- Rating
- 3
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 3
- Soundness
- 3
Read review
Paper summary
The paper proposes an approach for the unification of probe-based steering (which the authors call "writing"), and abstention-classification (which the authors call "reading"), into a single forward pass. Validation is done over several steering and classification probes.
Critical strengths
- The paper is exceptionally well-written
- The experimental procedure is exceptionally clear and, for the (insufficient) baselines and related work outlined, thorough
Critical weaknesses
1.The probing and abstention-classification literature is vast, and this work fails to reference much of it. Alternative probes cited/tested against in this work are old. The direction methodology exists already in the literature but the authors of this work fail to attribute it to any previous work at its introduction in the paper. Much of the steering literature, and even classification probing for MCQA tasks, SFT is generally compared against in the experimental results to validate the motivation of probing over more LLM training. 2. I think this work lacks any major novelty: linear abstention probes and MLP-based steering probes already exist in the form presented in this work. Perhaps the only novel contribution is the proposal of a single-forward-pass methodology which does both steering and classification (writing and reading). But, the impact of this is fairly low; it halves the inference-time cost. Also, being a simple log-linear model to recover the pre-edit representation, it’s not a particularly novel solution. With that said, it is surprising that the re-reconstruction map, which is shallow and presumably much less expressive than the hidden LLM layers whose effects it must predict, actually works. This is not explored in the work, though. 3. While the writing is good and the paper is clear, It is missing any real introductory motivation. While the contributions are laid out clearly, the authors make no effort to demonstrate their importance. Why is a writer abstraction useful? I don’t think it is, ``steering’’ is good enough. Why do we care about cross-domain generalization in this context? The one paragraph at the start of the section is insufficient, and also makes claims which go unsupported even by reference to existing work. Rather than presenting results for a methodology which has not yet been introduced in the introduction, I think the space should be allocated for a more careful explanation of the problem, the setting, and why the existing approaches are unsatisfactory. Even in the related work section at the end, the explanation of why these existing methods fall short is not given.
I'm rating this paper, as it stands currently, as a weak accept primarily based on the quality of the writing and the thoroughness of the experiments (with a couple exceptions). If the authors can (a) address the somewhat lacking grounding in the probing literature, (b) convince me that I have missed some novel aspect of their work, and/or (c) motivate their work more clearly, I will be open to raising my rating.
Questions for authors
- To my knowledge, the abductive-NLI dataset always has only two options, meaning random accuracy is 50% assuming a balanced dataset. The fact that Qwen 0.5B gets 50% accuracy without using a probe writer makes me wonder if, as often happens in LLM-based classification accuracy experiments, the LLM answers correctly but the output is not parsed properly. It seems hard to believe that a model that seems to have no ability to predict a task can improve based solely on a probe injection. Perhaps it is moreso the fact that the probe gets to train directly on the task. In that case, does the probe injection do better than SFT directly on the model’s weights?
- In table 1, are you training the probe and the LoRA adapter still as steering probes or as classification probes? That is, are you supervising directly on the probe output or are you still supervising on the eventual output through all the frozen LLM layers? I think the literature typically treats abstention classification as the former, but in l. 235 you say that neither LoRA or the probe architecture is changed.
- In your experiments showing that only the static direction reader generalizes cross-domain, did you replicate the median-based threshold calibration for the other methods as well, or only for the direction-based one?
2 comments
Response to Reviewer U9su (Part 2/2): Q1–Q3
Questions
Q1. Chance-level accuracy, parsing and comparison with SFT
Reply: Thanks for the question.
- Parsing. Accuracy is computed directly from the two option-token logits; no generated text is parsed.
- Supervision and hidden information. Our module receives supervised task training. A separate 1.5B diagnostic finds 70.6% accuracy from a supervised linear readout of frozen hidden states versus 50.8% from the model’s own logits, supporting the presence of underused task information (App. N).
- SFT comparison. Please see our response to W1, point 3, in Part 1.
Q2. Where is the supervision applied in Table 1?
Reply: Thanks for the question. The probe remains an activation-steering module, and LoRA remains a weight-adaptation module. Their architectures are unchanged; only the training target changes to context-sufficiency judgments. Both are supervised through the LLM’s final answer-token logits, with gradients flowing through the frozen backbone. We also added a logistic-regression classifier trained directly on frozen hidden states; this comparison further supports the direction’s more stable cross-domain performance (Table 1).
Q3. Was median calibration applied to all methods?
Reply: Thank you for highlighting this. Originally, only the direction used median calibration; we have repeated the comparison using the same calibration rule for all methods.
At 7B, calibration reduces cross-domain accuracy drops from 16.4 to 7.5 points for the activation-steering baseline and from 22.2 to 12.4 for LoRA; the direction drops 2.3 points. Although the gap narrows, matched calibration and threshold-free AUROC continue to support the direction’s more stable cross-domain performance across all four backbones (Table 1; App. J).
Thank you again for the time and care you have devoted to reviewing our work. Your feedback has been valuable in improving the paper, and we would be happy to address any remaining questions or concerns.
Response to Reviewer U9su (Part 1/2): W1–W3
We sincerely thank the reviewer for the thoughtful review and constructive suggestions. Your questions helped us clarify why this problem matters, what our work contributes, and how to present the paper’s overall argument more coherently.
Weaknesses
Reply: Thank you for highlighting these gaps. We have revised the introduction and contributions (§1) and expanded the related work (§6) to improve clarity and logical flow:
Reply: Thank you for highlighting these points. We have clarified three aspects:
Reply: Thank you for helping us clarify the motivation. We have revised the introduction (§1) to explain the problem, the gap in prior work and the supporting references more clearly.