- 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 ( 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 and ).
-
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 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, 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 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The models considered in the empirical analysis are not sufficient (only two models given), largely outdated (Qwen2.5 used, e.g., why not Qwen3?), and don't target a variety of model scales (only 6B and 7B).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
- Under label steering, how much of the gain does mean ablation of the task-unique FV component remove, relative to its other-task control?
- How well do non-label representations (query-token states, output-word embeddings) predict held-out under the same protocol, and does identification ablation still work when restricted to the band rather than layers 0–27?
- Why didn't the author consider more and recent models?