- Rating
- 2
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 3
- Soundness
- 3
Read review
Paper summary
The paper proposes criteria steering for adapting small LLM judges without updating their weights. It generates judging criteria from labeled examples, ranks them on validation data, and constructs an activation direction from stronger and weaker criteria. This direction is applied only to criterion tokens during prompt processing. Experiments on Bengali toxicity, Japanese safety, and financial compliance report improvements across several models.
Critical strengths
-
The intervention is simple and practical. Steering only the criterion tokens leaves model weights unchanged and requires no additional intervention during decoding. This offers a useful way to adapt existing judges to specific policies.
-
The improvements cover several models and domains. The Bengali results are particularly encouraging: Granite Guardian F1 increases from 0.882 to 0.951, with improvements across all five tested models (Table 1). This provides evidence that the method can improve judgment accuracy beyond one model or dataset.
Critical weaknesses
-
The label budget and confidence grouping need clarification. In Table 2, N counts only the examples used to generate criteria, while Sections 3.3-3.5 describe using additional validation labels to rank criteria and select steering settings. Section 3.5 specifies 353 validation examples for FinProof, so the N=16 result does not demonstrate adaptation from 16 labels overall. In Sections 3.1 and 3.3, assigning yes/no answers scores of 1/0 makes every parsed answer satisfy the confidence threshold. The grouping therefore cannot distinguish confident from uncertain valid answers. With no parsing failures, AUROC computed from these scores also reduces to balanced accuracy.
-
The comparisons do not establish superiority over prompt optimization or improvement on external data. In Section 3.5, the main prompt baseline uses a composed criterion, but the optimization budgets are not clearly matched. In Table 3, steering increases FinProof F1 from 0.957 to 0.964 relative to GEPA while also increasing the false-positive rate from 5% to 7.5%. This is a trade-off, and the manuscript provides no uncertainty estimates to establish whether the small F1 advantage is stable. In Table 4, the external Japanese result is below the published specialist average. Without an unsteered result using the same prompt, this comparison cannot establish whether steering improves performance on the external benchmark.
-
MINOR: The discussion of model abliteration should be expanded. In Section 2, refusal steering is briefly discussed through Arditi et al., but the connection to model abliteration and subsequent safety studies remains underdeveloped. The discussion could include Agnihotri et al. [1], Abu Shairah et al. [2], and Jain et al. [3]. A short comparison could explain the shared use of activation interventions, distinguish removing refusal-related components from adding a vector at criterion tokens, and clarify what is new about constructing that vector from ranked judging criteria.
References
-
Agnihotri, Shashank, Jonas Jakubassa, Priyam Dey, Sachin Goyal, Bernt Schiele, Venkatesh Babu Radhakrishnan, and Margret Keuper. "A granular study of safety pretraining under model abliteration." arXiv preprint arXiv:2510.02768 (2025).
-
Shairah, Harethah Abu, Hasan Abed Al Kader Hammoud, Bernard Ghanem, and George Turkiyyah. "An embarrassingly simple defense against LLM ablation attacks." arXiv preprint arXiv:2505.19056 (2025).
-
Jain, Samyak, Ekdeep S. Lubana, Kemal Oksuz, Tom Joy, Philip H. Torr, Amartya Sanyal, and Puneet K. Dokania. "What makes and breaks safety fine-tuning? A mechanistic study." Advances in Neural Information Processing Systems 37 (2024): 93406-93478.
Questions for authors
-
What is the total labeled-data budget for each method, including criterion ranking and all validation-based selection? Are these budgets matched across comparisons?
-
Were Bengali/Japanese confidence groups and AUROC computed from continuous probabilities or parsed yes/no outputs? If the latter, what distinguishes an uncertain answer from a parsing failure?
-
How were the prompt baselines and GEPA configured, and does steering retain an advantage at the same false-positive rate?
-
What is the external Japanese performance with the same minimal prompt but no steering? How was the reported variant selected from the two precomputed steering variants?
-
Are there parallels between the methodology of this work and works on model abliteration? If yes, could they please be discussed, and if not, then an elaborate response on why not would be very helpful.