- Rating
- 3
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 3
- Soundness
- 2
Read review
Paper summary
The paper frames automated red-team auditing as a Markov decision process, where the state is the conversation so far and the reward is an LLM judge's score for a target behavior. The authors replace the auditor-driven search of Petri with beam search, and rank branches with Value Probes: linear probes on the target model's mean-pooled activations at one layer, trained to predict discounted future judge scores. The method is evaluated on 18 behaviors with Gemma 3 4B, 12B, and 27B as targets, against a token-matched Petri audit, beam search with random selection, and beam search ranked by a Claude Opus 4.8 judge. The probe arms reach a mean peak judge score of 0.79 to 0.85, against 0.51 to 0.57 for Petri (Table 1). Probes trained on Gemma 3 4B rollouts are also tested on Gemma 3 12B, Gemma 3 27B, and Gemma 4 31B (Table 2).
Critical strengths
-
The MDP framing separates the search from the ranking signal, and the random-selection arm shows how much of the gain over Petri comes from the search structure alone (Section 5.2).
-
The Max Value Probe is within 0.04 of the Claude Opus 4.8 judge in mean peak score at every model size, using a single linear readout instead of 288 extra judge calls per audit (Section 5.3).
Critical weaknesses
-
The results suggest that much of the gain over Petri comes from beam search itself, and the probe's own contribution is not separated from simpler rankers. Beam search with random selection outperforms Petri by 0.24 at 4B and 0.15 at 12B and 27B, while the probes add 0.06 to 0.10 (Section 5.2). There is no ranker that predicts only the current node's judge score, and none trained on the same targets from the transcript text alone, so it is unclear how much the forward-looking target and the activations each contribute.
-
The transfer claim for Gemma 4 31B is not supported by the experiment. On 31B the transferred probes are compared only with random selection (Table 2), yet Section 5.5 states that transfer to a new model comes at no measurable cost relative to an on-policy probe. The Conclusion also compares the 31B scores with the Gemma 3 scores, although the Figure 3 caption says the 31B runs use a different judge and should be compared only within that panel.
Questions for authors
-
Could the authors compare the probe with a simpler ranker trained on the same data, such as a probe that predicts only the current node's judge score, or a small model that reads the transcript text? This would show how much the forward-looking target and the activations each contribute.
-
Would a Petri audit on Gemma 4 31B be feasible? This would help support the claim that transfer comes at no measurable cost for a new model.
Thank you very much for your review, these were very good comments. We attempted to acquire budget for these follow up experiments, but were unable to be approved before the deadline. However, although we do not have results from a full sweep, we do have some partial results from experiments we ran earlier we can share in our comment. We did not train probes for this, but we did find that using smaller judge LLMs with the blackbox method we detail, produced lower scores during auditing. We have notes from early experimental runs that suggest that a small enough capacity judge provides little to no lift over the random heuristic. As far as a full Petri audit on Gemma 4 31B, this is a good call out, we really want to include results for these experiments, but due to the above budget constraints, we weren't able to provide these results before the review. We've added an additional cost analysis section which gives a sense of the costs for conducting our audits, harvesting rollouts, and training new probes on new rollouts. If our budget request is approved later, we are happy to post results as a follow up for completeness. Thank you again!