acceptodds
Desk rejected at ICLR 2027

Multi-Agent Sycophancy Is a Pretrained Mid-Layer Vulnerability

Desk rejected at ICLR 2027, so left out of the paper lists and the map.

Abstract

Multi-agent LLM pipelines route model outputs between instances, where a peer asserting a wrong answer can flip a subject model from correct to incorrect at rates we term *yield*. This is widely attributed to RLHF-induced sycophancy, which would make better post-training the fix. We test this across four model families and find standard alignment is not the cause: pretrained base models exhibit the same substitution pattern as their Instruct variants and usually yield more, and which model family a system is built on explains roughly nine times as much of the variation in yield as whether that model was instruction-tuned. Activation patching on Llama-3.1-8B-Instruct and on its pretrained base localizes the corruption to the same narrow mid-layer window, where attention carries the causal weight and MLP contribution is negligible; patching above it restores 96% of the clean-to-pressured P(correct) gap on Instruct and at least 88% on base. On both, activation-space interventions show that pressure suppresses clean-reasoning features rather than activating a new sycophancy circuit. The attack surface decomposes into two independent factors (channel framing and consensus strength) whose interaction produces a 47.5 percentage-point yield gap at majority consensus, preserved across jury sizes ; yield is highest through tool returns, the channel deployed function-calling pipelines use. A single correctly-arguing dissenter reduces yield by more than 50 percentage points across all framings tested, whereas the strongest prompt-level defense fails on attack variants outside its design surface. Because the vulnerability is pretrained and mid-layer, mitigations should target the mechanism, structured dissent at the pipeline level, rather than prompt-level defenses.

Primary area:
interpretability and explainable AI
Keywords:
mechanistic interpretability, activation patching, sycophancy, multi-agent LLM systems, sparse autoencoders, residual stream, large language models, 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_xnsdOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies peer responses' influence over multiple-choice answering. Pretrained and instruction-tuned models are compared under peer responses of different presentation and agreement level. Activation patching and feature interventions are used to test internal changes. It reports that it happens at pretrained model and a mid-layer interventions can restore probability. Additionally, a correctly answering dissenter reduces the influence of incorrect consensus.

Critical strengths
  1. The Base/Instruct comparisons provide evidence that the observed susceptibility is present in pretrained models and is generally reduced by the post-training pipelines. The within-family matched question pools and the additional base-model experiment without a chat template strengthen this finding.

  2. The analysis includes interventions and correlations. Residual-state and component-level patching provide evidence that mid-layer states are important intervention points. The controls for prompt length, attribution, and consensus wording help characterize the results.

Critical weaknesses
  1. It is unclear how closely the main yield measure tracks what the model final answers. The paper identifies a calibration problem when the answer suffix is removed, but does not clearly show whether the defense results were rechecked after fixing it.

  2. The claim that there is no separate process making the model follow others seems too strong, such a process could be what weakens its original answer signals. And the dissenter always supplies the correct answer, so its benefit may come from providing that answer rather than simply disagreeing.

Questions for authors
  1. Could you report final answer changes alongside yield and show how often they agree? Were the defense results without the answer suffix rerun using the corrected calibration?

  2. Why do the results support the claim that no separate process makes the model follow others, rather than that process causing the original signals to weaken? For the dissenter, what happens if it supports a different wrong option, or answers independently without being given the correct answer?

5 comments
AuthorsOpenReview ↗

Response to Reviewer xnsd

We thank the reviewer for two sharp questions. Both are answered with new measurements below, and both sharpened how we state the result.

Q1a: "Could you report final answer changes alongside yield and show how often they agree?"

Table 1 now carries a flip-rate column. We re-ran all 16 main conditions and the N=4 sweep: every published yield reproduces within 1 pp, the flip rate lies within −0.75 to +6.00 pp of yield across all 26 conditions, and an independent second re-run gives per-question agreement of 85.00–100.00% per condition, 95.47% on average (Appendix F.1, Table 18; representative rows in Table R2 of our response to Reviewer DUNd).

Flip is slightly higher than yield because it counts a move to any wrong option while yield counts only a move toward the jury's target, so the yields reported throughout the paper are the conservative measure. The gradients the paper compares are changes in the answers the model gives, not an artifact of the representation-level readout.

Q1b: "Were the defense results without the answer suffix rerun using the corrected calibration?"

They were, and the rerun gave a cleaner measurement than the one we had. The position-matched basis does not resolve the problem for defended prompts: under these system prompts the model defers its decision to the reasoning it writes, so the readout taken before generation is close to chance. We therefore measured the defenses on the answer itself, which requires no calibration (Table R3 in our response to Reviewer DUNd; Appendix C.3, Table 7).

The strongest prompt does protect freely written answers: 73.25% correct against 25.25% undefended. The revision reports this in place of the earlier estimate, which was taken from a readout we can now show is uninformative for defended prompts. We have revised §5.6, the Discussion and Appendix G.1 accordingly.

What this does not change is the defense's actual limitation, which is the one the paper's argument uses: its effect falls from −65.00 pp on the attack it was written for to −45.50 on a count-matched anonymous claim, −28.00 on a bare assertion inside the jury, and −13.75 with no jury at all (Appendix C.3, Table 6). The contrast with the dissenter is coverage across phrasings, not the readout.

Q2a: "Why do the results support the claim that no separate process makes the model follow others, rather than that process causing the original signals to weaken?"

The reviewer is right, and we have sharpened the claim. Our evidence does not exclude a consensus-detecting process; our own SAE taxonomy describes consensus-signal features that rise under pressure (Appendix D.3, Family 2), so such a process plausibly exists. What our interventions show is where the causal weight sits at the output: clamping the features pressure turns on moves P(wrong) by −2.3 pp, against −15.6 pp for the features it turns off (Table R4 above).

A detector whose output drives the suppression is entirely consistent with that, and is arguably the natural reading of the two together. We have therefore replaced "pressure suppresses clean-reasoning features rather than activating a new sycophancy circuit" with "pressure acts mainly by suppressing the correct answer rather than by writing in the wrong one," restated the −2.3 vs −15.6 pp result as a bound on rather than an exclusion of an additive component, and changed "nothing to ablate" in the introduction to "no separable downstream component to ablate." The heading of §5.5 now states this directly.

This matters for the paper's recommendation, which is unaffected: whether or not a detector exists upstream, the thing that reaches the output is the removal of the model's own reasoning, so a downstream corrector would have to rebuild what was suppressed, and the effective intervention point is upstream.

AuthorsOpenReview ↗

Response to Reviewer xnsd cont.

Q2b: "For the dissenter, what happens if it supports a different wrong option, or answers independently without being given the correct answer?"

Both run, both in Table R1 of our response to Reviewer DUNd (Appendix C.4, Table 9).

A dissenter naming a different wrong answer is followed on only 3.50–5.50% of questions; accuracy rises from 20.75 to 79.50% under user-role framing and from 2.00 to 39.00% under assistant-role framing, because the model returns to its own correct answer rather than adopting the dissenter's. A dissenter carrying an independently queried model's answer, with no answer key anywhere in the pipeline, matches the correct-answer dissenter in every framing (84.00 vs 85.25% correct, user-role).

Weakness: "the dissenter always supplies the correct answer, so its benefit may come from providing that answer rather than simply disagreeing."

The wrong-option result is the direct test of exactly this, and it separates the two: if the benefit came from supplying an answer, a dissenter supplying a wrong one would move the model to that wrong answer. It moves there on 3.50–5.50% of questions and to the correct answer on 79.50%.

Although I appreciate the authors’ clarifications and additional experimental results, there are prior works including Zhu et al. (ACL 2025) showing that instruction-tuned models conform less than base and model is more likely to keep its own answer when peer disagrees. I will maintain my original score.

Reviewer_NjjAOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This work investigates the mechanisms underlying multi-agent LLM sycophancy, demonstrating that the vulnerability originates during pretraining rather than RLHF. Through cross-model comparisons and mechanistic analyses, the authors identify an attention-dominated vulnerability in layers 14–18, primarily involving the suppression of clean-reasoning features. Furthermore, it examines the effects of channel framing and consensus strength and proposes structured dissent as an effective mitigation strategy.

Critical strengths

Strengths:

  1. It challenges the conventional attribution of sycophancy to RLHF by comparing pretrained and instruction-tuned models. This provides valuable insights into the origins of multi-agent vulnerabilities.

  2. It combines activation patching, linear probes, and sparse autoencoders to investigate the underlying mechanisms. The identification of an attention-dominated vulnerability in middle layers strengthens the mechanistic contribution.

  3. It systematically examines how channel framing and consensus strength affect model behavior. The proposed structured-dissent strategy demonstrates substantial improvements across different attack conditions.

Critical weaknesses

Weakness:

  1. It focus on single-step injections and multiple-choice questions. It remains unclear whether the findings generalize to multi-round debates and open-ended reasoning tasks.

  2. Although the analyses support the feature-suppression hypothesis, the interventions only partially restore correct-answer probabilities. This limits the strength of the causal conclusions regarding the underlying mechanism.

  3. The structured-dissent experiments assume that the dissenting agent provides the correct answer. Its effectiveness under realistic conditions involving unreliable agents or correlated errors remains uncertain.

Questions for authors

Q1: Does the identified mid-layer vulnerability generalize to multi-round agent debates and open-ended reasoning tasks? Would similar mechanisms emerge in more realistic agent collaboration settings?

Q2: How do supervised fine-tuning and RLHF independently influence this vulnerability?

Q3: Can additional causal experiments establish whether feature suppression is the primary mechanism underlying sycophancy?

Q4: How effective is structured dissent when the dissenting agent provides an incorrect answer or shares correlated errors with other agents?

4 comments
AuthorsOpenReview ↗

Response to Reviewer NjjA

We thank the reviewer for separating the mechanistic and deployment concerns cleanly. We address all four questions and the three weaknesses.

Q4: "How effective is structured dissent when the dissenting agent provides an incorrect answer or shares correlated errors with other agents?"

We take this first because it is also Weakness 3. We added three dissenters that are never given the correct answer (Table R1 in our response to Reviewer DUNd; Appendix C.4, Table 9). A dissenter carrying an independently queried model's answer matches the correct-answer dissenter in every framing (84.00 vs 85.25% correct, user-role). A dissenter naming a different wrong answer is followed on only 3.50–5.50% of questions and still raises accuracy from 20.75 to 79.50% under user-role framing.

On correlated errors specifically: drawing the independent dissenter's answer from the same model as one of the two wrong-arguing agents gives the same result (84.50, 66.00 and 41.50% correct across the three framings). The rescue does not depend on the dissenter being an independent model, only on its answer being formed without seeing the consensus.

Q2: "How do supervised fine-tuning and RLHF independently influence this vulnerability?"

The base-versus-Instruct comparison contrasts pretraining with the complete post-training pipeline, because none of the four families releases an intermediate SFT-only checkpoint. What it does establish is that the vulnerability is already present before any post-training, and that the complete pipeline reduces it on average (base ≥ Instruct in 10 of 12 matched cells, mean +15.8 pp, positive family-level mean in all four families). Attributing the remaining difference between SFT and preference optimization would require a family that publishes both checkpoints, which we flag as future work in Appendix H.

Q3: "Can additional causal experiments establish whether feature suppression is the primary mechanism underlying sycophancy?"

Three results bear on this directly, and together they localize the causal weight to one direction.

Table R4: Directional SAE clamping at L19 (named peer jury, 50-question subset, ∆ vs reconstruction-only baseline).

Strategy∆P(wrong)∆P(correct)
Rising-only clamp to clean mean−2.3 pp+0.3 pp
Falling-only clamp to clean mean−15.6 pp+1.8 pp
Both-direction clamp to clean mean−21.9 pp+3.5 pp

The features pressure turns off carry essentially all of the causal weight; the features it turns on move P(wrong) by −2.3 pp. This is read from the sign of each feature's activation change, not from any labeling step, so it does not depend on the feature taxonomy. It is corroborated without any SAE by the difference-in-means direction at L25 (−32.5 pp on P(wrong_target)), and on the pretrained base by a DIM sweep that produces −14.7 and −15.1 pp at L17 and L18 and −1.6 to +2.1 pp at L20–L25, landing on exactly the two layer-local loci the patching identifies.

We have accordingly changed the claim from "pressure suppresses clean-reasoning features rather than activating a new sycophancy circuit" to "pressure acts mainly by suppressing the correct answer rather than by writing in the wrong one," and we now state the −2.3 vs −15.6 pp result as a bound on an additive component rather than its exclusion (§5.5). The new dissenter result is the behavioral counterpart: the correct answer is still present under pressure and is recovered once the consensus is broken.

Q1: "Does the identified mid-layer vulnerability generalize to multi-round agent debates and open-ended reasoning tasks?"

Multi-round debate we do not test, and Appendix H now says this is a scope boundary rather than a robustness claim. Our reading is that single-step injection is the worst case: it presents unanimous, persuasive, pre-committed wrong arguments with no opportunity for the subject to shift the exchange, which real debates rarely achieve. We offer it as a baseline against which iterative effects can be measured.

Within the setting, the effect is not tied to one configuration: it replicates across three MMLU domains (Appendix E.4, E.5), across jury sizes N ∈ {4, 5, 6} (Appendix C.5), across four model families and on Mixtral-8x7B (Appendix G.3).

On open-ended answers we now have direct evidence. With no forced-choice readout, the model writing a complete response, the named peer jury moves 70.50% of final answers to its wrong target against 0.00% unpressured (Appendix C.3, Table 7). The format of the answer is not what makes the attack work.

AuthorsOpenReview ↗

Response to Reviewer NjjA cont.

Weakness 2: "the interventions only partially restore correct-answer probabilities. This limits the strength of the causal conclusions."

We would offer the opposite reading. Partial restoration is what the suppression account predicts, and full restoration is what would have undermined it.

If pressure worked by writing the jury's answer into the representation, removing that contribution would hand the probability mass back to the correct answer. It does not. Clamping drops P(wrong_target) by 21.9 pp while P(correct) rises only 3.5 pp; the DIM direction drops it 32.5 pp while P(correct) rises 10.5 pp; on base, −15.1 pp against +9.9 pp. In every case the freed mass goes to the other wrong answers. That is suppression plus an argmax among whatever remains, not the reversal of a substitution. A mechanism that merely added a wrong-answer signal would not produce this signature.

What the partial restoration does limit is the claim that suppression is the whole story, and we have sharpened the wording accordingly in the abstract, introduction and §5.5.

Weakness 1: "It focus on single-step injections and multiple-choice questions."

On the multiple-choice half we now have direct evidence that the format is not what makes the attack work: in freely written answers the named peer jury still moves 70.50% of final answers to its wrong target (Appendix C.3, Table 7). On single-step injection, see Q1 above; Appendix H states it as a scope boundary.

Thanks for your reply! My several main concerns have addressed, and I will improve my final scores.

AuthorsOpenReview ↗

We thank Reviewer NjjA for confirming

We thank the reviewer for the positive engagement and for confirming that the response addressed the main concerns. We will carry the new dissenter conditions, the flip-rate validation and the written-answer defense measurements into the final version.

Reviewer_DUNdOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper investigates how incorrect peer responses lead LLMs to abandon initially correct answers. The authors compare Base and Instruct models across four model families, examining the effects of message roles and consensus strength. Through activation patching, component decomposition, SAE interventions, and steering, they identify an attention-dominated mid-layer region and attribute the effect to the suppression of clean-reasoning features. They further show that a correct dissenting response substantially reduces susceptibility. Experiments primarily use 400 high-confidence MMLU humanities questions.

Critical strengths
  1. The paper evaluates multiple model families and settings, with additional cross-domain, cross-model, and robustness experiments, providing broad empirical coverage. However, the relationship between the primary metric and actual behavioral changes requires stronger validation.

  2. Activation patching, attention/MLP decomposition, and feature-level interventions provide complementary evidence for the identified mid-layer effect, strengthening the proposed mechanistic explanation.

Critical weaknesses
  1. All tested dissenters endorse the ground-truth answer, even when providing only a bare assertion. This makes it difficult to distinguish the effect of disagreement itself from that of supplying the correct answer. More importantly, the proposed defense assumes access to the correct answer, which is generally unavailable in real-world settings. Evaluating independently sampled dissenters, including incorrect ones, would better establish whether dissent provides a reliable practical benefit.

  2. The LDA-based yield measures representational proximity to answer centroids rather than actual answer flips. Its sensitivity to calibration data, as shown in Appendix F.1, further raises concerns about the reliability of reported rates and whether defense improvements measured using this metric reflect genuine behavioral changes.

Questions for authors
  1. Does the rescue persist with an incorrect dissenter, uncertainty-only dissent, or an independently sampled model whose answer is not selected using ground truth?
3 comments
AuthorsOpenReview ↗

Response to Reviewer DUNd

We thank the reviewer for a precise question. We ran all three variants it names, and address both weaknesses below.

Question: "Does the rescue persist with an incorrect dissenter, uncertainty-only dissent, or an independently sampled model whose answer is not selected using ground truth?"

Yes to all three. We added three dissenters that are not given the correct answer, all in the same minimal format and the same 2v1 layout as the existing correct-answer dissenter.

Table R1: Where the answer goes when the dissenter is not given the correct answer (2v1, suffixed, n = 400, %). Each cell: top-1 correct / top-1 is the jury's target / top-1 is the dissenter's own wrong answer.

DissenterUser-roleAssistant-roleTool-role
None (3v0)20.75 / 78.75 / —2.00 / 98.00 / —0.25 / 99.75 / —
Correct answer (minimal)85.25 / 14.25 / —68.25 / 31.75 / —42.25 / 57.75 / —
Different wrong answer79.50 / 16.75 / 3.5039.00 / 55.50 / 5.5014.25 / 82.00 / 3.75
No answer named72.50 / 27.00 / —43.75 / 56.25 / —17.25 / 82.75 / —
Independent model84.00 / 15.00 / 0.7565.25 / 33.75 / 1.0041.75 / 57.75 / 0.50

The independent dissenter carries whatever answer a jury model gives to the bare question with no jury shown. It is correct on 377 of 400 questions, names another wrong option on 19, and agrees with the jury on 4, where the closing line becomes "All three models agree."

Three results follow. First, no answer key is needed: the independent dissenter matches the correct-answer dissenter in every framing (84.00 vs 85.25, 65.25 vs 68.25, 41.75 vs 42.25). Second, the rescue is not copying: a dissenter naming a different wrong answer is followed on only 3.50–5.50% of questions, and accuracy still rises from 20.75 to 79.50% under user-role framing, because breaking the consensus releases the correct answer rather than installing the dissenter's. Third, even content-free dissent works: "I disagree with the other models," with no answer named at all, raises user-role accuracy from 20.75 to 72.50%.

Under tool-role framing a wrong dissenter recovers less (0.25 to 14.25%), consistent with tool returns being the highest-trust channel. We report this rather than averaging it away. Drawing the independent dissenter's answer from the same model as one of the wrong agents gives the same result (84.50, 66.00 and 41.50% correct), addressing correlated errors.

This is Appendix C.4 (Table 9) and the new rows of Table 8 in the revision, with the headline in §5.6.

Weakness 1: "The LDA-based yield measures representational proximity to answer centroids rather than actual answer flips."

We now report both. The behavioral flip rate is the share of questions whose forced-choice top-1 answer is correct on the clean prompt and wrong under pressure: no LDA, no probes, no calibration set. We re-ran all 16 main conditions and the N=4 sweep.

Table R2: Yield versus behavioral flip rate, selected conditions (suffixed, n = 400, %). Full 26-condition table in Appendix F.1, Table 18.

ConditionPublished yieldRe-run yieldFlip rate
Tool-role jury (strong)98.0098.0099.75
Assistant-role jury (strong)97.7597.7598.00
Named peer jury (strong)75.7575.7579.25
Direct user assertion44.0044.2550.25
Named peer jury (weak)30.2530.2533.00
Anon. perspectives (weak)10.2510.2510.50
User-role sweep, 3v112.7512.0013.50
User-role sweep, 4v080.2579.2583.75
Assistant-role sweep, 3v160.2561.2562.25

Every published yield reproduces within 1 pp. Across all 26 conditions the flip rate lies within −0.75 to +6.00 pp of yield (mean +1.6 pp), and an independent second re-run gives the per-question picture: the yield call and the flip agree on 85.00–100.00% of individual questions in every condition, 95.47% on average. Flip is slightly the higher of the two because it counts a move to any wrong option while yield counts only a move toward the jury's target, so the reported yields are the conservative measure.

The gradients the paper rests on are therefore changes in the answers the model gives. The same holds with no forced-choice readout at all: in freely written responses the named peer jury moves 70.50% of final answers to its target, against 0.00% without pressure (Appendix C.3, Table 7).

Weakness 1 (continued): "Its sensitivity to calibration data, as shown in Appendix F.1, further raises concerns about … whether defense improvements measured using this metric reflect genuine behavioral changes."

This is well taken, and we measured the five defenses on the written answer, which needs no calibration at all.

AuthorsOpenReview ↗

Response to Reviewer DUNd cont.

Table R3: Defenses measured on freely written answers (named peer jury, no priming string, n = 400, %). "No answer": the response names no option, mostly an explicit rejection of the jury's claim.

ConditionCorrectJury's answerOther wrongNo answer
No pressure95.500.000.004.50
Named peer jury, no defense25.2570.500.503.75
skeptical73.257.002.7517.00
verify63.509.751.7525.00
cot53.5026.251.5018.75
ignore45.0038.750.7515.50
priors23.5068.000.258.25

Our submission reported that the strongest prompt loses essentially all effect without the priming string. On the answers themselves it does protect (73.25 vs 25.25% correct), and the revision reports that measurement in place of the earlier estimate. What survives, and is what the paper's argument actually needs, is the coverage failure: the same defense drops from −65.00 pp on the attack it names to −45.50 on a count-matched anonymous claim, −28.00 on a bare assertion inside the jury, and −13.75 with no jury at all (Appendix C.3, Table 6). A defender must anticipate every phrasing; the dissenter needs none.

Weakness 2: "All tested dissenters endorse the ground-truth answer, even when providing only a bare assertion."

Table R1 is our answer. We have also rewritten the corresponding Limitations entry from "The dissenter supplies the correct answer" to "How often the dissenter is right," which states the real boundary: our independent dissenters are right on 377 of 400 because the pool contains questions these models know well, and on harder questions they would be wrong more often, with assistant- and tool-role framing recovering less from a wrong dissenter (39.00 and 14.25% correct, against 68.25 and 42.25 for a correct one).

On presentation. We rewrote the first sentence of the abstract to state the phenomenon before any result, added a plain-language definition of yield in §4.1, and renamed §5.5 to the claim it supports. We would welcome any specific passage the reviewer found hardest to follow.

AuthorsOpenReview ↗

Following up on our response

We are writing to kindly remind you that we posted our response three days ago, with new experiments addressing both weaknesses and all three parts of your question. If you have any additional feedback, concerns, or questions regarding our response, we would greatly appreciate hearing from you.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Response to All Reviewers

We thank the reviewers for three careful and largely convergent reviews. Two concerns were raised independently by all three: that every dissenter in our mitigation is given the correct answer, and that yield is a representational measure whose relationship to the model's actual answers was not established. We ran new experiments for both. A revised PDF is uploaded.

The rescue does not require an answer key. We added three dissenters that are never shown the correct answer (Appendix C.4, Table 9). A dissenter carrying the answer of an independently queried model matches the correct-answer dissenter in every framing: 84.00 vs 85.25% correct under user-role framing, 65.25 vs 68.25 assistant-role, 41.75 vs 42.25 tool-role.

The rescue is not the dissenter's answer being copied. A dissenter naming a different wrong answer is followed on only 3.50–5.50% of questions. When the model leaves the jury's answer it returns to its own correct one, raising accuracy from 20.75 to 79.50% under user-role framing. This is the behavioral counterpart of the suppression mechanism: the correct answer survives under pressure and is released once the consensus is broken.

Yield tracks the answers the model gives. Table 1 now reports the behavioral flip rate beside yield. Across 26 re-run conditions every published yield reproduces within 1 pp, the flip rate lies within −0.75 to +6.00 pp of yield, and the two agree on 95.47% of individual questions (Appendix F.1, Table 18). Without any forced-choice readout, in freely written answers, the named peer jury moves 70.50% of final answers to its wrong target against 0.00% unpressured (Appendix C.3, Table 7).

A sharper measurement of the defenses. Re-running the defenses as the reviewers asked changed a result. Our submission reported that the strongest system prompt loses its effect without the priming string. On freely written answers it does protect (73.25 vs 25.25% correct). The earlier comparison used a readout that is near chance for defended prompts, because these prompts make the model defer its decision to the reasoning it writes. The defense's real weakness is coverage, not the readout: its effect falls from −65 pp to −13.75 pp when the same wrong answer is rephrased (Appendix C.3, Tables 6–7). Discussion, §5.6 and Appendix G.1 are revised accordingly.

Two notes for reading the revision: three tables were added (Tables 7, 9 and 18), so later table numbers have shifted; and in response to the presentation scores we rewrote the abstract’s opening and added a plain-language definition of yield in §4.1.

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

Related papers

Discussion (0)

Sign in to comment.