leave traces in the subsequent states the agent visits such that there is no uniform recovery time of the influence of the mistake.
- They also give conditions for fading/persistence of the effect and show with help of a simple non-linear feedback model how a parameter can control the threshold of where the influence persists or fades away.
- They run experiments with a small maths dataset and with a subset of the SWE-benchmark to benchmark the behaviour of LLM models and show reduced mistake reuse by doing repairs that go beyond just source correction.
Critical strengths
- Excellent presentation of simple ideas.
- Elegant explanatory examples and an interesting experimental setups for evaluating their claims.
- Interesting theoretical results which establish no uniform recovery time, bound differences between continuations sharing a common history, establish conditions for fading, and population persistence.
Critical weaknesses
- The experimental design is neat, it removes the Monte Carlo error and makes W, SWAP and C exact for the recorded policy. The binary choices keep the tree enumerable. However, it is unclear how agent-like this is. The model only makes choices rather than reasoning; and "agents reusing their own mistakes" is a bit of a stretch given that the first token scores may not align with what the agent would've actually done. Maverick's menu sensitivity (Appendix I.3) of 0/100 and 100/100 also suggests that suggests position bias may dominate in these cells. Similarly, depth 4 is a very short to be able to talk about agent's long term behaviour and 6 problems (the repair experiments) may not offer sufficient diversity.
- The SWE setup, while elegant, also don't address the "agents reusing their own mistakes" either because the actions are planted. The planted effect may be in-context instruction following?
Questions for authors
-
The notion of 'accessibility' is never formally defined; only Appendix B makes it concrete: A one-line definition, e.g., “the source is accessible at t if the rendered input under Q^1_t and Q^0_t differs for some k”, would help.
-
To be able to actually study "agents reusing their own mistakes" in the wild; I can only imagine an observational study done by means of data-mining a large repository of multi-step agentic traces. Given such a repository; can the authors describe how one may design the experiments to test their claims? This may be a future work.
-
A visual illustration of the Slacks mentioned in the results near line 394--397 would help; as would a tree schematic of W, SWAP and C, ideally annotated with the path probabilities with a real math problem overlayed on it. That would help in understanding D vs. M.
-
The details of the 'full repair' arm of the experiment are missing. It would help to have an example prompt showing what the repaired ledger and repair record look like. Having the full protocol as an dedicated appendix section would help.
-
The coding study doesn’t map onto D vs. M because is lacks the C and SWAP arms. Is it possible to design experiments which map directly on to the decomposition to make the experiments more coherent?
Minor Nit: the use of H to indicate a state was mildly confusing in Theorem 2.2
2 comments
Thank you for the detailed rebuttal and the additional experiments. I’ll keep my assessment, but a few points would benefit from clarification:
-
Repair effect: I might be wrong, but the original paper doesn't explicitly call out which model was used; so I am at a loss as to whether a model change caused the drastic drop from 32-57pp (values in abstract) to 9pp? Is it because DeepSeek had any particular property, like high trail effects? The model choices across experiments could use some justification.
-
Self-generated mistake: At turn 5, the violation rates imply D = 12.6pp and M = 6.2pp, meaning D > M? In the injected coding pilot, D is -ve (SWAP > W). Is that expected or can that be explained? Further:
The direct and trail effects went in opposite directions … in 0/720 Maverick settings with complete scores, 3/1,200 Qwen settings, and 236/1,200 DeepSeek settings. In these cases, the trail effect often outweighed the direct effect and reversed the overall direction.
Does “reversed the overall direction” mean that sign(Δ) ≠ sign(D)?
- Model behaviour: Do the authors have a hypothesis for why DeepSeek shows this pattern so much more often than the others? What might be causing the differences between models?
We appreciate your interest in our theory. We've conducted additional experiments and clarifications for your concerns.
1. Accessibility.
We agree that we should have defined source accessibility earlier. In the main paper, we will define a source as accessible at time t if changing only the source changes the agent's rendered input for some fixed state k. We will also state the assumption that the source cannot affect the agent's decisions through anything outside the rendered input and K_t. We will rename the states H/L/N to s+/s−/s0 to avoid using H for both a state and the horizon.
2. Studying mistakes in existing traces.
We agree with the idea that a large study of naturally occurring mistakes would be useful. Such a study would be outside the time scope of this rebuttal session, so it would indeed be a future work. What could be done is for a given repository, we would identify mistakes using a fixed rule before examining later outcomes, restoring the checkpoints before these mistakes, and run W and C. For future checkpoints, we would construct SWAP by correcting only the source while keeping later transcript and workspace. Recorded traces alone wouldn't establish these comparisons without executable replay.
3. Slack plots and the W/SWAP/C tree.
We made the requested figures and will add them to the revised paper, since we cannot attach them here. For the tree, we used a real root-sum problem to show the probabilities of each path and mistake, alongside how we calculate D and M. For the slack plots, we show which assumptions held, which failed, and whether the bounds gave useful information.
4. Full-repair prompt and protocol.
We checked the exported prompts against the original requests and made sure that they match exactly. The full repair adds a record to the source-correction prompt with the corrected workspace, old and new values, and labels showcasing which results were corrected. Take for example the root-sum problem, its record changes the stored sum from -1 to 0. It then marks the quantity as invalid and then recomputes, while keeping the earlier result for reference and stating that it has been replaced. The executor computes these corrections, without adding any additional model decisions or executing future operations. We will include the exact prompt and full protocol within the appendix for the revised paper.
5. Coding comparisons for D and M.
For this experiment, we implemented W, C, and SWAP on five SWE-bench tasks with Qwen3-Coder, with two W/C pairs for each task and 480 next-turn probes. W retains the injected command, C corrects its destination to /tmp before execution, and SWAP edits only the source command while keeping W’s later transcript and workspace.
C had no observed violations. Effects were concentrated in two tasks, and task-bootstrap ranges included zero. This pilot used a different provider route. Further details are in our response to reviewer m6qA.
6. Menu choices and repair diversity.
Here we compare the original menu, reversed options, and free-form responses at 12 different Maverick checkpoints. At 5 checkpoints, where the original menu retained its mistaken convention, 38/38 parseable free-form responses also retained it. However, 1/12 of the checkpoints showcased a position effect. This supports agreement within the tested cases, though does not establish reasoning over many steps without being given fixed answer options.
We also expanded the repair study to 30 new DeepSeek problems. The original package reduced target error by 9.1 pp as opposed to the source-only correction (95% CI: 4.72–13.66). Workspace correction alone increased errors, while combining it with a value-free repair record gave no clear improvement. The full comparison is in our response to reviewer m6qA.
7. Testing a mistake made by the agent itself.
We also ran Qwen3-Coder on SWE-bench tasks without injecting a mistake. While working on astropy-13977, the agent created a backup inside the repository instead of /tmp. We saved the checkpoint and applied W/C/SWAP to the command the agent had generated. After changing the source command to /tmp, SWAP still used the backup inside the repository in 4/16 probes on turn 15, yet C didn't use them at all. So, by fixing the command, the agent did not stop using the file created by the earlier mistake.
Using this backup was allowed. We also checked whether the agent broke the rule again: at turn 5, the violation rates were 18.8% for W, 6.2% for SWAP, and 0% for C. We found no violations at turns 15 or 30. So, the agent was still using the file from its own mistake at turn 15, even when we no longer observed new violations.
We hope these additions address your concern.