acceptodds
Under review as a conference paper at ICLR 2027

TraceTilt: Context-Supervised Model Editing and Steering

Abstract

Context changes what a language model does and knows; we study how to retain these changes after the context is removed. We introduce , a framework for context-supervised steering and model editing. We use model editing to construct steering from the predictions of a Context Teacher, the same model given additional context. The retained behavior can itself acquire information: a question-rewriting operation constructed from four reference contexts transfers to new records. For knowledge editing, four reference contexts with example questions initialize reusable task behavior; each new document then supplies its own editing supervision, without access to future evaluation questions. The resulting task and information updates are merged into ordinary weights. On three behavioral tasks, achieves 92.2% mean target-choice accuracy, versus 70.4% for the strongest evaluated baseline. On unseen questions without the document, the edited model recovers 28.6% of the Context Teacher's improvement over Base on BFCL and 26.9% on FictionalQA.

Primary area:
foundation or frontier models, including LLMs
Keywords:
model editing, activation steering, context distillation

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_eXEzOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper proposes TraceTilt, a framework that combines context distillation, activation steering, and model editing to retain context-induced behaviors and knowledge after the context is removed. The method constructs protected low-rank interventions by matching predictions from a context-conditioned teacher, and explores their application to behavioral steering and document-based knowledge editing. The paper also presents question-rewriting and activation-patching experiments to motivate its design. Experiments show strong performance on three behavioral steering tasks (92.2% mean accuracy), while the knowledge-editing experiments recover 28.6% and 26.9% of the Context Teacher's improvement on BFCL and FictionalQA, respectively.

Critical strengths

The combination of context distillation and model editing is reasonable and potentially insightful. Using contextual predictions as supervision for targeted parameter updates provides a promising direction for retaining information and behavior without requiring the original context at inference time.

Critical weaknesses
  1. The paper lacks the mandatory AI Use Statement. The ICLR 2027 Author Guidelines explicitly require an AI Use Statement section, but this section is missing from the manuscript.

  2. The paper suffers from serious clarity and organization issues, making it difficult to understand the main contributions and how different components relate to each other. In particular, the connections among behavioral steering, question rewriting, mechanistic analysis, and knowledge editing are not clearly established.

  3. The knowledge-editing evaluation is insufficient to support the paper's broader claims. Tables 4 and 5 contain substantial pending results, including the entire LiteraryQA evaluation and complete capability-retention measurements. The completed knowledge-editing evaluation covers only 10 BFCL questions and 50 FictionalQA questions, which is too limited to convincingly demonstrate the effectiveness and generality of the proposed method.

Questions for authors
  1. The discussion of related work appears insufficient. Given that this paper lies at the intersection of context distillation, activation steering, and model editing, could the authors better clarify its contributions relative to closely related works such as Learning by Distilling Context, SAKE, and Steer Like the LLM: Activation Steering that Mimics Prompting?

  2. What motivates the specific design choices of the knowledge-editing method? In particular, why are separate shared task updates and context-specific updates necessary, and what advantages does this design offer over simpler context-distillation approaches or established model-editing methods?

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

This paper employs model editing techniques to construct a guidance mechanism based on predictions made by a context teacher.The proposed TRACETILT method achieves well editing performance across the three evaluation tasks.

Critical strengths
  1. This paper employs model editing techniques to construct a guidance mechanism based on the predictions of a "contextual teacher."

  2. The proposed method demonstrates strong editing performance across three evaluation tasks.

Critical weaknesses

1.The experiments on knowledge editing were conducted using only a small number of samples, and some results are still pending; therefore, it cannot yet be concluded that this method is stable and effective.

2.The paper fails to clarify which specific design element contributes to the observed performance, nor does it verify whether each module is truly indispensable.

3.The evaluation is limited to mathematics, general knowledge, and conversational ability; it does not comprehensively assess whether editing knowledge negatively impacts other capabilities, such as multilingual proficiency, agentic behavior, or tool usage.

Questions for authors

no question

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

The paper proposes TRACETILT, a framework for context-supervised steering and model editing. This tries to retain the behavior even when context changes or is removed. The framework learns model editing for steering from the predictions of a Context Teacher, the same model given additional context. The authors retain a question-rewriting operation from four reference contexts containing sixteen question–answer records, then reuse it on new records without the instruction. They construct knowledge edits from complete contextual answers. They write task behavior and new information into ordinary weights using reference question examples and the target context. The resulting edits recover 28.6% and 26.9% of the Context Teacher’s improvement on BFCL and FictionalQA. On three behavioral tasks, TRACETILT achieves 92.2% mean target-choice accuracy, versus 70.4% for the strongest evaluated baseline (ICV - in context vectors).

Critical strengths

The paper studies an interesting problem of context-supervised steering and model editing. it combines some techniques in some interesting way.

Critical weaknesses

Most of the experimental results are incomplete. Tables 4 and 5 have most of the entries marked as dashes, indicating pending results. The paper as submitted does not seem to be ready for acceptance therefore. For the results presented, there seemed to a drop of 3.9-3 points in accuracy for the tested directions. Only one model is tested, hence it is not clear if the technique generalizes. In the presented results in Table 4, the model results are not clear intuitively. For BFCL, base is 0.0, shared is 0.0, Context is 10.0, but somehow Combined is 20.0, on the other hand for Fictional QA, base is 42.0, shared is 42.0, Context is 56.0 and Combined is the same as Context at 56.0. Is it the case then that shared method is mostly ineffective?

Also, in the current day scenario where context sizes are not really a bottleneck in most cases, I am not fully clear what is the practical relevance of this method of context supervised steering and model editing to preserve model behavior without supplying the context. The authors motivate this problem as follows "A document might say that Nora handed a key to Elias; with the document present, the model can answer who received the key. Once it is removed, that knowledge need not remain available. Our goal is to retain what the context changes, so that its instructions and information remain useful after removal." Why couldn't we simply use RAG or full model context to solve the same problem? On the other hand if the context is indeed removed, so that it cannot be used for RAG or as context, then perhaps it is not intended to retain that knowledge for answering. If it is behavioral information and not contextual details, could ICL or few shot examples not solve this problem?

Questions for authors

Why was only one model chosen for the evaluations? Was there a reason behind the choice of this model? The results are incomplete. In the presented results in Table 4, the model results are not clear intuitively. For BFCL, base is 0.0, shared is 0.0, Context is 10.0, but somehow Combined is 20.0, on the other hand for Fictional QA, base is 42.0, shared is 42.0, Context is 56.0 and Combined is the same as Context at 56.0. Is it the case then that shared method is mostly ineffective?

What is the practical relevance of this problem (refer Weaknesses).

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.