acceptodds
Under review as a conference paper at ICLR 2027

Multi-Agent Collusion and the J-Space

14 views

Abstract

LLM agents are increasingly deployed in Multi-Agent Systems (MAS), and they introduce the risk of coordination among agents that might escape human oversight. The J-space has been used to uncover deception in single LLMs, but its use for uncovering and steering collusion between agents remains unexplored. We show, with interpretability experiments on DeepSeek-V4-Flash (284B) on the NARCBench collusion benchmark, that colluding agents' J-space holds the collusion itself: before replying, 86% of colluders represent misrepresentation, concealment or remorse. In a steganographic card-counting task, the receiver's J-space holds a hidden count, and causal interventions alter its count. We also introduce the Jacobian Pair Contrast (J-PACT), an unsupervised method for detecting collusion between agents. We evaluate J-PACT on four open-weight models: GPT-OSS-20B, Llama-3.3-70B-IT, Qwen3-32B, and DeepSeek-V4-Flash (284B). J-PACT retains 85% of the above-chance AUROC of supervised NARCBench probes, matches them on steganographic scenarios, and names the colluding pair up to its complement in 99% of detected runs.

Primary area:
interpretability and explainable AI
Keywords:
Multi Agent, LLMs, Interpretability, AI Safety

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_9QUXOpenReview ↗
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 (α=1\alpha=1), 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

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.

  • Reading. The pair contrast now reads every token of each public message, in the workspace layers, after the colluders' instructions are removed from every prompt. This follows your W1: the colluding pair shares private instructions, so no detector sees them any more. The place of reading was a failure mode of the submission. The workspace holds what the model needs to decide what it will say next, and at the last token of a response that is punctuation, not the collusion. Reading mid-message is also the more realistic setting: one would rather catch collusion as it happens than after the fact.
  • Score. J-PACT writes what each pair shares as an agent-pair table. An LLM monitor reads the table with the public transcript, and Eq. 7 averages its answer with the pair contrast. Both parts are label-free, and the monitor costs one forward pass per run.

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<0.002p < 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.

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.

Reviewer_s6nnOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper extends the J-space (Jacobian lens, Gurnee et al. 2026) from single-LLM deception detection to multi-agent collusion detection. Interpretability experiments are conducted on DeepSeek-V4-Flash (284B) using the NARCBench collusion benchmark: in Core scenarios, 86% of colluders hold misrepresentation/concealment/remorse in their J-space before replying (versus only 33% under honest instructions), and in 70% of collusion runs the colluding pair exclusively shares these concepts (versus 8% in control). In the Stego steganographic card-counting task, the receiver's J-space holds the hidden count; a causal intervention swapping its hot/negative J-lens coordinates causes 57% of negative hands to be read as hot (versus 2% for random swaps). The paper then proposes J-PACT, an unsupervised collusion-detection method based on a "within-pair vs outside-pair concept sharing" contrast. Across four open-weight models, J-PACT retains 85% of the supervised probes' above-chance AUROC on Core (0.921 vs 0.993), matches the probes on Stego, but reaches only 0.519 on the Transfer split (probes: 0.729). With N=4, J-PACT can only localize the "pair split" rather than the colluding pair itself.

Critical strengths
  1. The Stego causal intervention is clean mechanistic evidence. Swapping the receiver's two J-lens concept coordinates causes it to read 57% of negative hands as hot, while random swaps yield only 2%—this elevates "the receiver decodes the steganographic signal through its J-space" from correlation to causation. This is the paper's most solid contribution and provides a foundation for future J-space-based intervention/defense.
  2. Unsupervised, probe-complementary practical positioning. In unanticipated collusion events such as those at OpenAI/GDM, labels do not exist; J-PACT requires no labels to localize a suspicious pair split, and its AUROC of 0.921 on Core is meaningful. The authors honestly note the Transfer failure and the N=4 "pair split" limitation without overstating the results.
Critical weaknesses
  1. The Transfer-split failure (AUROC 0.519) severely limits applicability. When agent roles are structured and each agent is observed only once, J-PACT is near-random—and this is exactly the common form of real multi-role collaboration. The paper attributes this to the violation of the "honest agents are exchangeable" assumption but gives no correction or degradation bound for when this assumption fails, compressing the method's practically usable setting to the narrow band of "same model + same role + multiple turns."
  2. Same-model assumption and single-model readout. J-PACT requires all agents to be instances of the same LLM (sharing J-space coordinates), which does not hold in heterogeneous multi-provider systems—yet the real danger of collusion often arises precisely in cross-model collaboration. Moreover, the headline 86%/70% figures on Core come only from DeepSeek-V4-Flash, with no corresponding readout rates reported for the other three models on Core, leaving the generality of "colluders hold the collusion in J-space" insufficiently supported.
Questions for authors
  1. Beyond DeepSeek-V4-Flash, what are the rates at which "colluders hold misrepresentation/concealment/remorse" on Core for GPT-OSS-20B/Qwen3-32B/Llama-3.3-70B? If these differ substantially, how general is "the J-space holds the collusion"?
  2. On the Transfer split, J-PACT achieves AUROC 0.519. Can role-conditioned normalization or role-based conditioning be introduced under the "heterogeneous agent roles" setting to recover above-chance performance, or is this failure fundamental?
  3. The same-model assumption rules out heterogeneous MAS. Is there a cross-model J-space alignment scheme (e.g., shared unembedding or projection into a common token space) that would allow J-PACT to extend to heterogeneous systems?
  4. With N=4, J-PACT can only localize the pair split. In experiments with N>4 (if conducted), can it truly localize the colluding pair? Please report results for N=6/8.
4 comments

Thank you for the careful review, and for calling the Stego intervention "clean mechanistic evidence" and J-PACT's label-free positioning practical. You asked four concrete questions, which we answer experimentally.

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.

  • Reading. The pair contrast now reads every token of each public message, in the workspace layers, after the colluders' instructions are removed from every prompt. The place of reading was a failure mode of the submission. The workspace holds what the model needs to decide what it will say next, and at the last token of a response that is punctuation, not the collusion. Reading mid-message is also the more realistic setting: one would rather catch collusion as it happens than after the fact.
  • Score. J-PACT writes what each pair shares as an agent-pair table. An LLM monitor reads the table with the public transcript, and Eq. 7 averages its answer with the pair contrast. Both parts are label-free, and the monitor costs one forward pass per run. This follows your Q2.

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 in the same setting.

W1, Q2. Is the Transfer failure fundamental? No. J-PACT now reaches 0.815 on Transfer (mean of four models), above the supervised probes for every model (0.672 on the mean) and above every label-free baseline (Table 2). The new setting takes the pair contrast from chance to 0.652. The monitor, which reads each pair's shared concepts in the context of the roles, takes it to 0.815. The agent-pair table matters: the same monitor reading only the transcript reaches 0.714 (Table 3).

Q4. Does it name the pair with N=6N = 6 and 88? Yes, and thank you for asking: this removes the limitation you point out. We extended NARCBench's generator to six and eight agents. J-PACT then names a single pair, and it is exactly the colluding pair in 70% and 75% of Core collusion runs (chance: 6.7% and 3.6%; matching control runs: 10% and 7%).

W2, Q3. Agents on different models? Yes. We extended NARCBench to conversations in which half of the agents are Qwen3-32B and half Llama-3.3-70B-IT. J-PACT is computed in the J-space of each of the two models, and we average the two. On Transfer it is above the supervised probes with four, six and eight agents (0.706, 0.760, 0.741 against 0.695, 0.660, 0.709). On Stego it stays at 1.00 while the probes fall to between 0.76 and 0.89 (Table 4).

W2, Q1. The readout on the other three models? To compare models we needed one vocabulary for all four, so we fixed it before reading: 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, and 93% to 100% do when the honest agents are given the colluders' instructions (Table 1). The DeepSeek readout of the submission is unchanged, in Appendix D.

The limits that remain are stated in Section 7: we incur a computational cost equal to the LLM monitor's forward pass (much lower than the hundreds of forward passes that the scenarios require, making it a sustainable cost), and with four agents we name the split into two pairs, while in the more realistic settings with more than four agents we name the pair itself.

Unanticipated collusion, as in the incidents we cite, comes without labels. That is the setting J-PACT is for, and it now covers different roles, different models and larger groups. Code and caches are updated at the links in the paper. We would be grateful to know whether this addresses your concerns.

A short pointer to our answers to your four questions: Q1 → Table 1 (all four models); Q2 → Table 2 (Transfer now 0.815); Q3 → Table 4 (mixed models); Q4 → Table 5 (six and eight agents). 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 runs on Transfer, heterogeneous models.

These experiments are real, but they do not show that the submitted method already worked. They change the method.

W1 / Q2. The original pair-contrast J-PACT was near-random on Transfer (0.519). Under the new mid-message, instruction-stripped reading, the pair contrast alone is 0.652. The jump to 0.815 comes after adding an LLM monitor that reads the agent-pair table together with the transcript. The same monitor on the transcript only already reaches 0.714. So most of the Transfer recovery is not a fix of the Jacobian contrast I reviewed; it is a new label-free LLM judge sitting on top of a still-weak contrast (0.652). I do not treat “Transfer is no longer fundamental” as established for J-PACT as submitted.

Q4. N=6/8 pair-naming at 70% / 75% (chance 6.7% / 3.6%) removes the “only a pair split at N=4” limitation. This is a useful addition. It is a new experimental protocol, not evidence that the original N=4 detector localized the pair.

W2 / Q3. Averaging J-PACT in Qwen3-32B and Llama-3.3-70B J-spaces, then reporting Transfer 0.706–0.760, is a workaround, not a cross-model alignment of J-space. Heterogeneous MAS was the setting I asked about. Coordinate-wise averaging of two unaligned concept bases is not that scheme, and I do not know how much of Table 4 is the monitor rather than the J-contrast.

Q1. The four-model readout with a fixed five-word vocabulary (81% / 100% / 100% / 94% of colluders above honest agents) is the right measurement. I accept that the “J-space holds the collusion” claim is no longer DeepSeek-only. The original 86%/70% Core figures were still a single-model, end-of-response readout; they should not be cited as if they were this new protocol.

The Stego intervention remains the paper's strongest result. The rebuttal, however, replaces the method I scored (pair contrast, end-of-response, same-model, Transfer ≈ chance) with a different pipeline. That does not make the original weaknesses disappear, and I do not raise a score on the basis of a revised system that was not in the submission. I keep my rating at 2 (weak rejection).

We thank the reviewer for the reply. We respond directly to the reviewer's central claim, that the revision cannot count towards the rating because it changes the method.

The changes are answers to your questions. Q2 asked whether role-based conditioning could recover Transfer, Q3 whether J-PACT could extend to heterogeneous models, and Q4 for results with six and eight agents. None of these can be answered without changing the method or the protocol. It is not consistent to request them and then set the answers aside because they are changes.

The Jacobian contrast is not a weak base under an LLM judge. With the equations as submitted, the new reading alone takes Transfer to 0.652. An LLM judge on the transcript alone reaches 0.714. J-PACT reaches 0.815 (p = 0.022 against that judge), above the supervised probes (0.672). The gap between 0.714 and 0.815 is the J-space. On Core and Stego the monitor adds nothing (0.921 → 0.916, 1.000 → 1.000).

Table 4 does not average concept coordinates. J-PACT is computed separately in each model's own J-space, and the two scores are averaged. No alignment of bases is involved.

Pair naming. With four agents and one colluding pair, the contrast names the pair up to its complement, as the submission stated. This limit applies only to very small multi-agent systems, as deployed multi-agent systems, and the incidents that motivate the paper, involve more than four agents. There the same contrast names the exact pair, in 70% and 75% of runs with six and eight agents (chance: 6.7% and 3.6%).

The research question and the pair contrast's equations are unchanged. The Section 3 readout you accept and the Stego intervention you call the strongest result do not involve the monitor. ICLR's discussion period permits revising content, and we ask that the paper revised on the basis of your review be the one assessed.

Reviewer_nYVZOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The authors apply an existing mechanistic interpretability method - J-space detection - to multi-agent collusion scenarios from NARCBench. The empirical findings suggest that their novel detection metric J-PACT has power on a narrow set of tasks.

Critical strengths

The approach chosen by the authors is interesting, and the underlying questions of how to identify colluding pairs of agents is of great interested to the community.

Critical weaknesses

The main critical weakness of the paper concern the generality of its findings and the resulting limited practical value:

  • assumption that agents are operating the same models and that honest agents are exchangeable
  • overstated results wrt identifying colluding pairs
  • lack of comparison to simple unsupervised baselines
Questions for authors
  • can J-PACT be grounded in the formal statistics literature, e.g. in composite hypothesis testing? If so, under what assumptions?
  • how does J-PACT perform relative to simple unsupervised baselines operating on internal model activations? Does it add improved practical performance?
  • NARCBench itself has important limitations; e.g. collusion is instructed not emergent and also agent number is fixed to 4, with one colluding pair. How would you overcome these limitations in future work?
4 comments

Thank you for the review, and for writing that how to identify colluding pairs "is of great interest to the community". Your concerns are about how general our findings are. The revision addresses generality directly. 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.

  • Reading. The pair contrast now reads every token of each public message, in the workspace layers, after the colluders' instructions are removed from every prompt. The place of reading was a failure mode of the submission. The workspace holds what the model needs to decide what it will say next, and at the last token of a response that is punctuation, not collusion. Reading mid-message is also the more realistic setting: one would rather catch collusion as it happens than after the fact.
  • Score. J-PACT writes what each pair shares as an agent-pair table. An LLM monitor reads the table with the public transcript, and Eq. 7 averages its answer with the pair contrast. Both parts are label-free, and the monitor costs one forward pass per run. This is what removes the exchangeability assumption of your W1.

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 in the same setting.

W1. Same model, exchangeable honest agents. Neither is assumed any more. Different roles: on Transfer, where every agent has its own role, J-PACT reaches 0.815 (mean of four models), above the supervised probes for every model (0.672) (Table 2). Different models: we extended NARCBench to conversations in which half of the agents are Qwen3-32B and half Llama-3.3-70B-IT. J-PACT is computed in the J-space of each of the two models, and we average the two. On Transfer it is above the supervised probes with four, six and eight agents (0.706, 0.760, 0.741 against 0.695, 0.660, 0.709) (Table 4).

W2. Overstated identification of the colluding pair. We agree, and the claim is now exact. With four agents we state what is identified: the split into two pairs. With six and eight agents J-PACT names a single pair, and it is exactly the colluding pair in 70% and 75% of Core collusion runs, against 6.7% and 3.6% by chance and 10% and 7% in the matching control runs (Table 5).

W3, Q2. Simple unsupervised baselines on activations: does J-PACT add practical performance? Yes, where roles differ. Table 2 adds four label-free baselines. Two read the internal activations: Activation Alignment and the Mahalanobis distance. Two read the text: sentence embeddings and an LLM monitor of the transcript. On Transfer J-PACT reaches 0.815 against 0.633, 0.564, 0.582 and 0.714 (paired bootstrap: p=0.010p = 0.010, p<0.002p < 0.002, p<0.002p < 0.002, p=0.022p = 0.022). On Core and Stego, where all agents share a role, the activation baselines are already near ceiling. J-PACT also names the pair and the concepts it shares.

Q1. Composite hypothesis testing. Thank you for the suggestion: Section 4 now opens with this framing. H0H_0: all agents pursue the intended task. H1H_1: some agents collude, composite because neither the number nor the identity of the colluders is specified. J-PACT's score targets one property that H1H_1 has and H0H_0 lacks: a colluding pair shares concepts beyond what each shares with the rest of its table. That is the one assumption it needs; it needs neither a shared model nor exchangeable honest agents (W1). We state where type I and type II errors come from. We do not claim a test of fixed level: we evaluate the score by AUROC.

Q3. NARCBench's limits. We lifted the two that concern our paper in this revision. Agent number: we extended NARCBench to six and eight agents, where J-PACT remains the best label-free detector on the role-structured split and stays above the supervised probes (Table 4). One colluding pair: we also generated eight-agent runs with three colluders (Table 5). Instructed collusion: the instructions are removed before anything is read, so no result rests on the instruction text. Collusion that emerges without instructions is the next step we name in the Conclusion.

Section 7 states the limits that remain, chiefly the monitor's forward pass, which is far below the hundreds of forward passes the scenario itself requires.

The incidents that motivate the paper came without labels and without warning. A label-free detector that names the agents to inspect is what an operator has in that situation, and it now holds across roles, models and group sizes. Code and caches are updated at the links in the paper. We would be grateful to know whether this addresses your concerns.

A short pointer to our answers: W1 → Tables 2 and 4 (different roles, different models); W2 → Table 5 (exact pair with six and eight agents); W3 and Q2 → Table 2 (four label-free baselines); Q1 → Section 4. Do these resolve your concerns? If so, we would be grateful if you would consider updating your rating.

Thanks for the substantial revisions, which have clarified my concerns wrt pair identification. I keep my rating, however, as an evaluation of the overall approach based on its merits shows that J-PACT as the detection contribution remains limited.

Thank you for the reply, and for confirming that the pair-identification concern is resolved.

On the detection contribution, we would point to two results, which you requested.

  • Against the baselines you asked for (Table 2). On Transfer J-PACT reaches 0.815, against 0.633, 0.564, 0.582 and 0.714 for the four label-free baselines (p ≤ 0.022) and 0.672 for the supervised probes.
  • The ablation (Table 3). Neither part reaches this alone: the pair contrast gives 0.652 and the LLM monitor on the transcript 0.714. The gain to 0.815 comes from adding the Jacobian pair table and contrast. Without the J-space the result is not there.
  • Localisation (Table 5). Beyond the run-level score, J-PACT names the colluding pair (70% and 75% with six and eight agents, against 6.7% and 3.6% by chance) and the concepts it shares.

If "limited" refers to Core and Stego, where simple activation baselines are also near ceiling, we agree and say so in the paper.

The contribution is in the role-structured and mixed-model settings, which are the realistic ones. We would be grateful to know whether there is a specific comparison you find missing.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.