- Rating
- 2
- Confidence
- 2
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
The paper applies the Jacobian lens to multi-agent collusion on NARCBench, using the fact that agents running the same model share a J-space. On DeepSeek-V4-Flash, it reports that colluders hold misrepresentation, concealment or remorse concepts before replying (86% vs. 33% for honest counterfactuals), and that the colluding pair often shares such concepts with no other agent. It also reports that a steganographic receiver holds the hidden count, and that swapping its HOT and Negative J-lens coordinates changes the range it decodes. Finally, the paper proposes J-PACT, an unsupervised pair-contrast detector. Across four models, J-PACT reaches 0.921 AUROC on Core (vs. 0.993 for supervised probes) and is at chance on Transfer. When it flags a run, it names the colluding pair up to its complement in 99% of cases.
Critical strengths
- Novel angle. Comparing J-space readouts across agents gives human-readable evidence of what a suspicious pair shares. J-PACT turns this into a label-free detector that also localizes the pair, which run-level probes cannot do.
- Candid about scope. The paper clearly states its assumptions and its failure on Transfer, includes a random-pair steering control, and releases code.
Critical weaknesses
- The interpretability evidence has a gap. Appendix D states that the concept words and range signatures were selected from the same readouts on which they are later evaluated, so the 86% vs. 33% and the Stego results are in-sample. The colluders' system prompts also describe the deception explicitly, and the honest counterfactual removes these instructions. The paper therefore cannot separate "holds the collusion" from "holds its instructions". The same question applies to J-PACT, since the colluding pair shares private instructions.
- The causal claim is stronger than the data in Table 6. Even at the exact exchange (), HOT hands are never decoded as NEGATIVE, and NEUTRAL hands drift toward HOT. The result therefore looks like a one-directional shift rather than a swap. The intervention also edits the directions of the answer tokens themselves, so the alternative explanation of a lexical effect is not ruled out.
- The detection evidence is narrow. Stego is at ceiling for every method. Transfer, the only split with heterogeneous roles, is at chance. With four agents, localization only identifies the split into two pairs. These assumptions differ from the real-world incidents used as motivation.
Questions for authors
Same as wekanesses above.
3 comments
A short pointer to our answers: W1 → Table 1 (vocabulary fixed in advance; honest agents given the colluders' instructions); W2 → Table 6 (swap in both directions); W3 → Tables 2, 4 and 5. Do these resolve your concerns? If so, we would be grateful if you would consider updating your rating.
I thank the authors for the new results and clarifications. Based on the method’s contribution I maintain my original assessment.
Thank you for the review, and for noting the readable evidence of what a suspicious pair shares and that the paper is candid about its scope. Each of your three weaknesses led to a new experiment. First, we state what changed in the method.
What changed in J-PACT. In the submission, J-PACT was the pair contrast alone, read at the end of each response with the colluders' instructions still in the prompt. The revision makes two changes, each prompted by the reviews, and keeps the pair contrast's equations as they were.
Table 3 separates the two changes. Under the new reading, the submitted pair contrast alone reaches 0.921 on Core, 0.652 on Transfer and 1.000 on Stego. With the monitor, J-PACT reaches 0.916, 0.815 and 1.000. Every baseline and the supervised probes are evaluated under the same settings.
W1. Were the concept words selected from the readouts they are scored on? Not any more. Section 3 now rests on a vocabulary fixed before reading and shared by four models: the five deception words of Gurnee et al. 81%, 100%, 100% and 94% of the colluders of GPT-OSS-20B, Qwen3-32B, Llama-3.3-70B and DeepSeek-V4-Flash score above the honest agents of their run before replying (Table 1; Holm-adjusted p<0.002 on every model).
W1. "Holds the collusion" or "holds its instructions"? We now test this directly. The honest agents of the same run are given the colluders' instructions, quoted in their own prompt, and are read at the same token. 97%, 95%, 100% and 93% of the colluders still score above them (Table 1, Control 2). Reading the instructions and carrying them out are separated.
W1. The same question for J-PACT. Answered by the first change above: J-PACT, the baselines and the probes all read prompts without the instructions.
W2. Is the swap one-directional, and could it be lexical? You were right that the submitted edit acted as a one-directional shift, which led us to the fix. The submission exchanged the HOT and Negative coordinates symmetrically on every hand. The revision swaps as Gurnee et al. do, on the concept the receiver holds: the hand's own range is subtracted and the same amount added to the other, and NEUTRAL hands are not edited. 92% of NEGATIVE hands are read as HOT and 95% of HOT hands as NEGATIVE, against 2% without it (Table 6). On a lexical effect: the swap is never applied where the answer is emitted, and twenty random pairs swapped the same way read at most 2% and 10% of hands as the other range.
W3. Is the detection evidence narrow? Three additions. Transfer is no longer at chance: J-PACT reaches 0.815 (mean of four models), above the supervised probes for every model (0.672) and above four label-free baselines (Table 2). Stego separates methods once the agents run different LLMs: the probes fall to between 0.76 and 0.89 and J-PACT stays at 1.00 (Table 4). With six and eight agents J-PACT names exactly the colluding pair in 70% and 75% of Core runs, against 6.7% and 3.6% by chance (Table 5). For this we extended NARCBench itself to different LLMs, to six and eight agents and to three colluders. The assumptions you list as separating our evidence from the real incidents are the ones we removed: J-PACT now detects collusion among agents with different roles, on different models, in groups larger than four, without labels and without seeing the colluders' instructions.
Section 7 states the limits that remain: the monitor's forward pass, far below the hundreds the scenario requires, and with four agents the split into two pairs.
The incidents that motivate the paper came without labels. The revision shows that colluders are separated from honest agents who read the same instructions, and that a label-free detector built on this names the agents to inspect. Code and caches are updated at the links in the paper. We would be grateful to know whether this addresses your concerns.