- Rating
- 3(+1)
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 3
- Soundness
- 3
Read review
Paper summary
This paper looks at how we measure sycophancy and belief revision in two-turn LLM evaluations. Usually, we ask a model for an initial judgment in turn 1, push back in turn 2, and compare that against an empty control where the pushback is removed. The authors point out two flaws with this setup: it lumps together revising because of a good argument (which we want) and revising because of who is speaking (which we don't), and it assumes the empty control is a true zero.
To separate these effects, the authors introduce Matched-Presence Decomposition (MPD). Instead of deleting elements in the control, MPD swaps each manipulated piece of turn 2 (the argument text, the user's reputation record, or the user's stated position/shares) with the corresponding element from a different episode in the pool, and adds a bare reask_only condition that drops even the framing sentence. They also define six audits to check whether the controls and metrics are actually valid before reading the coefficients.
Testing this on ~1000 comment episodes across 86 Manifold prediction markets and seven model families, they find three main things. First, simply asking the model a second time (reask_only) causes it to change its probability estimate in 65.6% to 86.6% of episodes across the four open-weight models, and greedy decoding barely reduces this. Second, when scored on whether the revision moves in the direction of the pushback (toward), argument content has a large positive effect across all seven models stated position has a small effect in five of seven , and the user's reputation record has zero effect across the board. Third, when testing prompt mitigations to reduce re-ask drift, telling the model it can repeat its previous answer (license_to_repeat) gives the biggest drop in drift (down to 4.2% on Qwen3-8B), but does so by making the model ignore real evidence (retaining only 43.3% of the evidence effect). Asking the model to first output a boolean for whether turn 2 contains new information (evidence_gate) cuts drift by 19 to 47 points while keeping the evidence response intact.
Critical strengths
-
I really appreciate the methodological discipline in this paper. The core point of Audit 1 is simple and important: an empty control (
zero_info) and a presence-matched donor control (argument_shuffled) do not behave the same way. Because showing an unrelated comment from another market actually makes the model revise less often than showing an empty prompt (and on Gemma-4-12B, the empty control triggers a 23.4% refusal rate because the prompt says another trader responded without showing the comment), standard evaluations that subtract the empty control end up understating the true effect of the argument by 2 to 17 points. -
I also respect how the authors handled their negative results and pre-registered checks. When they tried to transfer MPD to 50,339 ICLR peer review episodes, Audit 2 caught that the model's initial review score (6.46) sits above the human mean (5.52), so 78% of both real and donor reviewer scores pull the model downward (). Rather than trying to salvage those numbers, they voided the directional coefficients from that domain and used it as an example of what the audit catches. Same goes for dropping the "ungrounded revision" metric when Audit 4 showed the LLM judge agreed 100% of the time with a trivial string check.
-
the mitigation analysis in Section 5.3 makes a great practical point. If you only evaluate anti-sycophancy interventions on how much they reduce flipping, you end up picking prompts that make the model stubborn to everything, including good evidence. Requiring both drift reduction and evidence retention is the right way to evaluate these fixes.
Critical weaknesses
- My biggest issue with the paper is that the headline claim about "bare re-ask drift" (Figure 1 and Section 5.2) is confounded by how the prompts were built for reply comments. In the main text,
reask_onlyis described as asking the model a second time with "nothing attached." However, if you read Appendices B, G, and N carefully, 73.6% of the Manifold stimuli (1,014 out of 1,378) are replies to an earlier comment, and in all of those episodes, turn 2 prepends that parent comment (Earlier in the comment thread, another trader wrote: ...) even inreask_only. The model never saw that parent comment in turn 1. So for nearly three-quarters of the data,reask_onlyis not a bare re-ask at all; it is feeding the model a brand new comment from the thread and asking for an updated probability.
You can see how much this matters in Appendix G. When the authors run reask_verbatim (which repeats the turn 1 question exactly and drops both the word "updated" and the parent comment), Qwen3-8B's revision rate drops from 0.784 all the way down to 0.179, whereas reask_neutral (which only drops the word "updated" and keeps the parent comment) stays at 0.623. That means roughly 44 percentage points of Qwen3-8B's drift comes from showing the parent comment in turn 2, not from asking the question twice. Since 26.4% of the dataset (364 episodes) has no parent comment, the authors already have the clean data to separate this. Those 364 episodes should be reported separately in the main text, and Figure 1 needs to be caveated or updated so readers aren't misled into thinking a pure re-ask causes an 80% flip rate when most of those prompts injected new thread context.
-
My second concern is about how the presence-matched control (
argument_shuffled) interacts with the unconditionaltowardmetric. Putting a comment about one prediction market into a prompt about a completely different prediction market is not a neutral placebo; it is a weird, off-topic non-sequitur. As Appendix G shows, the less topically related the donor comment is, the more it suppresses any revision at all (slope of +0.645 across similarity quartiles). Becausetowardis defined unconditionally as , any episode where is scored as 0. When an off-topic donor comment causes the model to freeze and keep its old probability (),towarddrops mechanically, even if the donor comment's direction is random. So part of ontowardis just measuring the fact that models freeze up when shown off-topic text from another market, rather than measuring persuasion by the real argument. Reportingtowardconditional on (which is mentioned briefly in one sentence in Section 5.2) alongside the unconditional metric would make this much cleaner. -
the primary
towardresults rely on only 32.1% of the dataset (443 of 1,378 stimuli) where two LLM labellers (DeepSeek-V4-Pro and GPT-5) agreed on the comment's direction, and the authors note in Appendix K that they cancelled the human validation check on these labels. Throwing out 68% of the comments (48% as directionless, 20% for disagreement) selects for the most strongly opinionated comments in the pool. While the post-hoc Brier score analysis against market resolution in Appendix L is a great label-free check on 1,158 stimuli, it also reveals a problem withevidence_gatethat gets brushed past: on Qwen2.5-7B,evidence_gatehas the highest directional retention in Table 3 (1.377), yet in Table 5 its Brier score gain (real - donor) drops from -1.53 (significant) to -0.64 (not significant). Also, in Section 5.3, models underevidence_gatestill claim"new_information": trueon 32% to 53% of episodes where no comment was shown at all. Soevidence_gateis still hallucinating new information nearly half the time on empty re-asks. -
I am not convinced by how the reputation channel () and its linear probe ( in Appendix O) are set up. In the prompt (Appendix N), reputation is a multi-field metadata string (account year, lifetime mana deposited, number of traders, trustworthy flag, comment likes). When you swap a commenter's metadata with a random donor's metadata from another episode, how is the "direction" of that swap defined for
towardor for the linear probe? Does a donor have higher or lower reputation than the real commenter? If you don't condition on the sign of the reputation difference between the real user and the donor user, then sometimes the real user has higher reputation and sometimes the donor does, which will average out to zero ontowardeven if the model cares about reputation. The paper never explains how direction is assigned to or what binary label the linear probe is trained to predict.
Concers on writing and presentation
the paper is written in a very compressed, telegraphic style where key definitions are tucked into the appendix. Section 3.3 summarizes Audits 3 through 6 in four sentences without giving their formulas, and the nine experimental conditions are never laid out in a clean table in the main text. Also, in Section 4.2 and Figure 3, the authors repeatedly state that coefficient point estimates cannot be compared between (open-weight) and (API) runs because per-stimulus noise differs. Changing the number of Monte Carlo draws per stimulus changes the standard error of the mean, not the expected value of the sample mean contrast. Unless temperature changed (as in Claude-Sonnet-5), the point estimates are estimating the exact same quantity.
Questions for authors
-
On the 26.4% of Manifold episodes (364 stimuli) that are top-level comments with no parent comment injected in turn 2, what are the revision rates for
reask_only,reask_neutral, andreask_verbatimacross the four open-weight models? -
In Table 1, how is scored on
toward? Specifically, is still just the direction of the comment text? If so, why would replacing a random donor's reputation record with the real commenter's reputation record increase alignment with unless the real commenter systematically has higher reputation than the donor? Shouldn't the contrast be conditioned on whether the real commenter's reputation exceeds the donor's? -
What exact binary target or contrast is the linear probe for and trained on in Appendix O?
-
Why does
evidence_gatereport"new_information": truein 32% to 53% of episodes where no comment was shown, and is that rate coming primarily from the 73.6% of episodes where the parent comment appears in turn 2? -
Why does Qwen2.5-7B show 137.7% retention on
towardunderevidence_gate(Table 3), but lose more than half of its Brier score improvement (real - donordropping from -1.53 to -0.64 in Table 5)?
4 comments
Response to Reviewer zSYz (2/2)
W4 / Q2 / Q3. Direction of the reputation channel; split by who outranks whom; probe target.
toward is always scored against the comment's direction; the reputation block has none. The reputation contrast asks whether the same argument is followed more closely under a different author. We split it by lifetime deposits, which every record shows; the real commenter's deposits are higher on 50% of stimuli:
| β_R on toward | real > donor | real < donor |
|---|---|---|
| Qwen3-8B | +0.012 [−0.005, +0.032] | +0.013 [−0.003, +0.033] |
| Qwen2.5-7B | +0.000 [−0.024, +0.022] | +0.009 [−0.013, +0.033] |
| Gemma-4-12B | −0.022 [−0.043, −0.004] | −0.009 [−0.030, +0.015] |
| GPT-OSS-20B | −0.005 [−0.040, +0.037] | +0.019 [−0.030, +0.066] |
| DeepSeek / Claude / GPT-5.5 | −0.039 / +0.029 / +0.010 | −0.013 / +0.021 / +0.004 |
The null holds in both halves for every family. The platform's "trustworthy" badge differs between the real commenter and the donor on 400 stimuli; split that way, each cell of the labelled subset holds only 36–136 pairs, and the contrasts scatter on both sides of 0 across families (App. I). Probe target: the real reputation block of a stimulus is labelled 1 and its donor-swapped block 0; the train/test split is by stimulus, so both rows of a pair fall on the same side (App. P).
Official comment by Reviewer zSYz
I thank the authors for their detailed response and extended experiments. Raising my score accordingly and also posting few followups:
- In your response, you noted that you were hand checking a sample of the direction labels against the two LLM labellers and would post the agreement rates once complete. Could you share those human agreement numbers before the discussion period closes?
Update of human agreement numbers
We sincerely thank you for your thoughtful and constructive feedback. We especially appreciate your willingness to raise your score in light of our response and additional experiments.
We have now completed the human annotation study, with the results reported in Appendix L. Specifically:
- Human annotation: Two annotators independently labelled all 1,378 comments using blind sheets and the same annotation guidelines provided to the model labellers.
- Inter-annotator agreement: The two human annotators achieved 83.3% agreement on the three-way labels (Cohen's κ = 0.74), compared with 80.0% (κ = 0.66) between the two model labellers. When both human annotators assigned a direction, their directional agreement was 96.1%.
- Human–model agreement: On the 443 stimuli in the primary subset, Human Annotators 1 and 2 achieved 88.5% and 88.7% agreement with the two LLM labellers' consensus labels, respectively.
- Directional consistency: Among the 396 stimuli for which both human annotators and both model labellers assigned a direction, both human annotators matched the LLM labellers' consensus direction in 95.2% of cases.
- Conservative subset selection: The human annotators agreed on a directional label for 644 comments, but 443 were selected by the model labellers. This suggests that the model labellers were more conservative in assigning directional labels.
- Example of ambiguity: A comment questioning the authenticity of a Netanyahu livestream was labelled NO by one annotator, who interpreted it as genuine skepticism, and YES by the other, who read it as sarcasm mocking such skepticism. This illustrates how different interpretations of sarcasm can lead to disagreement.
- The original 443-stimulus subset remains unchanged.
We hope these additional results address your concern, and we sincerely appreciate your valuable feedback and continued engagement.
Response to Reviewer zSYz (1/2)
Thank you for the detailed and thoughtful reviews.
Writing. We revise accordingly: the conditions table is added (Table 2, App. B); Audits 3–6 are defined in full in §3.3, with the realized values in App. N; the claim that k = 1 and k = 2 estimates are not comparable is removed from §4.2, and the Table 3 note, the Fig. 3 caption, App. D, App. M and Limitations now state that k changes precision only.
W1 / Q1. Parent comment; the three re-asks on the 364 no-parent stimuli; Fig. 1.
The parent accounts for 6–8 points of re-ask drift, not the 44 points inferred in the review. The verbatim fall is a wording effect, and the claim survives removing the parent. The four-placement table (full analysis in CQ1), then the requested numbers on the 364 no-parent stimuli:
The verbatim fall occurs where no parent exists, so the 44-point attribution does not hold. The fall is wording, and it is specific to Qwen3-8B. Fig. 1 now hatches the parent's share and gives both numbers.
W2. Donor comments are not neutral; counting "no change" as 0 mixes suppression into toward.
We agree that the donor comment suppresses revision. Conditional on revising, the donor direction is at chance, and the content contrast survives. The paper reports the suppression gradient. Conditional on revising, the donor direction is at chance (0.50–0.57) in all seven families, and the real comment's is 0.76–0.87; the content contrast is +0.24 to +0.36 on that reading (Appendix I, added in the revision):
W3 / Q4 / Q5. The 443 subset; the Qwen2.5-7B Brier advantage halved while direction-scored retention is 1.38; the gate reports new information on 32–53% of empty re-asks.
The 443 subset is an acknowledged limitation. Both retentions are intact: the Brier advantage shrank because the gate removed the donor comment's harm, not the response to evidence. The 32–53% is the parent block.
Labels (W3): we agree; the submission states this limitation in Limitations (App. A) and in App. L. We are hand-checking the direction labels against the model labels now and will post the agreement rates as soon as this is done.
However, the limitation does not reach the paper's conclusions. The two labellers agree on the three-way label (YES, NO, no direction) in 80.0% of the 1,378 stimuli (Cohen's κ = 0.66). Of the 227 disagreements, 179 have one labeller reading a direction where the other reads none, and 48 have opposite directions. Every result that needs no direction (revision rate, |Δ|, re-ask drift, each mitigation's drift reduction) uses all 1,378 stimuli. The label-free Brier results on 1,158 resolved stimuli agree in sign with the direction-scored results in all seven families. If the agreed subset over-represents clearly argued comments, that can make the evidence effect look larger than it is; it cannot create a reputation effect of zero where there is one: the reputation contrast compares the same argument under a different author, and selecting on how clearly the argument is made does not touch that comparison. Brier (W3, Q5): the paired test is in CQ2; the outcome-scored retention ratio fell because its baseline value included the donor comment's harm:
Gate (W3, Q4): it reports "new information" on 39.5–72.0% of empty re-asks with a parent and on 0.1–10.9% without, so the gate sees text that is new; §5.3 is corrected.