acceptodds
Under review as a conference paper at ICLR 2027

Tracing Context Ghosts in Agentic Replay Training

Abstract

Leveraging past interaction trajectories as training data provides a scalable way for agents to improve from their own execution histories. Yet data reuse pipelines often sanitize or shorten the recorded context while retaining the continuation produced from the original context as the training target. This practice introduces a hidden mismatch where the retained target still encodes behavioral effects triggered by the absent context. We term this residue a context ghost. We isolate this mechanism by independently varying the context used to generate the target and the context retained to train the model. This explicitly contrasts models fine-tuned on identical sanitized prompts but supervised by targets generated with and without extraneous metadata. Comprehensive evaluations reveal that target provenance shapes learned behavior even when the original influencing context is absent. Furthermore, simply cleaning the training prompt while keeping old targets amplifies these unintended behavioral shifts. Regenerating targets under cleaned contexts largely mitigates the problem and improves task success. Ultimately, robust agent replay training requires alignment between training inputs and the exact conditions that produced their supervision.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Large Language Models, LLM Agents, Agent Training, Learning from Experience, Replay Training

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_fmwkOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

When leveraging past interaction trajectories for training, preserving the subsequent target while modifying its context may lead to a hidden mismatch, where the retained target still encodes the behavioral effects triggered by the absent context. The paper identifies this phenomenon as a "context ghost" entering the target. To isolate the contribution of target provenance, the paper proposes an experimental framework that decouples the context used for target generation from the context retained during fine-tuning by only controlling the presence of metadata conditions (open: "Recovery status for the record: open"; closed: "Recovery status for the record: closed"; neutral: no recovery status is inserted). The paper demonstrates with experiments that (1) "context ghost" exists; (2) naive prompt sanitization can amplify such biases; (3) regenerating targets under cleaned contexts can effectively mitigate "context ghosts"; (4) "context ghosts" vary across metadata types, models, and environments.

Critical strengths
  1. The paper identifies and formalizes a subtle and interesting failure mode specific to agent replay training: modifying training contexts may not eliminate their behavioral influence in the corresponding targets that were generated under the original contexts.
  2. The experimental design is logically structured, progressively establishing the existence of context ghosts, analyzing their causes, evaluating mitigation strategies, and characterizing their generality.
Critical weaknesses
  1. The conceptual novelty of "context ghost" is not fully established. The phenomenon appears closely related to existing observations that supervision targets inherit information from their generation context, such as context distillation and biased imitation learning. The paper would benefit from a clearer distinction of what is fundamentally new in the agent replay setting.

  2. The evidence for real-world prevalence of context ghosts remains limited. Although the paper studies multiple metadata types, models, and environments, the evaluated metadata are still manually designed perturbations rather than naturally collected signals from deployed agent systems. It remains unclear how frequently such context-target mismatches arise in practical trajectory curation pipelines.

  3. The proposed mitigation may introduce additional data generation cost. Target regeneration requires rerunning agent trajectories under cleaned contexts, which can be expensive in large-scale replay pipelines. The paper demonstrates effectiveness, but the practical trade-off between mitigation benefits and additional rollout costs is not fully explored.

Questions for authors
  1. How do the authors distinguish "context ghosts" from existing concepts such as context distillation, biased imitation learning, or target leakage? Specifically, what property of agent replay training makes this phenomenon fundamentally different from these existing mechanisms?

  2. In practical agent training pipelines, how common are metadata fields that influence agent behavior but are later removed during curation? Do the authors have evidence from real-world trajectory datasets showing that this issue occurs beyond the controlled experimental settings?

3 comments
AuthorsOpenReview ↗

To Reviewer fmwk (1/2)

We thank the reviewer for the time and effort devoted to our paper, and for recognizing the identified failure and the structure of our experiments. We respond to each point below.

W1 / Q1. The novelty of the context ghost is not fully established, since supervision targets are known to inherit information from their generation context, as in context distillation and biased imitation learning. How does it differ from these and from target leakage, and what property of agent replay makes it different?

A. We agree that context distillation is the closest neighbor. It compresses useful information into the weights on purpose, while a context ghost is an accidental distillation of administrative metadata that agent replay curation performs. The known strength of intentional distillation is exactly what makes this accidental variant a credible threat.

What makes agent replay different from the concepts the reviewer names lies in what the learner observes and what the hidden variable carries.

In context distillation and privileged demonstrators (Weihs et al., 2021; Ortega et al., 2021), the hidden information is useful for the task. In agent replay, the hidden field, such as a recovery status label, carries no task information.

Furthermore, in causal confusion (de Haan et al., 2019), the learner observes a nuisance variable and latches onto it. In a context ghost, the learner never observes the nuisance variable during training because the pipeline scrubs it from the prompt.

Finally, while target leakage moves label information into the input features, a context ghost does the exact opposite. Context information leaks into the target actions and stays there even after the input is sanitized.

Because the ghost is baked into the actions during the original rollout, scrubbing the field's wording from the target's rationale (Sanitize, Table 4) leaves the bias in place. This is why a context ghost requires counterfactual rollouts to measure and regeneration under the cleaned context to remove.

W2 / Q2. Evidence for real-world prevalence is limited, since the evaluated metadata are designed perturbations. How common are fields that influence agent behavior but are later removed during curation, and is there evidence from real trajectory datasets?

A. This is a central question for practice. The curation steps that create this mismatch are already documented standard practice. Agent SFT recipes routinely truncate histories to retain only recent states, and deployed agents frequently prune or summarize their past observations (Section 2). Each step alters the training input while keeping the continuation written under the original context. Our experiments use controlled metadata intentionally because starting two rollouts from otherwise identical snapshots is strictly required for causal measurement.

To show that these conditions occur in the wild, we have added new evidence from real datasets and operators to Appendices F and G.

First, framework-written text routinely sits outside the loss in released data. Across four public agent SFT releases (nebius/SWE-agent-trajectories, SWE-bench/SWE-smith-trajectories, zai-org/AgentInstruct, and SWE-Gym/OpenHands-SFT-Trajectories), such text appears in almost every trajectory within user or observation turns. These include explicit behavioral directives like SWE-smith's pre-submit review request and OpenHands' strict prohibition against asking for human help.

Crucially, this text directly shapes the next action. On hundreds of SWE-smith prefixes, at the exact point where SWE-agent issues its pre-submit review message, we kept the diff and removed the review instructions. This single deletion drastically shifts behavior. Without the instructions, both Qwen3-4B and Qwen3-30B-A3B immediately submit or end the task in the vast majority of cases. With the message intact as released, both models instead delete their reproduction scripts nearly every time, mirroring the released trajectories.

These results show that the ingredients of a context ghost are common in real data, and Appendix F identifies compaction as one curation step that removes such a field. For any specific pipeline, the paired probe in Algorithm 1 can directly measure this prevalence rate prior to training.

AuthorsOpenReview ↗

To Reviewer fmwk (2/2)

W3. Target regeneration may add data-generation cost in large replay pipelines, and the trade-off between its benefit and the extra rollouts is not explored.

A. Regeneration replaces each affected target with one new continuation. When the environment can be restored to the failure snapshot, the continuation simply starts there. On our FeedMerge corpora, it starts from a median context of 1.2k to 1.5k tokens and writes between half and two thirds of the output tokens of the original rollout. Without restorable snapshots, the cost equals one full rollout per affected record, exactly what the pipeline already paid to collect it.

Two properties keep this cost bounded. The admission test runs only once per field type, so regeneration applies strictly to records carrying a field that measurably moves behavior. Furthermore, cheaper remedies fall short. Paired dropping requires an extra rollout as well, yet its trained model still exits after one step in 31.1% of episodes, compared to just 1.8% for regeneration in the same paired run (Table 4 and Section 4.4).

Thank you for your detailed responses and experiments. My concern in W3 has been addressed, but my main concerns still remain. The added experiments demonstrate that framework instructions can influence agent behavior (which seems to be an expected consequence), but do not establish how frequently real-world curation pipelines actually remove such context while retaining the original continuations. More fundamentally, I think the conceptual novelty of context ghosts beyond existing context-conditioned supervision mechanisms is not fully established. Therefore, I still prefer to maintain my original rating.

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

This paper identifies the “context ghost” phenomenon in agent replay training: metadata present during trajectory generation can shape the continuation, and its behavioral influence may remain in the training target even after the metadata is removed from the prompt. Using controlled counterfactual rollouts from identical failure states, the authors separate target provenance from training-input conditioning. Experiments across six models and three agent environments show that prompt sanitization alone can amplify unintended behavior, while regenerating targets under cleaned contexts usually mitigates it.

Critical strengths
  • Addresses an important problem for agent self-improvement pipelines.

  • Uses a clean factorial design to separate target provenance from input conditioning.

  • Controls environment snapshots and sampling seeds across metadata conditions.

  • Evaluates six models across three distinct interactive environments.

  • Uses six independent data/training pipelines and preregistered statistical tests.

  • Evaluates target regeneration and several cheaper mitigation strategies.

Critical weaknesses
  • The experiments apparently train on all generated continuations, including many with extremely low task success. This differs from the terminal-success filtering emphasized as common practice and may itself cause degradation.

  • The metadata is synthetic and semantically strong; “recovery status: closed” naturally encourages termination. Evidence from real operational logs is missing.

  • Training budgets were selected to move the closed-trained policy substantially toward its targets, potentially exaggerating effect sizes relative to standard replay training.

  • Several central analyses rely on continuation rate or episode length rather than task success. Longer execution may represent waste rather than recovery.

Questions for authors
  • Were training trajectories filtered by terminal success or verifier acceptance? Does the effect remain when only successful trajectories are retained?

  • What are the target success rate, average length, and termination frequency under each metadata condition?

  • How sensitive are the findings to weaker, more conventional fine-tuning budgets?

  • How were the nine cells used for the full factorial analysis selected?

  • Can the authors provide examples from real operational logs rather than constructed metadata?

  • Can target length or termination frequency be matched to separate metadata-specific transfer from generic short-trajectory imitation?

3 comments
AuthorsOpenReview ↗

To Reviewer Ecgw (1/2)

We thank the reviewer for the precise and constructive questions. We are glad that the reviewer found our experimental design clean and the evaluation broad. We address each question below.

W1 / Q1. The experiments appear to train on all generated continuations, including many with very low task success, unlike the terminal-success filtering common in practice. Were trajectories filtered by success or verifier acceptance, and does the effect remain when only successful trajectories are kept?

A. The paper trains on every continuation, which reflects realistic pipelines that retain unsuccessful trajectories for their training signal (Wang et al., 2024). Crucially, training on unsuccessful trajectories is not what causes the degradation. On Qwen3-4B, neutral targets succeed in only 7.8% of cases, yet the neutral-trained model improves the untrained model's success rate from 2.0% to 3.5% (Table 3). The degradation stems from the context under which the closed targets were generated.

When we apply a success filter to Qwen3.5-4B (Table 6), the gap narrows only because the filter discards the affected records. It retains four-fifths of open-generated targets but just over half of the closed ones, dropping below a tenth on gemma-4-31B. After filtering, the success gap is 3.6 points and stays positive in four of the six pipelines. By contrast, regeneration preserves these records entirely and recovers most of the lost success rate without any filter.

W2 / Q5. The metadata is synthetic and semantically strong, since "recovery status: closed" naturally encourages termination, and evidence from real operational logs is missing. Can the authors provide examples from real logs?

A. We use a constructed status label intentionally because causal measurement requires snapshots that differ in exactly one field.

However, its semantic strength is moderate. Switching the label from closed to open changes the untrained model's step count only one-sixth as much as explicit stop/continue instructions (Table 8). More importantly, a ghost forms even when the metadata says nothing about stopping. When a temporary metadata sentence classifies a failure as "external," the trained model's diagnosis rate plummets from 86.5% to 20.4%, and it spontaneously cites external causes when stopping, a behavior never seen otherwise (Table 5).

Regarding real operational logs, new evidence in Appendices F and G confirms that these conditions are common.

Across four public agent SFT datasets, framework-written workflow text (e.g., SWE-smith's pre-submit review request or OpenHands' "never ask for human help" rule) appears outside the loss in almost every trajectory.

This text directly controls behavior. On hundreds of SWE-smith prefixes, removing the review instructions causes Qwen3 models to immediately submit or end the task in the vast majority of cases. With the instructions intact, they instead delete their reproduction scripts, mirroring the released trajectories. Furthermore, a Llama-3.3-70B compactor summarizing failure histories to 100 tokens drops our status sentence in a third of the summaries while keeping the failure notice itself.

These results show that the ingredients of a context ghost are common in real data and that aggressive compaction removes such a field. Algorithm 1 can then measure this rate in any given pipeline.

AuthorsOpenReview ↗

To Reviewer Ecgw (2/2)

W3 / Q3. Training budgets were selected to move the closed-trained policy substantially toward its targets, which may exaggerate effect sizes. How sensitive are the findings to weaker, more conventional budgets?

A. The halfway criterion is meant as a floor on learning, not an amplifier. It ensures that the closed-trained model has genuinely absorbed its targets, guaranteeing that any small ghost observed does not simply result from undertraining.

In fact, the ghost is largest under the briefest budget. After a single epoch, the neutral-trained model continues 62.6 points more often than the closed-trained one, compared to a 46.5-point gap after the four epochs used in the paper. All six pipelines agree in direction at every training budget (Table 6).

Furthermore, the four Qwen bases meet this learning criterion using exactly one default recipe across the board. They use 400 records for four epochs at a learning rate of 1e-4, as detailed in Appendix B.

W4. Several central analyses rely on continuation rate or episode length rather than task success, and longer execution may represent waste rather than recovery.

A. Continuation rate serves as the primary metric because it directly records the decision affected by the label change: whether to keep working after a failure. The paper explicitly separates genuine recovery from wasted execution using two evaluation suites (Table 4).

On sealed failures (Suite B), where every extra call represents waste, the regenerated model makes only 0.27 more calls per episode than the base model. In addition, Tables 2 and 3 report the success rate for every cell. On Qwen3.5-4B in FeedMerge, regeneration successfully recovers 19.4 of the 26.9 points of success rate that closed-generated targets initially cost.

Q2. What are the target success rate, average length and termination frequency under each metadata condition?

A. Table 12 reports the open and closed success rates for all 18 cells. We list the mean steps and early exits here alongside the neutral targets for four FeedMerge cells.

Target success rate (%)

Modelclosedopenneutral
Qwen3-4B0.12.47.8
Qwen3.5-4B56.180.472.2
Qwen3.5-27B30.495.394.9
gemma-4-31B7.695.482.8

Mean steps

Modelclosedopenneutral
Qwen3-4B2.04.35.5
Qwen3.5-4B3.96.57.3
Qwen3.5-27B3.46.98.2
gemma-4-31B1.35.04.5

Exits within two steps (%)

Modelclosedopenneutral
Qwen3-4B78.235.59.6
Qwen3.5-4B45.08.52.8
Qwen3.5-27B55.93.01.8
gemma-4-31B91.84.216.8

Q4. How were the nine cells used for the full factorial analysis selected?

A. The 2x2 design was first run on Qwen3-4B on FeedMerge under criteria rigorously frozen before its first evaluation. It was then replicated systematically along environment, model family, and scale.

The full analysis covers 11 cells, and the main paper displays the nine covering its six base models. The other two cells (Qwen3-8B and gemma-3-4B on FeedMerge) show the same pattern. Provenance shifts by +19.7 and +43.4 points, conditioning by -26.5 and -6.5, making the sanitized configuration nc the lowest configuration in both. The revision adds these two cells to Table 13.

Q6. Can target length or termination frequency be matched to separate metadata-specific transfer from generic short-trajectory imitation?

A. Yes, we have matched both. We created closed- and neutral-generated training sets with exactly the same step-count distribution, set size, and early-exit rate. Even with perfect matching, the closed-trained model still continues 28.7 points less often (Table 6).

The Filter remedy in the main paper provides further evidence against generic length imitation. The closed targets under Filter are as long as the neutral ones and never stop within two steps. Despite this, the Filter-trained model exits after a single step in 14.9% of episodes, compared to just 1.8% for the neutral model in the same paired run (Section 4.4). If the model were merely imitating target length, this ordering would be exactly reversed.

Thanks for the response, and i will keep my score.

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

The paper identifies a failure mode in agent replay training which they call a context ghost, where the metadata removed from the training prompt still affects the trained model through the target that was generated with it. The authors propose a controlled setup that branches rollouts from the same failure snapshot with open, closed, and neutral status labels, and fine-tunes the model on the cleaned prefix with each target to separate the effect of the training prompt from the target provenance. The method is evaluated on three environments: FeedMerge, τ²-bench, and AppWorld with six open models from the Qwen3 and Gemma families. The authors show that models trained on the closed-generated targets quit early on 16 out of 18 settings and removing the label from the prompt makes it even worse. The authors also claim that regenerating the target under the cleaned context recovers most of the performance, e.g., the success rate of gemma-4-31B on FeedMerge goes from 4.2% to 77.6%.

Critical strengths

The experimental design is clean and isolates the effect well. Branching from the identical failure state and varying the generation context and the training context separately makes it clear that the behavior shift comes from the target itself. The statistics are also careful with six seeded pipelines per cell and pre-registered metrics. The finding that sanitizing the prompt alone makes the problem worse is useful for the people who build these pipelines since prompt-loss masking and context pruning are the common practice. The proposed fix is simple and only costs one extra rollout per record.

Critical weaknesses
  • The paper lacks novelty. The paper indicates this in Appendix E that the NC configuration is the same as context distillation, and the result that keeping the label in the prompt binds the behavior to the label is the inoculation prompting effect from Tan et al. (2025) which the paper also cites. However, regardless of the indication, the authors claims their method to be newly discovered hidden mechanism in the abstract and the introduction. The actual contribution is the measurement in the agent replay setting and the regeneration fix which is not clearly written.

  • The experimental setup is a worst case scenario. The metadata is a single injected sentence, the training budget is tuned so that π_closed moves at least halfway toward its targets, and the training set only consists of the affected records. Furthermore, the large success rate gains are only on FeedMerge which is the authors’ own environment.

Reference: Tan, D., Woodruff, A., Warncke, N., Jose, A., Riché, M., Africa, D. D., & Taylor, M. (2025). Inoculation Prompting: Eliciting traits from LLMs during training can suppress them at test-time. arXiv:2510.04340. https://arxiv.org/abs/2510.04340

Questions for authors
  • The first thing that I want to ask is whether the continuations are filtered by the task success before training. Standard replay pipelines filter by success and this might remove a large portion of the ghost by itself.
  • How does the effect scale when the closed-generated records are mixed with the clean records, e.g., 10% or 50%? This would tell whether this is a worst-case demonstration or a practical issue.
  • Can the authors explain more about why πneutral is worse than πopen for gemma-4-31B on FeedMerge in Table 4? Does this mean that choosing the generation context matters more than removing the metadata?
3 comments
AuthorsOpenReview ↗

To Reviewer sHqZ (1/3)

We are grateful for the reviewer's careful reading of our paper, and we appreciate the recognition of our clean design and the practical value of our findings. Our responses follow.

W1. The paper lacks novelty. Appendix E notes that configuration nc is context distillation, and the binding effect of a retained label is the inoculation prompting effect of Tan et al. (2025), yet the abstract and introduction present a newly discovered mechanism. The actual contribution, the measurement in agent replay and the regeneration fix, is not clearly written.

A. Appendix E and the abstract describe the same phenomenon at two different levels. Appendix E identifies the learning route, which is mechanistically identical to context distillation. We explicitly stated this connection because the known strength of distillation is exactly what makes an accidental version of it such a credible threat.

The abstract, however, uses "hidden" for the mismatch this route creates between a cleaned training input and a target that still encodes the removed context. In agent replay, this route is triggered unintentionally and remains hidden in three senses.

First, it is hidden in the record. The administrative field is removed from the prompt, but its effect stays entrenched in the target's actions. Scrubbing every mention of the status from the target's rationale does nothing, leaving the trained model at the uncleaned level (Sanitize, Table 4).

Second, it is hidden in the logs. A static log always couples a continuation with the context that produced it, failing to separate task difficulty from an inherited ghost. Counterfactual rollouts from identical snapshots make this separation. In fact, the shift a field causes at generation accurately predicts the ghost left after training across 18 cells (r = 0.92, slope near one), providing a test pipelines can run before training (Algorithm 1).

Third, it is hidden from intent. Context distillation and inoculation prompting intentionally add instructions to elicit or bind traits. Replay curation, by contrast, deletes an administrative field simply for hygiene. Our 2x2 design measures the transfer and the binding at once. It reveals that this hygiene step inadvertently keeps the transfer but loses the binding, moving raw replay down to nc, the worst of all four configurations across 11 factorial cells.

The contributions the reviewer describes are exactly the ones listed in our introduction. The first contribution identifies and formalizes the context ghost in agent replay training. The second provides the controlled measurement branching from identical failure states. The third shows that standard cleaning amplifies the ghost, whereas regeneration mitigates it when compared against filtering, paired dropping, and rationale sanitization (Table 4).

AuthorsOpenReview ↗

To Reviewer sHqZ (2/3)

W2. The setup is a worst case: the metadata is a single injected sentence, the training budget is tuned so that the closed-trained model moves at least halfway toward its targets, and the training set contains only affected records. The large success-rate gains also appear only on FeedMerge, the authors' own environment.

A. We appreciate this concern and have rigorously tested each condition to ensure our setup reflects realistic challenges.

Regarding the single sentence, one sentence is the smallest field a log can carry. Its semantic strength is also moderate. Switching the status label from closed to open changes the untrained model's step count only one-sixth as much as switching explicit stop/continue instructions (Table 8). Real framework text is often longer and more forceful. For example, SWE-smith's pre-submit review request contains four instructions, and simply removing them raises Qwen3-4B's direct submission rate from 3.2% to 92.2%, as shown in Appendix G.

Regarding the training budget, the halfway criterion acts strictly as a floor to rule out undertraining. In fact, a smaller budget yields a larger ghost. After just one epoch, the continuation rate gap between neutral- and closed-trained models reaches 62.6 points across all six pipelines. This is substantially larger than the 46.5-point gap observed after the four epochs used in the main paper (Table 6).

Since framework-written fields pervade real corpora, removing one during curation affects most records simultaneously. In four public agent SFT releases, such text recurs in almost every trajectory (Appendix F). Furthermore, the ghost persists even at much lower shares. When closed-generated targets make up only half of the training set, continuation still falls by 14.2 points in all six pipelines. At a quarter and even a tenth, one-step exits still rise significantly, as detailed in Q2 below.

Finally, success-rate gains extend well beyond FeedMerge. On tau2-bench, regeneration raises the success rate over the sanitized configuration in five of six bases by up to 6.6 points, while the continuation rate for gemma-4-31B recovers dramatically from 61.3% to 97.7%. On AppWorld, the untrained success rate peaks at only 6.8% for all six bases, leaving almost no room for success-rate gains. Consequently, the pre-registered primary metric there is continuation rate, where the post-training effect remains significant in five of six cells (Table 14).

AuthorsOpenReview ↗

To Reviewer sHqZ (3/3)

Q1. Are the continuations filtered by task success before training? Standard replay pipelines filter by success, which might remove a large portion of the ghost by itself.

A. The paper trains on every continuation, which reflects realistic pipelines that retain unsuccessful trajectories for their training signal (Wang et al., 2024). Such pipelines inherit the full context ghost.

We have added a success filter for Qwen3.5-4B in FeedMerge, keeping only snapshots where all three continuations succeed (Table 6). The filter indeed narrows the gap, but it does so strictly by discarding the records affected by the ghost. Because every success criterion in our environments requires further work after the failure, early-exit targets inherently fail. Consequently, the filter heavily penalizes closed-generated targets. It retains 80% of open targets but only 56% of closed targets on this base, dropping to just 7.6% on gemma-4-31B, and leaving only 8 out of 7,150 on Qwen3-4B.

Even with these massive exclusions, the success-rate gap between the open- and closed-trained models remains 3.6 points, keeping the same direction in 4 of 6 pipelines.

By contrast, regeneration preserves all records. Without applying any filter, regeneration recovers 19.4 of the 26.9 points in success rate, achieving 77.2% compared to 57.8% for closed and 84.7% for open.

Q2. How does the effect scale when the closed-generated records are mixed with clean records, for example at 10 or 50 percent?

A. This is a useful test of practical relevance, which we conducted on Qwen3-4B in FeedMerge. We mixed closed-generated targets to make up 10%, 25%, or 50% of the training set, filling the remainder with neutral targets from the exact same snapshots (Table 6).

At a 50% share, the continuation rate drops by 14.2 points consistently across all six pipelines.

At 25% and 10% shares, the context ghost manifests first in when the model quits. One-step exits rise significantly from the 1.5% baseline, reaching 5.5% with a tenth mix and 8.4% with a quarter mix (Table 6).

These results demonstrate that even a small minority of affected records changes when the trained model gives up. Given that framework-written fields reach almost every trajectory in real corpora, as detailed in W2 above, this phenomenon poses a practical issue rather than a worst-case demonstration.

Q3. Why is πneutral worse than πopen for gemma-4-31B on FeedMerge in Table 4? Does this mean that choosing the generation context matters more than removing the metadata?

A. The difference already exists in the targets. Without any label, gemma-4-31B stops within two steps in 16.8% of its continuations, whereas under the open label it stops in only 4.2%. The trained models directly inherit this, yielding 20.6% against 3.5% in one-step exits.

The open label is metadata too, and on this model its ghost favors persistence. Behavior is fixed the moment the target is generated, which is exactly the paper's central claim. Regeneration intentionally uses the clean context because that is the exact context the deployed model will face. While generating under a favorable label is possible, it constitutes a deliberate context distillation of that label's behavior, a choice practitioners can now make knowingly.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.