acceptodds
Under review as a conference paper at ICLR 2027

Task identification features causally drive execution features in in-context learning

Abstract

Language models learn patterns from demonstrations in context, but how information at the demonstrations comes to drive the output has not been established. An identification feature for a pattern forms where the pattern is demonstrated (we call these “label tokens”) and drives an execution feature at subsequent locations where the model must continue the pattern (“cue tokens”). We establish two results: 1) an identification and an execution feature together mediate this effect on behavior, and 2) a single affine map shared across patterns relates each identification feature to its execution feature. In simple patterns generated by input-output tasks in GPT-J and Qwen2.5-7B, where the execution feature is the task's function vector, steering the identification feature raises cue-token alignment with the execution feature more than with a generic one, ablating it lowers that alignment, and ablating the execution feature removes the accuracy gained from identification steering. A single affine map fitted on training tasks predicts held-out tasks' execution features from their identification features ( of 0.60 and 0.53, versus below zero for a mean predictor). The same pathway appears in a naturalistic setting, in-context learning of coding conventions, with no explicit labels, features at evidence and cue tokens steer and ablate generated style, and identification steering moves the cue-token state toward the execution feature, although the linear map is weaker.

Primary area:
interpretability and explainable AI
Keywords:
in-context learning, mechanistic interpretability, function vectors, task vectors, activation steering, causal interventions, linear representations, large language models, few-shot 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 2 reviews, above 27% of the others.
Reviewer_rQHUOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper argues that an identification feature at demonstration label tokens causally drives an execution feature, the FV, at the cue token in GPT-J and Qwen2.5-7B, and that a shared ridge map predicts held-out FVs from identification features (R2R^2 of 0.60 and 0.53). A coding-convention extension repeats the interventions. The paper doesn't point to code.

Critical strengths
  • The controls are careful. Every intervention has an other-task control, all fitting uses training tasks only, and Appendix D reports results that cut against the story (between-task alignment–accuracy correlations of −0.36-0.36 and −0.24-0.24).

  • The 53-family paired-document coding suite tests the account outside Q:/A: prompts, with steering in both directions (Table 5).

Critical weaknesses
  • The central result is largely known. Kharlapenko et al. (2025) showed that task-detection features on demonstration output tokens causally drive task-execution features at the token before completion, on similar tasks, yet the paper cites them only in passing. The shared map, the remaining novelty, isn't specific to identification features. Raw label-token states predict held-out FVs as well as uAu_A at nearly every layer from 6 to 27, and at roughly 0.35–0.4 already at layer 0 (Figure 16, right), while a scaled partial isometry nearly matches ridge (0.588 vs. 0.641 in GPT-J). In code, dfidd^{\text{id}}_f is the difference at tokens that literally differ, ablated from the embedding output onward (Appendix E), which is closer to editing the input than to isolating a feature.

  • The mediation test isn't task-specific. Without any steering, zero-ablating vAv_A at the cue drops real six-shot accuracy to 1.3% and 5.6%, and zeroing only the shared FV direction gives 13.8% and 34.1% (Table 2), so removing "essentially all" of the steering gain shows the cue needs this subspace for any answer, not that the gain flows through it. The other-task control already removes 39% and 51%. The specific test, mean-ablating the task-unique component, which leaves 57.9% on real Qwen prompts, is never combined with steering.

  • Critical and highly overlapping works from the literature aren't cited/discussed (e.g., [1]). The work [1] should be included in the experiments as well as it improves over (and usually outperforms) FVs.

  • The models considered in the empirical analysis are not sufficient (only two models given), largely outdated (Qwen2.5 used, e.g., why not Qwen3?), and don't target a variety of model scales (only 6B and 7B).


[1] Saglam et al., "Learning Task Representations from In-Context Learning," ACL Findings 2025.

Questions for authors
  1. Under label steering, how much of the gain does mean ablation of the task-unique FV component remove, relative to its other-task control?
  2. How well do non-label representations (query-token states, output-word embeddings) predict held-out vAv_A under the same protocol, and does identification ablation still work when restricted to the band BB rather than layers 0–27?
  3. Why didn't the author consider more and recent models?
Reviewer_jgRmOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2(+1)
Presentation
3
Soundness
2
Read review
Paper summary

This paper investigates how information within in-context demonstrations influences a language model’s subsequent task execution. For simple input–output tasks, identification features in-context are constructed from label-token activation means after removing a shared component, while execution features are function vectors extracted from selected attention heads. The paper has several ablations and mechanistic interventions to show these are causally related within models and have a meaningful impact on naturalistic downstream tasks.

Critical strengths

Consistent evidence supports the author's hypothesis of a task-specific causal relationship between demonstration representations and downstream execution beyond simple linear readouts. The mean and zero ablations are particularly nicely done and informative on this point.

The ecologically valid examples in the coding experiments also point towards a meaningful realization of this hypothesis.

Critical weaknesses

The combined intervention establishes a downstream requirement for task execution, but does not isolate mediation of the identification-steering effect to the execution feature. The central experiment in Figure 5 shows that the behavioral improvement from identification steering disappears when the full function-vector direction is repeatedly zero-ablated at the cue token - so the improvement depends on this. However, it remains unclear whether the execution feature transmits the steering-induced task information or whether its ablation disables some other downstream process required to express the improvement. We'd need an intervention on the steering-induce task as well to establish that its disruption is the causal mechanism behind the downstream performance shift.

Questions for authors

I think a little more framing around the novelty would be helpful: the residual directions, shared predictive relationship, etc... many of the topics here have been addressed in other works in various parts or forms. This work is still quite novel in how it all comes together, but that should be contextualized better.

  • Dmitrii Kharlapenko, Stepan Shabalin, Arthur Conmy, Neel Nanda Proceedings of the 42nd International Conference on Machine Learning, PMLR 267:29819-29851, 2025.
  • Roee Hendel, Mor Geva, and Amir Globerson. 2023. In-Context Learning Creates Task Vectors. In Findings of the Association for Computational Linguistics: EMNLP 2023, pages 9318–9333, Singapore. Association for Computational Linguistics.
  • Lean Wang, Lei Li, Damai Dai, Deli Chen, Hao Zhou, Fandong Meng, Jie Zhou, and Xu Sun. 2023. Label Words are Anchors: An Information Flow Perspective for Understanding In-Context Learning. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 9840–9855, Singapore. Association for Computational Linguistics.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.