acceptodds
Under review as a conference paper at ICLR 2027

From Target Selection to Condition Reuse: Auditing Residual Interventions in Tool-Use Models

Abstract

Activation interventions can make a language model produce a target answer, but can the model use the intended condition in other computations? We study this question in Qwen3.5-27B on tool-use tasks where time or permissions determine eligibility and selection. Across 32 fresh worlds, two case-specific residual-writing variants raise familiar target-selection correctness from 35/128 to 119/128 and from 33/128 to 114/128. Yet these gains disappear when familiar facts and identifiers are recombined within the same selection task, while selection of previously rewarded identifiers increases. The gains persist across output formats but yield little improvement in independent judgments about whether a candidate is eligible or should be selected. Internal edits can nevertheless change how a source fact affects a judgment: in a prespecified secondary comparison, an early-layer source patch changes 50/64 judgments when that source is used, with one adverse change among 64 cases where it is not. This selective effect does not ensure correct complete actions. Together, these results separate target-selection control, condition-sensitive judgments, and complete-action correctness. Evaluating how an edited state is used, rather than only its trained answer, exposes these differences.

Primary area:
interpretability and explainable AI
Keywords:
large language models, residual stream interventions, activation steering, mechanistic interpretability, tool use, counterfactual evaluation

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

The paper investigates whether activation interventions that produce a desired selection also support reuse of the condition governing that selection. In procedural time- and permission-based tasks, the authors train case-specific residual vectors in Qwen3.5-27B using target-identifier supervision. These interventions substantially improve selection on training-exposed menus, but the improvement disappears when familiar facts and identifiers are recombined. The interventions also provide little benefit to independently elicited eligibility and winner judgments, despite transferring across output wrappers. Additional experiments substitute retained history and patch source facts. An early source patch changes eligibility judgments primarily when the edited source is relevant, but does not reliably produce correct complete actions. The paper distinguishes successful target control, condition-sensitive judgments, and complete-action correctness.

Critical strengths
  • The factual-recombination experiment provides a convincing test beyond familiar-answer accuracy. It preserves the selection task while changing which identifier the current facts support. The disappearance of selection gains, together with increased repetition of the old target, identifies a specific failure pattern rather than merely showing poor transfer to a different task. The perfect-training-fit analysis further shows that this pattern is not confined to unsuccessful optimization.

  • The independent continuations and capability controls make the behavioral comparisons substantially more informative. Selection outputs are not passed into the judgment branches, and the wrapper experiments help separate serialization changes from changes in the requested computation. I also appreciate the explicit separation of fresh confirmation, reused worlds, and exploratory analyses, as well as the reporting of collateral errors and the unsuccessful primary source-patch result.

Critical weaknesses
  • The evidence remains specific to a restrictive intervention design. The learned writer operates at block 60 of 64, changes only a small portion of retained state, and receives supervision solely on target identifiers. This setup permits a shortcut that reproduces rewarded identifiers without updating the condition used elsewhere. The results establish that this shortcut occurs, but do not show how sensitive the failure is to intervention depth, token support, or supervision. The source-patching experiment does not resolve this because it changes all three: it uses natural donor activations, covers many tokens, and acts earlier. The paper acknowledges these confounds, but a more closely matched intervention comparison would substantially strengthen the contribution.

  • Weak native performance limits the interpretation of some reuse and action results. Natural B correctly answers only about one-third of target-true winner-verification questions, and its selection and complete-action accuracy also remain modest. Consequently, the absence of intervention gains can reflect both intervention limitations and limitations of the consumer itself. The 27 illustrative permission records demonstrate that the separation can occur when controls succeed, but they were selected partly for successful writer selection and failed writer judgments, so they cannot establish its prevalence among capable cases. The same-task crossing experiment is stronger here; the broader verification and complete-action conclusions need more systematic capability-conditioned evidence.

Questions for authors
  • Does the selection-versus-reuse gap persist when the same target-supervised writer is trained at earlier sites, with comparable intervention magnitude? This would help distinguish a limitation of the supervision from a limitation of the late write location.

  • Can you report results on a subset qualified solely by successful Native A and Natural B behavior, before considering writer outcomes? In particular, how often do crossed-selection and verification failures occur when the relevant unedited consumers demonstrably perform both conditions correctly?

  • Does the successful block-15 source patch retain its selective judgment effect across the full source-intervention cohort, beyond the prespecified 64-case subset? Extending this comparison would clarify the robustness of the paper’s positive condition-use result.

Reviewer_CDs2OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper studies whether activation intervention not only produces target answer but also learns the underlying intended conditions when used in other computations. The authors developed experiments using Qwen3.5-27B where a time or a permission condition determines which candidate is the cheapest eligible option.

The experiments show that on 32 fresh worlds, familiar-menu target selection is raised. However, when facts and IDs are recombined, the gain disappears. It is also found an early (block-15) patch changes 50/64 "used-source" judgments with little effect on unused sources. The paper then concludes that interventions should be evaluated by how edited states are reused, not only by their output.

Critical strengths

The paper delievers an important message that training a model to output something is not equal to training it to learn the reasons behand the output.

The findings of the paper are important for works on knowledge steering and editing.

The paper proposes a sensible audit protocol.

The experiments are carefully designed and controlled.

Critical weaknesses

The main finding seems to implied by the constructions. The writing is applied at one token which leaves the prior ones untouched. Therefore it can also affect the output through a small set of states. Therefore it is unsurprising the independent eligibility queries which probably read from earlier representations, are unchanged.

The models perform relatively poorly at the tasks. It is unclear whether to attribute the failures to the intervention of the basic capabilities of the model. A stronger base model would be suggested.

Lack of comparisons with standard intervention methods. The authors are encouraged to survey and compare with other intervention methods.

The experiments focus on rather simple and stylish settings. Generalizations to more realistic settings will be appreciated.

The paper is generally hard to follow, with some general terms being redefined and used heavily...it was hard to understand without consistently going back and forth with the paper.

Questions for authors

See weaknesses.

Reviewer_tz5zOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The authors study whether an internal intervention can make an LLM change its answer and then use the new condition in other tasks. On tool-use tasks with Qwen3.5-27B, they find that learned residual vectors change the answers on training menus, but do not generalize to recombined facts or change the model’s independent judgments.

Critical strengths
  • The research topic is valuable. It can help us understand how LLMs change their answers under internal interventions.
  • The experimental design and reporting are very rigorous.
Critical weaknesses
  1. The authors have not demonstrated that the injection is effective beyond the training menus, leaving the core conclusion without a necessary premise. The injection needs to be shown to generalize before any further claims can be made. More specifically, the writer's gains hold on Anchor. When only the order is shuffled, the gains are largely preserved. However, once the IDs are changed or the mappings between facts and IDs are rearranged, the gains disappear. On Cross, the writer even performs slightly worse than Native and more often selects the old training IDs. The logical order should be to first demonstrate that the injection can generalize, and then examine whether it helps the model understand the conditions. Based on the current evidence, the paper can only conclude that "this particular injection does not help with understanding," rather than support a general claim. In fact, the simplest injection method, directly providing condition B in the prompt (equivalent to layer 0), performs substantially better than the writer on Cross. Although the authors attempt to change the model's decisions through source patching, this is based on natural counterfactuals rather than learned injection.

  2. The design is insufficiently justified, and the conclusion is almost expected. The paper does not explain why the injection is performed at layer 60 or why the action ID token is chosen. It is hardly surprising that a vector trained only with target ID supervision and injected into the final few layers fails to learn the conditions. Moreover, the writer is a black-box optimized vector, and the paper provides very limited analysis of what it actually changes. The authors only report which cache components change, without analyzing how the injected information propagates to later tokens through attention and recurrent states. This makes evaluation on multiple models or models with different architectures even more necessary, rather than relying solely on Qwen3.5, which uses a hybrid attention architecture.

  3. Limited novelty. Evaluation methods such as shuffling facts and asking forked queries are already common. The two supervision methods, Path-cover and Random-fact, produce almost identical results. This is largely redundant and adds little to the argument.

  4. Narrow scope and very difficult to follow. The practical relevance and generalizability of the conclusions are questionable. The paper introduces too many custom terms, and the main text is filled with numerical results, making it difficult to follow.

Questions for authors

Please refer to Weakness

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.