acceptodds
Under review as a conference paper at ICLR 2027

SteerSim: De-identifying Developer Behavior in Agent Traces via User Simulation

Abstract

Developer–agent interaction logs are valuable training data, but releasing them poses privacy risks beyond personally identifiable information. We observe that how long developers let an agent run, how closely they scrutinize proposed actions, and how they phrase corrections can form a distinctive behavioral fingerprint that conventional anonymization leaves intact. We introduce SteerSim, an agentic user simulator that replaces the developer in a live coding session to enable privacy-preserving trace synthesis. Operating within the coding agent's environment, SteerSim inspects the repository and takes actions to steer the agent toward task completion. It is conditioned on a persona extracted from real developer sessions, which can be transformed to obscure the original developer's behavioral patterns. We further propose SteerRL, a training algorithm that combines supervised fine-tuning with agentic reinforcement learning using rubric-based rewards. We evaluate both on real sessions from SWE-Chat along three dimensions: human-likeness, re-identification risk, and downstream utility. With Opus 5 as its backbone, SteerSim produces interactions that are nearly indistinguishable from human interactions, as measured by the Turing score. Transforming personas reduces re-identification risk to near chance while preserving downstream utility, achieving a favorable privacy–utility tradeoff. Finally, applying SteerRL to Qwen3.5-4B yields a simulator with performance comparable to Opus 5, offering a practical and low-cost alternative.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Privacy-preserving data synthesis, user simulation, reinforcement learning.

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_UbNEOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies privacy risks in developer-agent interaction logs. The authors argue that conventional anonymization does not remove behavioral fingerprints, and propose SteerSim, an agentic user simulator that replaces the original human developer in a live coding session. The simulator is conditioned on a persona extracted from real developer traces, and the persona can be transformed to reduce behavioral re-identification. The paper further proposes SteerRL, a post-training method for a smaller open-weight simulator using SFT and agentic RL with rubric-based rewards.

Critical strengths
  1. The paper identifies an important and timely privacy problem: developer-agent traces can leak identity through behavioral patterns even after standard anonymization.
  2. The proposed simulator is more ambitious than next-utterance user simulation, since it acts inside a live coding environment, can inspect the repository, and decides when to intervene or stay silent.
Critical weaknesses
  1. The central assumption behind the paper is not fully validated: replacing a real human developer with a simulator may remove exactly the interaction signals that make the original logs valuable. The downstream utility experiment only shows comparable performance in limited settings, and the gains appear mainly on FeatureBench, while SWE-bench Verified and SWE-Skills-Bench do not show clear improvement.

  2. The persona is extracted by an LLM from a single session and evaluated largely by LLM-based rubric judgments. The paper does not provide enough evidence that the induced persona accurately represents the real developer.

  3. The comparability between the simulator and human developers remains unclear. Human-likeness and protocol validity are useful metrics, but they do not directly test whether the simulator has comparable software-engineering judgment, debugging ability, or intervention quality. A simulator may look human-like while still steering the agent in ways that are less informative for training future agents.

Questions for authors

How do the authors verify that an induced persona is an accurate representation of the original developer, beyond LLM rubric scores? Were any human developers or independent annotators asked to validate the extracted personas?

Since the goal is to release useful training traces, can the authors evaluate whether models trained on synthetic traces learn the same intervention patterns as models trained on real human traces, rather than only reporting benchmark pass@1?

3 comments

We thank the reviewer for the thoughtful feedback and for recognizing both the importance of behavioral privacy and the value of simulation inside a live coding environment. We address the preservation of useful interaction signals with additional downstream experiments, and clarify what our persona and behavioral evaluations establish.

W1: Replacing the developer may remove the interaction signals that make the original sessions useful; the downstream gains are limited.

Thank you for identifying the central utility question. Our aim is to reduce behavioral identification while retaining the usefulness of the sessions for training coding agents. Figure 5 tests this through agent-side SFT and skill distillation: synthetic and real sessions provide comparable downstream performance in the evaluated settings, and both improve FeatureBench. We do not interpret the absence of clear gains on the other benchmarks as evidence that the sessions improve every kind of software-engineering task.

We added two experiments to examine the value of the generated interactions more directly. First, we regenerated 350 persona-swap sessions with Qwen3.5-4B as the coding agent and SteerSim (Opus-5) as the user simulator, then trained Qwen3.5-4B on its own coding trajectories. Second, following the feedback-annotation approach in The verification horizon [1], we labeled the simulated-user turns and used their feedback in Span-KTO training. All results below use Qwen3.5-4B with opencode on the first 100 SWE-bench Verified tasks, three evaluation runs, and the same environment.

Tables R2–R3: Downstream utility of self-generated sessions and simulated user feedback.[^hardware-UbNE]

Training data and objectivepass@1pass@3
None (untrained)0.3930.56
Real SWE-Chat sessions, SFT0.400—
Self-generated SteerSim (Opus-5) sessions, SFT0.4100.58
Self-generated SteerSim (Opus-5) sessions, Span-KTO0.4100.58
Self-generated SteerSim (Opus-5) sessions, class-balanced Span-KTO0.4170.57
Real sessions and real user feedback, Span-KTO0.402—

Self-generated trajectory SFT improves mean pass@1 from 0.393 to 0.410 without stronger coding-agent outputs as targets. With standard Span-KTO, simulated feedback reaches 0.410, compared with 0.402 using real feedback. Class balancing further increases pass@1 on simulated feedback to 0.417. This tests a use of the simulated interventions themselves: their feedback is converted into supervision for the coding agent. It extends the utility evaluation beyond imitation and skill distillation, although it does not show that every useful human interaction signal is preserved.

The 350 sessions provide 1,130 judged verdicts without requiring live human participation during rollout or feedback annotation. The pipeline still starts from real session tasks and personas. We also observe a substantial feedback-distribution difference: 87% of simulated feedback is negative, versus 32% for real feedback. Class-balanced Span-KTO yields the highest pass@1 in this comparison, while SFT and standard Span-KTO retain the highest pass@3. We will report this distinction rather than treating simulated and real feedback as interchangeable distributions.

W2/Q1: Does a persona induced from one session accurately represent the developer? Was it validated by humans?

Thank you for asking us to clarify the intended meaning of a persona. Our personas are session-specific behavioral summaries, not estimates of a developer's complete or stable personality. Section 4.2 explicitly attaches personas to sessions: the same developer can have different personas across sessions. The axes are motivated by prior studies of developer interaction styles, including [2], and the induction prompt summarizes observable steering habits from the session. Prior work motivates this representation but does not independently validate each extracted profile.

The paper provides several checks beyond the persona-alignment rubric. In the re-identification evaluation, the attacker's profile comes from a different session by the same developer; its 0.309 accuracy on real user turns, against 0.100 chance (Table 8), shows that the profiles capture some cross-session identifying information. Appendix I.5 further removes task, repository, product, and technology specifics from the conditioning personas. On the same 100-session subset, this changes Turing score from 4.295 to 4.240, persona alignment from 0.862 to 0.870, and profile re-identification from 0.240 to 0.220 (Table 14). These results support the role of transferable behavioral descriptions rather than dependence solely on task-specific details. Appendix G also traces how a generated session reflects the original developer's corrections and plan-gating behavior.

We have not asked the original developers or independent human annotators to validate the extracted personas. We will state this explicitly and distinguish evidence that a profile captures useful behavioral information from validation of every inferred trait.

W3: Human-likeness does not establish comparable engineering judgment, debugging ability, or intervention quality.

Thank you for making this distinction. We agree that a human-likeness score is not a measure of engineering competence. SteerSim is designed to model developer steering, including when to inspect, interrupt, question, or approve the agent, rather than to claim expert-level debugging from a Turing score. Its live access to the repository makes such decisions possible, but does not by itself establish that their quality matches a human developer's.

We therefore evaluate behavior and downstream usefulness separately. Table 11 directly measures protocol actions and reveals remaining differences: real developers average 0.35 interrupts and 0.44 plan approvals/rejections per session, while SteerRL averages 0.17 and 0.07. SteerRL is also more talkative (7.47 versus 5.44 messages). The new feedback-training experiment above adds evidence that interactions generated by SteerSim (Opus-5) contain usable supervision, evaluated through downstream tests. We will distinguish this demonstrated utility from the stronger, untested claim of matching human engineering judgment.

Q2: Do models trained on synthetic traces learn the same intervention patterns as models trained on real traces?

Thank you for this question. There are two different learned roles in the paper. SteerRL learns to produce user interventions; the downstream coding agent learns to respond to them. In downstream SFT, user messages remain in the context, but the loss covers only the coding agent's assistant turns; user turns and tool results are masked (Appendix H). We also understand the question as asking whether agents trained on synthetic and real traces respond similarly to human corrections and feedback. We have not directly compared their responses to matched interventions; user turns still condition the agent's responses even though they are masked from the loss.

For simulator behavior, Tables 2 and 11 report persona alignment and directly counted interaction patterns. Relative to simulator SFT, SteerRL increases own-persona alignment from 0.790 to 0.867, validity from 0.770 to 0.921, and interrupts from 0.04 to 0.17 per session, although the human distribution is not fully matched. For downstream agents, the new Span-KTO experiment uses feedback extracted from SteerSim (Opus-5) interactions to train the coding agent and evaluates its task performance. This provides related evidence that simulated feedback is useful for learning, but does not establish that agents trained on synthetic and real traces respond equivalently to human interventions. We will make this distinction explicit.

[1] The verification horizon: No silver bullet for coding agent rewards. arXiv 2026.

[2] No two developers think alike: How problem-solving styles and experience shape needs in conversational interaction with Copilot. ICSME 2026.

Reviewer_L6zxOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

Coding-agent logs reveal a developer's identity through how they steer the agent, and that standard anonymization doesn't remove this. To resolve this issue, the paper proposes SteerSim, a user simulator that replays each session in the live environment under a different developer's persona, and SteerRL, which trains a 4B model as the simulator with SFT followed by RL on persona-fit and human-likeness rewards. The paper shows that applying the persona brings re-identification close to chance with only a small loss in human-likeness, and the synthetic sessions help a 4B coding agent about as much as the real ones.

Critical strengths

The training and evaluation of the user simulation are sound and comprehensive. 3 metrics, Validity ( protocol-validity rate of the simulated developer), Turing, and Persona are proposed to evaluate the 4B model's closeness to a human developer. And the model has went through SFT and RL training with carefully designed rewards for better simulating the user.

Critical weaknesses
  1. The result seems to not include the performance of such user simulator driven sessions, besides the human-likeness. All three metrics, validity, turing, and persona, do not evaluate whether the simulator driven sessions actually achieves the same outcome as the original human driven sessions. Similarly, the RL reward design does not reward performance besides the human-likeness. Only mimicing the style of a human written prompts, without consideration of the performance does not seem to be a valid human developer simulator design.

  2. The downstream usefulness evaluation of the simulated sessions seems to be limited to only 4B coding agent. The simulated session contain not only the simulated user prompt, but also a strong coding agent's output, (GPT 5.6 luna). Using this session trajectories to improve a 4B coding agent, the main effect might be from learning from this stronger coding agent, instead of from the user prompts. It would be more meaningful if the downstream usefulness can be demonstrated on the similar coding agent that produces the session trajectories (could be both strong agents or both weak agents).

Questions for authors
  1. Do simulator-driven sessions reach the same outcomes as the human-driven originals, in terms of produced code, artifact, or passing tests of the code?
  2. Is the downstream gain from the simulated steering or from the stronger coding agent?
2 comments

We thank the reviewer for the thoughtful feedback and for recognizing the comprehensive simulation evaluation and reward design. We clarify the distinction between simulating open-ended developer sessions and solving fixed benchmark tasks, and add the requested same-model downstream experiment.

W1/Q1: Do simulated sessions reach the same outcomes as the human originals, and why does the simulator's reward not include task performance?

Thank you for this question. SWE-Chat records open-ended developer interactions rather than tasks with a fixed success specification and reference test suite. A developer may revise the objective, reject a plan, revert a change, or end the session before completion. After the first intervention, a simulated rollout can also enter a different state from the original session. Consequently, reproducing the original final patch is not a general correctness criterion, and the dataset does not supply a uniform task-success verifier analogous to SWE-bench's tests.

The simulator's role is to reproduce how a developer steers under a given persona, including decisions that need not maximize immediate task completion. Optimizing only successful completion could instead encourage it to become a second coding agent. Our rewards therefore target human-likeness and persona-consistent steering in the live environment. This goes beyond mimicking prompt wording: the simulator can inspect the repository, observe the agent, interrupt execution, and approve or reject plans according to the current state (Section 4.3).

We agree that validity, Turing score, and persona alignment do not establish equal engineering competence or identical task outcomes. We assess the practical usefulness of the resulting sessions separately by training a coding agent and testing it against executable benchmark tests (Section 6.1 and Appendix H). Figure 5 evaluates both agent-side SFT and skill distillation; the new experiments below additionally examine self-generated trajectories and learning from simulated feedback. We will clarify this separation and the absence of a session-level outcome-equivalence claim.

W2/Q2: Are downstream gains due to simulated steering or to distillation from the stronger coding agent? Can the same model generate and learn from the trajectories?

Thank you for suggesting this control. We regenerated the same 350 persona-swap sessions using Qwen3.5-4B itself as the coding agent, with SteerSim (Opus-5) as the simulated user. We then fine-tuned Qwen3.5-4B on its own agent-side trajectories using the paper's unchanged recipe. Evaluation uses Qwen3.5-4B with opencode on the first 100 SWE-bench Verified tasks, with three runs and the same environment for every row.

Table R2: Same-model trajectory generation and downstream SFT.[^hardware-L6zx]

Training datapass@1pass@3
None (untrained)0.3930.56
Real SWE-Chat sessions, SFT0.400—
SteerSim (Opus-5), self-generated trajectories, SFT0.4100.58

Self-generated sessions improve the observed pass@1 by 1.7 percentage points and pass@3 by 2 points over the untrained model, with pass@1 comparable to training on real sessions. The improvement therefore does not require stronger-model coding outputs as SFT targets. A frontier model still generates the simulated user interventions, and this comparison does not isolate their contribution against a matched no-steering rollout baseline.

To test a more direct use of those interventions, we also annotated simulated-user turns using the judge prompt from The verification horizon [1] and used their feedback labels for Span-KTO training on the same self-generated sessions.

Table R3: Learning from simulated user feedback on SWE-bench Verified under the same evaluation setup.

Training data and objectivepass@1pass@3
None (untrained)0.3930.56
Self-generated SteerSim (Opus-5) sessions, SFT0.4100.58
Self-generated SteerSim (Opus-5) sessions, Span-KTO0.4100.58
Self-generated SteerSim (Opus-5) sessions, class-balanced Span-KTO0.4170.57
Real sessions and real user feedback, Span-KTO0.402—

The 350 sessions yield 1,130 judged user verdicts. With standard Span-KTO, simulated feedback achieves pass@1 of 0.410, compared with 0.402 for real feedback, and matches SFT. Class balancing further increases pass@1 on simulated feedback to 0.417. Its pass@3 is slightly lower than SFT's, so we do not claim uniform superiority across metrics. This experiment shows that the simulated interactions can supply feedback labels for training the coding agent, in addition to providing trajectories for imitation. The feedback is predominantly negative in this setting (87%, versus 32% for real users), making class weighting an important design consideration. We will add these results and distinguish the role of the user simulator from the model generating the coding responses.

[1] The verification horizon: No silver bullet for coding agent rewards. arXiv 2026.

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

The authors introduce the risk of agentic logs carrying behavioral fingerprints that can be identified by LLM attackers at non-trivial rates and generic PII anonymisation does not fix this. The paper introduces SteerSim as a defense where an agentic user simulator conditioned on a 10-axis persona extracted from real sessions operates in the live repository. It regenerates the entire session to protect priovacy. SteerRL trains Qwen3.5-4B to reduce re-identification to near chacne and increase human likeness.

Critical strengths
  • The simulator design is concrete. It operates inside the live environment, and under state-conditioned actions only.
  • I really like the privacy risk studied here, which is quite novel and well motivated. Behavioral re-identification is an important risk and the results show it can be significant even after redactio, subsitution, etc.
  • The expeirments are clear and rigorous. SFT vs RL experiments are useful
Critical weaknesses
  • My main concern is that the evaluation and training pipelines have the same judge LLM (gpt luna). The same model generates the training reward and the evaluation which would most probably lead to reward hacking.
  • The privacy eval feels circular right now: attacker is an LLM matching against personas written by the same persona inducer the simulator was conditioned on.
  • There is no human evaluaiton reported on any results or evaluations.
  • There is no performance gain on SWE-Bench Verified/SWE-Skills-Bench from real or synthetic data.
  • I'm confused and would like clarification: isn't SFT on the released data partyl distillation from a frontier agent and doesnt reflect the real data anymore
  • There should be discussion around formal privacy guarantees and if there is a possibility to include them in the defense.
Questions for authors

See above

2 comments

We thank the reviewer for the thoughtful and constructive feedback, and for recognizing the importance of behavioral re-identification, the concrete simulator design, and the value of the SFT and RL experiments. We clarify the evaluation controls below and provide additional downstream experiments.

W1: Using the same LLM for training rewards and evaluation may lead to reward hacking.

Thank you for raising this concern. We have already included an evaluation with a second judge from a different model family in the paper (Appendix I.3, Table 12). Although GPT-5.6 Luna supplies the training rewards and primary evaluation, we independently re-evaluate human-likeness using Qwen3.5-397B-A17B, keeping the pairwise prompts, real reference sessions, and side randomization fixed.

Table R1: Human-likeness under two judge models, reproduced from Table 12.

SimulatorLuna: ownLuna: transformedQwen: ownQwen: transformed
Opus 53.8463.6804.0394.077
Qwen3.5-4B base2.9762.8862.7832.654
Base + SFT3.8013.5863.9503.835
Base + RL3.0682.9302.8492.713
SteerRL3.8973.6964.5544.505

The judges' per-session scores correlate at 0.70–0.82 across backends. Both support the main conclusions: SFT substantially improves human-likeness, RL from the raw base provides little improvement, and SteerRL matches or exceeds the frontier backends on this metric. The second judge is more favorable to SteerRL; as discussed in Appendix I.3, scores above the real-vs-real anchor of 4.0 indicate judge preference, not that a simulator is “more human than human.” These results address dependence on a particular reward model, although agreement between LLM judges cannot rule out shared biases. Protocol validity, the behavioral counts in Table 11, and downstream benchmark tests additionally provide evidence that does not depend on the human-likeness judge. We will make these controls more prominent in the main text.

W2: The privacy evaluation appears circular because the attacker matches against personas produced by the same inducer.

Thank you for asking us to clarify the information available to the attacker. The attacker does not receive the same session persona that conditions the simulator. Under the own-persona setting, the simulator receives a persona extracted from the target session, whereas the attacker's profile of the true author comes from a different session by that developer (Section 3 and Appendix A.4). Under persona transformation, the simulator instead receives another developer's persona. Naming that donor does not count as identifying the original author.

The same induction procedure does provide a shared representation, so we also evaluate an attacker that uses no induced personas. Appendix I.4, Table 13 matches generated user turns against candidates' raw user-side transcripts using embedding similarity. For SteerRL, re-identification decreases from 0.251 to 0.123 under the profile attacker and from 0.184 to 0.161 under the raw-transcript attacker when moving from own to transformed personas. The latter remains above the 0.100 chance level. Thus, the privacy result is not solely an artifact of matching induced profiles, but its strength depends on the adversary. We will explicitly distinguish the near-chance result under the profile threat model from the residual leakage under raw-transcript matching.

W3: No human evaluation is reported.

Thank you for highlighting this distinction. Our Turing evaluation follows the approach of Learning user simulators with Turing rewards [1], which evaluates user simulation with both LLM and human judges in conversational settings. This provides methodological precedent, but we do not claim that its human evaluation validates our coding-session personas or trajectories. We have not conducted a new human annotation study in this work.

Our evidence combines a second judge (Table 12), directly counted interaction behaviors (Table 11), a qualitative comparison against the original developer's behavior (Appendix G), and downstream execution-based evaluation (Figure 5). These assess different aspects of simulation beyond a single rubric score. We will clarify that human-likeness is an LLM-judged measure and distinguish it from direct human validation of developer fidelity.

W4: Neither real nor synthetic data improves performance on SWE-bench Verified or SWE-Skills-Bench.

Thank you for pointing this out. Our utility question is whether transformed sessions preserve the usefulness of the real sessions they replace. Figure 5 compares real and synthetic data under the same downstream uses, agent-side SFT and skill distillation. Both improve FeatureBench, while neither gives a clear improvement on the other two benchmarks. This is consistent with SWE-Chat's emphasis on feature-development work, although it does not establish a causal explanation for the difference between benchmarks. We will clarify that the result concerns comparable downstream utility in the evaluated settings, rather than improvement on every benchmark.

We additionally evaluated training on sessions generated by the downstream coding model itself. With Qwen3.5-4B as both the trajectory-generating coding agent and the model subsequently fine-tuned, SteerSim (Opus-5) data raises mean SWE-bench Verified pass@1 from 0.393 to 0.410, compared with 0.400 for SFT on real sessions. With standard Span-KTO, simulated feedback reaches 0.410, compared with 0.402 for real sessions and real user feedback; class balancing further increases the simulated-feedback result to 0.417. These new comparisons use the first 100 tasks, three evaluation runs, and the same environment across all rows. They extend the evidence to self-generated trajectories and feedback-based training; we will report them separately from the original Figure 5 experiments.

W5: Is downstream SFT partly distillation from a frontier coding agent, rather than evidence about the simulated users?

Thank you for this important question. The original synthetic trajectories contain frontier coding-agent outputs, so we added a control that removes that source of stronger-model code supervision. We regenerated the same 350 persona-swap sessions with Qwen3.5-4B as the coding agent, paired with SteerSim (Opus-5) as the user simulator, and then fine-tuned Qwen3.5-4B on its own agent-side trajectories using the unchanged SFT recipe.

Table R2: Downstream training with self-generated coding-agent trajectories. SWE-bench Verified, first 100 tasks; three evaluation runs.1

Training datapass@1pass@3
None (untrained)0.3930.56
Real SWE-Chat sessions, SFT0.400—
SteerSim (Opus-5), self-generated trajectories, SFT0.4100.58

The observed improvement persists when the coding responses used for training come from the same model being trained. It therefore does not require distillation of a stronger coding agent's responses. SteerSim (Opus-5) still supplies the simulated steering, so this experiment does not remove all frontier-model assistance or isolate steering from every other aspect of trajectory generation. It directly addresses the narrower concern about the source of the coding-agent training targets. As specified in Appendix H, user turns remain in the context but are masked from the SFT loss; the downstream agent learns how to respond within the interaction.

W6: Can the defense provide formal privacy guarantees?

Thank you for suggesting this discussion. Our current claim is empirical mitigation of behavioral re-identification under the specified adversaries, not a formal privacy guarantee. In particular, the raw-transcript result in Table 13 demonstrates residual identification above chance, and the generator deliberately retains the opening request and repository state to preserve task utility.

A formal extension would require specifying which information is protected and which inputs remain public. For example, a guarantee for a learned persona distribution would not by itself protect identifying information retained in the task, repository, or a directly extracted session persona. A useful direction is to privatize the persona-learning or sampling mechanism under a clearly defined neighboring-dataset relation, and analyze the remaining inputs and generated trace as part of the complete release mechanism. We will add this distinction and explicitly state that persona swapping alone does not establish differential privacy.

[1] Learning user simulators with Turing rewards. EMNLP 2026.

Footnotes

  1. Rebuttal experiments were re-run on different hardware under heavier serving load (approximately 38% of tasks reached the 1,800 s agent wall-clock limit), so the untrained baseline is 0.393 here versus 0.44 in Figure 5. All rows in the rebuttal comparisons use the same hardware and resource settings. ↩

Comments on the paper

1 comment

We thank the reviewers for the constructive feedback. Below, we summarize the additional experiments and clarifications in our responses.

Additional experiments

  1. We evaluated using the same coding model to generate trajectories and train on them. SteerSim sessions improve downstream performance without stronger coding-agent responses as training targets (Table R2 in the responses to Reviewers rxWS, L6zx, and UbNE).
  2. We evaluated simulated user feedback as supervision through Span-KTO, finding performance comparable to training with real feedback. Class-balanced weighting further improves pass@1 on simulated feedback (Table R3 in the responses to Reviewers L6zx and UbNE).

Clarifications

Reviewer rxWS

  1. We highlighted the second-judge evaluation and judge-free behavioral statistics, and clarified the evidence supporting the human-likeness evaluation (Tables 11–12).
  2. We clarified that the attacker's profile comes from a different session and highlighted the raw-transcript adversary, which does not use induced personas (Tables 8 and 13).
  3. We explained the scope of the privacy claim and distinguished empirical re-identification mitigation from formal privacy guarantees.

Reviewer L6zx

  1. We explained why open-ended SWE-Chat sessions do not provide a uniform task-success criterion, and distinguished persona-consistent steering from maximizing task completion.
  2. We clarified the separate roles of the coding agent and SteerSim (Opus-5) in the new same-model experiment.

Reviewer UbNE

  1. We clarified that personas describe session-specific behavior and highlighted the abstracted-persona control as evidence for transferable steering habits (Table 14).
  2. We distinguished learning to generate interventions from learning to respond to them, and explained how the feedback-training experiment provides related utility evidence without directly establishing equivalent responses to human corrections.
  3. We clarified the distinction between human-likeness, observed steering behavior, and engineering judgment.

We hope this summary facilitates the reviewers' evaluation and discussion. We welcome further questions and would appreciate reconsideration of the scores if our responses have satisfactorily addressed the concerns. Thank you again for your time and consideration.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.