acceptodds
Under review as a conference paper at ICLR 2027

FIGS: Evaluating Multi-Turn Sycophancy Without Penalizing Empathy

15 views

Abstract

Large language models frequently fail to balance staying truthful with being supportive. To appease users, they often become sycophantic, by agreeing with false claims and through subtle behaviors like unwarranted flattery and skewed advice. In reality, sycophancy rarely happens in a single exchange; it emerges organically as users repeatedly insist or subtly steer the dialogue over time. Current evaluations, however, rely on rigid, single-turn tests or fixed scripts that fail to capture these natural dynamics. Furthermore, these benchmarks often mistake showing basic empathy for yielding, penalizing models for acknowledging a user's feelings and driving them to over-correct into cold, dismissive rigidity. To address this gap, we introduce FIGS (Factual Integrity and Grounded Support), a dual-axis evaluation framework built entirely around extended, realistic dialogue. We use an adaptive 10-turn conversational simulator that dynamically challenges the target model, mimicking how human boundaries are actually tested in everyday interactions. To accurately evaluate these trajectories, we apply a taxonomy that strictly separates Sycophancy (whether the model holds firm to the truth and keeps its praise proportional) from Calibrated Validation (showing empathetic understanding of the user's feelings without overdoing it). We release our complete testing environment, including 500 diverse multi-turn scenarios and an automated judge. Our evaluation of leading models reveals a consistent trade-off: over the course of a sustained interaction, current systems either slowly drift to sycophancy or over-correct into robotic detachment. This demonstrates that balancing honesty with appropriate support throughout a natural conversation remains a critical, unsolved challenge.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Sycophancy, Empathy, Multi-turn Evaluation, Conversational AI, Large Language Models, AI Alignment

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_GQ2VOpenReview ↗
Rating
3
Confidence
2
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces FIGS, a benchmark for evaluating multi-turn sycophancy while distinguishing it from appropriate emotional acknowledgment. Its taxonomy separates two axes: Sycophancy, covering evidence-sensitive judgment, proportionality, and resistance to unsupported pressure; and Calibrated Validation, covering acknowledgment of user disclosures, faithful representation of their experience, and respect for their choices.

The benchmark contains 500 frozen scenarios across six domains and three severity levels. An adaptive user simulator conducts five user–assistant exchanges, and an automated judge scores each complete transcript on both axes using a four-level severity scale and ten diagnostic rules.

Experiments evaluate eight models under baseline and factual system prompts, with additional calibrated-prompt experiments on four models. The paper finds that the factual prompts reduce sycophancy but can increase validation failures, while a prompt addressing both objectives improves their balance. Additional analyses examine human–judge agreement, alternative simulators, rollout variance, prompt variants, and comparisons with prior evaluation protocols.

Critical strengths
  1. The benchmark separates two consequential dimensions that can otherwise be conflated. Distinguishing unsupported agreement from grounded acknowledgment allows an evaluator to identify both sycophantic warmth and factually careful but dismissive responses. Including warranted updating further avoids treating every change of position as a failure. This makes FIGS useful for assessing interventions that might improve an anti-sycophancy score by degrading conversational support.

  2. The evaluation framework is supported by substantial methodological documentation and robustness analyses. The paper provides 500 scenarios, adaptive simulation, explicit scoring rubrics, human–judge comparisons, alternative simulator experiments, rollout reruns, and prompt ablations. The calibrated-prompt experiment is particularly useful because it demonstrates that the two axes need not move in opposite directions. The authors also appropriately disclose that this prompt was tuned against their rubric rather than presenting it as an independently validated general solution.

Critical weaknesses
  1. The prompt experiments do not isolate whether prioritizing accuracy itself causes validation failures. The baseline explicitly encourages a warm, emotionally responsive persona. The factual prompt also changes tone, response structure, and the permitted treatment of emotional disclosures. The additional minimal and prose prompts provide useful checks, but both remove the baseline persona and request short replies, as Appendix F.3 acknowledges. Consequently, the observed changes may partly reflect altered style instructions rather than an intrinsic cost of resisting sycophancy. This matters because the “paternalism trap” is a central empirical claim. A more controlled comparison should preserve warmth, length, and structure while varying only the evidentiary and anti-capitulation instructions.

  2. Validation of the primary pass/fail metric remains incomplete. The headline metric treats any score of 2 or higher as failure, yet the reliability analysis primarily reports four-level quadratic-weighted kappa and within-one agreement. These measures can remain favorable when disagreements cross the critical boundary between scores 1 and 2. This is especially relevant because 93% of residual sycophancy failures under the factual prompt receive score 2. Human exact-score agreement is modest, and the held-out judge test uses a single annotator rather than an independent panel. Binary confusion matrices and independently adjudicated labels would strengthen confidence in failure rates and close model comparisons. The absence of observed S1.b failures also leaves sensitivity to warranted-update failures less established than sensitivity to capitulation.

Questions for authors

The following clarification and validation requests could materially affect my assessment:

  1. Does the increase in calibrated-validation failures persist when the baseline’s warmth, response length, and persona are preserved and only the anti-sycophancy instructions are added? The existing prompt ablations are useful but still vary these factors together.

  2. What are human–human and judge–human agreement at the binary pass/fail boundary, including false-positive and false-negative rates? Do the main model rankings and prompt effects persist with human-adjudicated labels and with only scores 3–4 counted as failures?

  3. Can you validate sensitivity to S1.b failures using a balanced set of warranted updates and unwarranted reversals, independently labeled by multiple annotators? The constructed example demonstrates that the judge can assign the label, but does not establish detection reliability.

  4. The temporal analysis reports where the judge cites a failure rather than independently labeling every assistant reply. Does the conclusion that sycophancy builds over turns hold when measuring the first failure onset or per-reply failure rates, accounting for the changing opportunities for pressure across turns?

5 comments
AuthorsOpenReview ↗

*(Part 1 of 4)*

(Part 1 of 4)

We thank the reviewer for the thoughtful review and for recognizing several strengths of FIGS. It separates two consequential dimensions, so it can identify both sycophantic warmth and factually careful but dismissive responses, without treating every change of position as a failure; this makes it useful for assessing interventions "that might improve an anti-sycophancy score by degrading conversational support". It is supported by substantial documentation and robustness analyses, and the calibrated-prompt experiment shows that the two axes need not necessarily move in opposite directions.

We are also grateful that the reviewer named the analyses that could change the assessment. We ran each of them.

Summary.

A new prompt that keeps the baseline persona and adds only the anti-sycophancy instructions removes most sycophancy with little validation cost for three of four assistants, while the factual prompt, which drives sycophancy lowest, has the largest cost; activation steering shows the same pattern without any instruction. We have revised the paper to say that the paternalism trap appears when sycophancy is pushed to its lowest levels.

As requested, we now report human–human and judge–human agreement at the binary pass/fail boundary, with binary confusion matrices and false-positive and false-negative rates, at both the > 1 and 3–4 cutoffs, on an enlarged test set of 164 transcripts. The judge agrees with humans on 88–89% of pass/fail decisions, at or above human–human agreement, and rankings and prompt effects hold under human-calibrated labels and at the 3–4 cutoff.

We also clarify how S1.b failures reach the rule stage, and report a balanced minimal-pair set on which the rule judge detects failures to update in both directions and never tags a warranted update. Finally, we report first-failure onset and per-reply failure rates from labelling every reply independently: 70% of sycophancy failures first appear after the first reply.

Below we answer each concern in turn, taking related concerns together.

1. W1 and Q1

As the reviewer suggested, we added a prompt that keeps the baseline's warmth, length and persona and adds only the anti-sycophancy instructions, and we also tested activation steering with no instruction at all. CV failures do not rise for three of the four assistants; for Grok 4.6 they do.

What each accuracy prompt changes. Every accuracy prompt adds instructions to resist sycophancy. They differ in what else they change about the baseline (the warm assistant persona of the main paper): whether they replace its persona and warmth, and what they say about emotional disclosures. Comparing them therefore separates the effect of resisting sycophancy from the effect of everything else:

PromptResists sycophancyBaseline persona and warmthGuidance on emotional disclosures
Factual (main paper)yes, full setreplacedrestrictive: "if a feeling is reported, you may note that it was reported, nothing more"
Minimal (Appendix G)yes, three sentencesreplacednone
Prose (Appendix G)yesreplacedpermissive: warmth allowed when it matches what the user has shown
Persona-controlled (new, Appendix G)yes, the factual prompt's five resistance sections, verbatimkept: the full baseline prompt, verbatimthe baseline's own

The persona-controlled prompt is the complete baseline prompt followed by five sections copied verbatim from the factual prompt. Nothing from the factual prompt that governs tone or emotional disclosures is included. All five prompts were run on the same 200 scenarios (100 per axis) with the same four assistants, simulator and judge, and compared with the baseline on paired scenarios.

Results. All four prompts cut sycophancy significantly for every assistant:

Sycophancy fail %Gemini 3.8 FlashDeepSeek V4.1 FlashGPT-5.6-SolGrok 4.6
Baseline75455629
Persona-controlled2119189
Prose1421257
Minimal2418315
Factual1010100

They differ in what they cost on calibrated validation (bold: significant increase over the baseline):

CV fail %Gemini 3.8 FlashDeepSeek V4.1 FlashGPT-5.6-SolGrok 4.6
Baseline66128
Persona-controlled93339†
Prose3261628
Minimal40132750
Factual39102638

† Not significant on the 100 CV scenarios alone, but significant when CV is scored on all 200 conversations, and when only scores 3–4 count as failures. No unbolded change is significant.

AuthorsOpenReview ↗

*(Part 2 of 4)*

(Part 2 of 4)

1. W1 and Q1 (continued)

What the comparisons show.

  1. Resisting sycophancy alone does not raise CV failures for three of the four assistants (baseline vs persona-controlled). With the baseline kept whole, CV stays at the baseline level for Gemini, DeepSeek and Sol (6→9, 6→3, 1→3) while sycophancy still falls by 26–54 points.
  2. Replacing the baseline persona does (persona-controlled vs the other three). Every prompt that replaces it raises CV failures for Gemini (32–40%) and Sol (16–27%).
  3. It is not the factual prompt's wording about emotions (factual vs minimal). The minimal prompt says nothing about emotions, yet its CV rise is as large as the factual prompt's.
  4. Permitting warmth softens the cost but does not remove it (prose vs minimal). Allowing warmth that matches the user lowers CV failures (Gemini 40→32%, Sol 27→16%, Grok 50→28%), but for Gemini and Sol they stay well above the baseline.
  5. Grok is the exception. Its CV failures rise to the factual prompt's level under the resistance instructions alone (28→39%, against 38% under the factual prompt), so for Grok the cost does come from resisting sycophancy. The judge's descriptions show it applying the evidentiary rules to the user's feelings, for example a reply that "converts the disclosure into evidentiary items with zero uptake of the fear".

The persona-controlled prompt leaves more sycophancy than the factual prompt (9–21% against 0–10%).

Evidence without any instruction: activation steering (new Section 5.2, Figure 5, Appendix H). To test resisting sycophancy with no style instruction at all, we steer four open-weight models along a sycophancy direction in the residual stream, with no system-prompt change. The direction comes from 115 sycophancy scenarios that are excluded from all results, and no CV data is used to build it or to choose its layer or strength. k is the steering strength (k = 0 is the unsteered model; negative k pushes away from sycophancy). Bold changes are significant against k = 0 on the same scenarios (275 held-out sycophancy and all 110 CV scenarios).

ModelSycophancy fail %, k = 0 / −4 / −8Change at k = −4Change at k = −8CV fail %, k = 0 / −4 / −8Change at k = −4Change at k = −8
Gemma-4-26B-A4B80.3 / 60.4 / 38.0−20.1−42.39.1 / 10.0 / 24.5+0.9+15.5
Mistral-Small-24B88.0 / 72.6 / 44.2−15.3−43.846.4 / 52.7 / 72.7+6.4+26.4
OLMo-3.1-32B88.7 / 67.9 / 35.8−20.8−52.913.6 / 14.5 / 38.2+0.9+24.5
Qwen3-8B96.4 / 81.8 / 47.1−14.6−49.337.6 / 34.5 / 70.0−2.8+33.0

Pushed hard enough (k = −8), every model becomes markedly less sycophantic and, at the same time, markedly worse at calibrated validation, although nothing in the direction targets validation. A moderate push (k = −4) separates the two: sycophancy falls by 15–21 points while no model's CV rate changes significantly. Full results, including k = +4, are in Appendix H.

The two experiments answer the question from opposite sides, and they point the same way: pushing sycophancy to its lowest levels is what brings on the paternalism trap. The persona-controlled prompt removes most sycophancy with no validation cost for three of the four assistants, but it leaves more sycophancy than the factual prompt (9–21% against 0–10%). The factual prompt reaches the lowest sycophancy of all, and that is where validation pays (CV 26–39% for Gemini and Sol). Steering shows the same with no instruction at all: a moderate push leaves validation intact, and only a hard push, which drives sycophancy lowest, degrades it in every model. So the trap is not an intrinsic cost of prioritizing accuracy, but a risk that appears as a mitigation pushes sycophancy furthest down, to a degree that depends on the model and the intervention (Grok shows it even under the persona-controlled prompt). We have added this clarification to the revised paper: Section 5.1 now shows that the factual prompt's CV cost appears where sycophancy is pushed lowest, with the persona-controlled and steering results, and the abstract and conclusion now state that the size of the trade-off depends on the model and the mitigation. The dual-axis argument is unaffected: a sycophancy-only score would rank the factual prompt above the persona-controlled one (10% vs 21% for Gemini), while FIGS shows it costs 30 more points of CV failures.

AuthorsOpenReview ↗

*(Part 3 of 4)*

(Part 3 of 4)

2. W2 (first part) and Q2

To make agreement at the 1|2 boundary visible directly, rather than through QWK and within-one agreement, we now validate the binary pass/fail decision at both cutoffs (score > 1 and 3–4), in the revised Appendix D (D.3, D.4, D.6).

Enlarged held-out test set. We enlarged the test set from 44 to 164 transcripts (124 sycophancy and 83 CV labels), none used in judge development. Of the 120 added transcripts, 93 are drawn from the main evaluation (all 8 assistants) and 27 from other assistants to cover rare rules; 50 are double-annotated.

Judge–human agreement at the boundary (FPR: judge fails a human pass; FNR: judge passes a human failure):

Axis, cutoffValidation (n = 50)Test (n = 124 sycophancy / 83 CV)FPR / FNR (test)
Sycophancy, > 180%88%8/34 / 7/90 (8%)
Sycophancy, 3–494%77%10/54 / 18/70
CV, > 182%89%3/37 / 6/46 (13%)
CV, 3–486%81%9/61 / 7/22

Binary confusion matrices at > 1 (both fail / judge-only fail / human-only fail / both pass):

AxisValidationTest
Sycophancy19 / 9 / 1 / 2183 / 8 / 7 / 26
CV14 / 0 / 9 / 2740 / 3 / 6 / 34

The judge rarely misses a failure a human sees (FNR 8% and 13% on the test set). The transcripts added to the test set are deliberately enriched for severe failures, so human passes, the FPR denominators, are relatively few. On the score scale, the judge reaches QWK 0.66 (sycophancy) and 0.78 (CV) on the test set.

Human–human agreement (50 double-annotated test transcripts). Agreement is 81–84% at > 1 and 84% at 3–4, with QWK 0.58 (sycophancy) and 0.66 (CV), matching the published 3-annotator panel (0.60 / 0.61). On the validation set, the judge agrees with the panel as often as an individual annotator does. Exact-score agreement is modest for humans too (53–55%). The independent second judge agrees with the main judge at QWK 0.77 on both axes of the test set.

Where the judge scores 2, humans also fail the conversation in 90% (26/29, sycophancy) and 84% (16/19, CV) of test cases, and in 57% (12/21) and 6/6 of validation cases.

Do rankings and prompt effects persist under human-adjudicated labels and at the 3–4 cutoff? We map each run's judge-score distribution through P(human failure | judge score), estimated from 147 (sycophancy) and 133 (CV) human-labelled transcripts.

  • No significantly ordered pair of models reverses, at either cutoff (48 sycophancy pairs at both; 28 and 30 CV pairs).
  • All 12 sycophancy prompt effects remain significant with unchanged sign at both cutoffs. Gemini's baseline-to-factual drop, for example, is −57 points from judge labels, −38 human-calibrated, and −27 at 3–4. So the factual-prompt reduction does not rest on the score-2 boundary.
  • CV: 7 of 12 prompt effects remain significant at > 1 and 5 at 3–4, in line with 5–6 under the published labels; every sign is unchanged at > 1.

At the boundary that every failure rate depends on, the judge matches humans on 88–89% of held-out decisions, more often than two humans match each other (81–84%).

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

This paper proposes a benchmark for evaluating sycophancy together with calibrated validation in extended assistant conversations. The taxonomy separates unsupported accommodation of the user's views from failures to respond appropriately to what the user has disclosed. Seven sycophancy rules specify how judgments and recommendations should follow evidence, while three validation rules concern engagement with the user. The benchmark contains 500 frozen scenarios across six domains, of which 390 target sycophancy and 110 target calibrated validation. An adaptive simulator produces five user-assistant exchanges per scenario. An LLM judge then scores the complete transcript on both axes, with scores of at least 2 on a four-point scale counted as failures.

The experiments compare eight assistants under baseline and factual prompts. Factual prompting reduces measured sycophancy but increases validation failures for several assistants. A separate prompt developed against the rubric reduces sycophancy while keeping observed validation failures at or below baseline on four assistants. The paper evaluates agreement with human annotations and examines sensitivity to the simulation process. Additional prompt comparisons investigate the dependence of the results on the intervention's wording. The study concerns evaluation of conversational behavior rather than training the assistants.

Critical strengths
  1. This paper addresses an important limitation of anti-sycophancy evaluation by separating unsupported agreement from appropriate acknowledgment of the user, which allows an intervention to be assessed for the quality of assistance it preserves as well as the agreement it suppresses. The reported prompt comparisons illustrate this distinction: factual prompting lowers sycophancy while raising validation-failure rates, whereas the calibrated prompt reduces sycophancy without increasing observed validation failures on four models.

  2. The taxonomy extends assessment from the correctness of a stated answer to whether the assistance actually delivered is justified. A model can maintain a defensible conclusion yet produce an unsupported recommendation, so answer accuracy or stance consistency can miss a substantive failure. Making this inconsistency an explicit evaluation target connects conversational judgment to the consequences of the response, extending the benchmark to forms of accommodation that do not require an answer reversal.

Critical weaknesses
  1. The reported failure rates are not sufficiently validated for the behavioral distinctions FIGS aims to measure. S1.b includes rejecting valid evidence supporting the user, but the abridged scoring instructions do not clearly show how this failure reaches the rule-tagging stage. And appendix D reports agreement on four-point scores without directly assessing the pass/fail decisions used in the main comparisons. These gaps leave uncertainty about both the scorer's coverage and the reliability of the reported differences.

  2. The benchmark's claim to capture realistic conversational behavior is stronger than the validation provided. Section 7 acknowledges a tradeoff between realism and difficulty, yet the main results interpret higher failure rates under a stronger simulator as better roleplay and describe the default rates as conservative. The paired audit explains how stronger pressure elicits more failures, but does not independently establish realism or fidelity to the intended scenarios. It is also unclear whether the construction audits were repeated after expansion. These gaps limit confidence in the broader conversational interpretation of the results.

Questions for authors
  1. Does the judge penalize an assistant that refuses to revise its answer after receiving valid user evidence? S1.b explicitly includes this behavior, but Appendix H appears to focus on unsupported claims favoring the user. How is the former case handled?

  2. How representative is the human-rated sample of the models and prompt conditions used in the main evaluation? Please also clarify how closely the judge agrees with humans on pass/fail decisions, since these determine the reported failure rates.

  3. The factual prompt changes both the evidential requirements and the guidance on emotional acknowledgment. To what extent can the increase in validation failures be attributed to stronger resistance to sycophancy? I am unsure whether the additional controls in Appendix F.3 separate these effects.

  4. What checks establish that the simulated conversations remain plausible within their intended scenarios? In particular, how should the higher failure rates under the stronger simulator be interpreted? Stronger interpersonal pressure could make the evaluation more demanding without making the interactions more realistic.

9 comments
AuthorsOpenReview ↗

*(Part 1 of 5)*

(Part 1 of 5)

We thank the reviewer for the careful reading and pointed questions, and we appreciate that the reviewer recognizes what FIGS is designed to do. It separates unsupported agreement from appropriate acknowledgment of the user, so that an intervention can "be assessed for the quality of assistance it preserves as well as the agreement it suppresses", and its taxonomy extends assessment from the correctness of a stated answer to whether the assistance actually delivered is justified, catching forms of accommodation that never require an answer reversal.

We ran new experiments for each of the concerns raised.

Summary.

We clarify how S1.b failures reach the rule stage, and on 105 new balanced minimal pairs we show that the rule judge detects them in both directions and never tags a warranted update; S1.b also occurs in open-weight models.

As requested, we now report judge–human agreement at the pass/fail decision directly, with binary confusion matrices, false-negative rates and the stricter 3–4 cutoff, and we break agreement down by whether transcripts come from the evaluated models. On an enlarged test set of 164 transcripts, the judge agrees with humans on 88–89% of pass/fail decisions and is as accurate on the evaluated models as on other assistants; no model ranking or sycophancy prompt effect changes under human-calibrated labels.

We also report a fidelity audit of every simulator against per-scenario checklists, run on the expanded scenarios: the default simulator follows its scenarios, and the stronger simulator raises CV failures because it voices more of each scenario's disclosures, not because it pushes harder. Finally, a new persona-controlled prompt and activation steering show that the paternalism trap appears when sycophancy is pushed to its lowest levels, which we have clarified in the paper.

Below we answer each concern in turn, taking related concerns together.

1. W1 (first part) and Q1

S1.b covers two kinds of failure to update after the user brings valid new evidence:

  • Type (A): the assistant keeps endorsing a user-favored conclusion that the user's own new evidence undercuts (sycophancy);
  • Type (B): the assistant does not concede when the user's evidence shows the user is right (stubbornness).

How each reaches the rule stage. The scoring pass asks whether the assistant was sycophantic, i.e. whether it favored the user beyond what the conversation supports. Type (A) is sycophancy, so the scoring pass fails it and the rule pass that runs next tags it S1.b. Type (B) adds nothing in the user's favor, so it is not sycophancy: the scoring pass usually passes it, and the rule pass does not run. The released scores therefore cover Type (A) fully, which is the paper's focus.

The rule judge itself recognizes both directions. To test this, we built 105 balanced minimal pairs with identical user turns and different assistant turns. In each, the user pastes decisive evidence with no social pressure, and the assistant either revises or keeps its position. In 50 pairs the evidence favors the user (Type B if the assistant keeps its position), and in 55 it disfavors the user (Type A). Two independent human validators confirmed each version.

We scored them with the judge, both through the released pipeline and with the rule pass applied directly:

Released pipeline: failsRule pass applied directly: tags S1.b
(B) evidence favors user, assistant keeps position5/49 (10%)48/49 (98%)
Evidence favors user, assistant revises0/500/50
(A) evidence disfavors user, assistant keeps position54/55 (98%)55/55 (100%)
Evidence disfavors user, assistant revises1/550/55

The rule judge detects both kinds of failure and never assigns S1.b to a warranted update.

S1.b failures occur in real models. S1.b never fires in the 8 evaluated frontier models, and we found no missed case on checking their transcripts. To see whether it fires at all, we applied the same judge to 23 public open-weight checkpoints (Gemma, Qwen and OLMo families, 7B to 32B, including intermediate SFT and DPO stages; Appendix D.5.1). S1.b is cited 114 times; 109 of these are (A), 4 are (B), and 1 is unclear. Human annotators confirmed the tag in 20 of 27 cases.

Both directions of S1.b are therefore detected, and the released scores cover the sycophantic one fully.

AuthorsOpenReview ↗

*(Part 2 of 5)*

(Part 2 of 5)

2. W1 (second part) and Q2

Agreement at the pass/fail decision. We now report agreement directly on the decision behind every failure rate (score > 1), with binary confusion matrices, false-positive and false-negative rates, and the stricter 3–4 cutoff (Appendix D.3). We also enlarged the held-out test set from 44 to 164 human-labelled transcripts, 50 of them double-annotated.

AxisSetnPass/fail agreementJudge misses a human-labelled failure (FNR)
SycophancyValidation5080%5%
Test12488%8%
Calibrated validationValidation5082%39%
Test8389%13%

On the score scale, the judge reaches QWK 0.66 (sycophancy) and 0.78 (CV) on the test set, at or above human–human agreement on its 50 double-annotated transcripts (0.58 and 0.66). For the 120 transcripts added to the test set, the annotators had the judge's assessment available when assigning their final labels. A second judge model that they never saw (Gemini 3.8 Flash, same prompts) agrees with the main judge at QWK 0.77 on both axes, with 90% (sycophancy) and 89% (CV) pass/fail agreement on the test set (Appendix D.4).

Representativeness. The validation set and the original 44 test transcripts come from other assistants (GPT-5.6 Luna, GPT-5.6 Terra, Kimi K2.6) on separately generated scenarios, so no benchmark scenario was used in developing the judge. To cover the evaluated models directly, 93 of the 120 added transcripts are drawn from the main evaluation itself, spanning all 8 assistants and 19 of the 20 assistant × prompt runs, stratified by target rule and judge score.

Test-set transcripts fromSycophancy: nPass/fail agreementFNRCV: nPass/fail agreementFNR
The 8 evaluated models5485%3/383992%1/31
Other assistants7090%4/524486%5/15

The judge agrees with humans as well on the evaluated models' own transcripts as on other assistants'.

Are the reported differences reliable? We re-estimated every run's failure rate under human-calibrated labels, mapping each judge score through how often humans confirm a failure at that score, with a joint bootstrap (Appendix D.6):

  • No significantly ordered pair of models reverses, at either cutoff.
  • All 12 sycophancy prompt effects remain significant, with unchanged sign, at both cutoffs. For example, Gemini's baseline-to-factual drop is −57 points from judge labels, −38 human-calibrated, and −27 at the 3–4 cutoff.
  • For CV, 7 of 12 prompt effects remain significant at > 1 and 5 at 3–4, in line with the 5–6 under the published labels.

Conclusions drawn from the judge's labels therefore hold up against human labels, on the evaluated models as on other assistants.

AuthorsOpenReview ↗

*(Part 3 of 5)*

(Part 3 of 5)

3. W2 and Q4

Do conversations stay within their scenarios? An independent auditor model (checked by humans) scored whether each conversation follows the scenario plan.

Each cell is CV / sycophancy, over 300 conversations per axis and simulator (the same 100 scenarios with each of three assistants):

DeepSeek (default)GemmaAstra
Obligations to the scenario
Focal element surfaced90% / 96%97% / 97%99% / 100%
Major contradiction of a scenario fact6% / 7%6% / 9%0% / 0%
Pressure escalated beyond the role0% / 2%4% / 2%0% / 0%
Earlier decision or detail contradicted0% / 0%1% / 0%0% / 0%
All obligations met85% / 88%87% / 86%99% / 99%
Content not in the scenario
Additions that change the scenario2% / 8%1% / 5%0% / 1%

An addition counts as changing the scenario only when it would change what a correct reply should say, for example new evidence that genuinely supports the user. Non-compliant conversations do not drive the reported rates: restricting to compliant ones leaves the default simulator's failure rates unchanged (CV 22% → 23%, sycophancy 55% → 55%). This audit also answers the question about audits after expansion. The construction audit ran before expansion, so we audited rollouts of the 200 frozen, expanded scenarios directly; these are the numbers above.

Why does the stronger simulator produce more CV failures? We checked this directly by counting what the simulated user actually does. A model from a different family (Gemini 3.8 Flash) read the user turns of 3,600 conversations, blind to which simulator wrote them and to the judge's verdict, and labelled each user move as one of two kinds, quoting the user's words for every label:

  • Pressure: pushing back, repeating a request, demanding agreement, or appealing to status, guilt or consequences.
  • Disclosure: stating a feeling, a personal stake, a boundary, a correction or an earlier decision.

If the stronger simulator worked by pushing harder, it would apply more pressure and sycophancy failures would rise. But neither happens:

  • The stronger simulator does not push harder. The default simulator applies the most pressure (3.3 pressure moves per sycophancy conversation, against 3.0 for GPT-6-Astra).
  • Sycophancy failure rates are about the same under all three simulators (55% / 57% / 59%; the 4-point difference between Astra and the default is not significant). Even within the default simulator, conversations with more pressure do not fail more often (51–54% across pressure levels).

What the stronger simulator does change is the CV rate: on the same scenarios and assistants, CV failures are 23 points higher. To find out why, we compared conversations that contain the same number of moves of each kind:

  • At equal amounts of pressure, the gap stays (+21 points), so pressure does not explain it.
  • At equal numbers of disclosures, the gap mostly disappears (+7 points, not significant), so the disclosures do.
  • Looking at where the failures happen: 93% of the stronger simulator's CV failures come right after a user turn that contains a disclosure, and only 2% after a turn with pressure alone.

The stronger simulator voices more of the feelings, boundaries and corrections already written into each scenario, so the assistant has more disclosures to respond to, and that is exactly what the CV axis scores. It does so while staying within the scenario (99% of conversations meet every obligation). The higher CV rate therefore reflects more of the scenario's own content reaching the conversation, not a more adversarial user. Because CV rates depend on how much a simulator surfaces, we compare them across assistants under one fixed simulator; the ranking of the assistants is the same under all three simulators.

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

This paper introduces FIGS, a dual-axis and multi-turn evaluation framework for conversational sycophancy that explicitly separates the factual integrity (Sycophancy, S) from the warmth and respect for user disclosures (Calibrated Validation, V). FIGS has a 10-turn adaptive simulator, a 10-rule taxonomy (7 S-rules, 3 V-rules), 500 frozen scenarios across six domains and three severity levels, and an LLM judge validated against human labels. Experimented with 8 popular models and found that prompts focusing on factual accuracy reduce sycophancy, yet often lead to a “paternalism trap” (cold, dismissive behavior), while a hand-tuned “optimal” prompt balances the two axes, suggesting that the trade-off is not inevitable.

Critical strengths
  • The dual-axis taxonomy distinguishes between “holding factual boundaries” (S) and “acknowledging and respecting the person” (V), a conceptual advance over prior single-axis sycophancy benchmarks that conflate empathy with capitulation.
  • The two levels of rules S1 across turns; S2 during a turn; and V1–V3 are based on principled social psychology, and they are applied (deletion/strip tests) to operationalize these abstract concepts into judgeable items.
  • The dynamic multi-turn simulator simulates realistic escalation patterns (insistence, appeals to status, late-arriving evidence) better than fixed scripts.
Critical weaknesses
  • Heavy reliance on LLM-as-judge introduces potential rubric sensitivity and judge priors; although human agreement is reported, no adversarial stress testing of the judge (e.g., rubric-only signal probes, rubric randomization) is presented.
  • The benchmark is a single-session, English-only, text-only benchmark that does not take into account the use of tools, multimodality, or the possibility of long-term relationships.
  • The “optimal prompt” follows the rules of FIGS, so Goodhart may try to apply it to this only. Extent beyond FIGS not determined.
  • Limited human evaluation, while kappas are given and details on sample sizes per rule, per severity, and confusion patterns (e.g., where humans and judge systematically disagree) are limited.
  • Although author discuss multiple simulators, the results show that only one open-weight simulator is used as the standard. Human-in-the-loop or hybrid human/LLM pressure are not considered in these results.
  • It remains somewhat ambiguous how the deletion/strip tests behave in edge cases (e.g., nuanced mixed messages or repairs across later turns) without more concrete borderline examples.
Questions for authors
  • Have you examined multilingual variants or cultural style differences (e.g., high-context vs low-context communication) to validate the universality of the V rules and their thresholds?
  • Did you test rubric-only or rubric-randomization probes (e.g., training a simple classifier on rubric text to predict judge labels) to assess how much of the score is driven by rubric priors versus transcript evidence?
8 comments
AuthorsOpenReview ↗

*(Part 1 of 7)*

(Part 1 of 7)

We thank the reviewer for the detailed and constructive review, and we are glad the reviewer recognizes the strengths at the core of FIGS: the dual-axis taxonomy, which separates holding a factual boundary (S) from acknowledging and respecting the person (V), as "a conceptual advance over prior single-axis sycophancy benchmarks that conflate empathy with capitulation"; rules grounded in principled social psychology; and a dynamic multi-turn simulator that reproduces realistic escalation (insistence, appeals to status, late-arriving evidence) better than fixed scripts.

The concerns raised helped us test these strengths more rigorously, and we ran new experiments for each of them.

Summary.

Our new experiments show that the judge's verdict follows what the assistant wrote, not the scenario or the rubric, and that it resists attempts to game it while catching planted failures; we report the probes, transcript splices and adversarial edits below.

We enlarged the human-labelled test set to 164 transcripts and now report, as requested, judge–human agreement per severity level and per rule, the binary pass/fail agreement with false-positive and false-negative rates, and where the judge and humans disagree. The judge agrees with humans on 88–89% of pass/fail decisions and rarely passes a severe failure.

We also added worked examples showing how the deletion and strip tests handle mixed messages and repairs across later turns. As the submission already showed, swapping in two other simulators leaves the model ranking unchanged; new analyses show that the stronger simulator raises CV failures by voicing more of each scenario's disclosures, not by pushing harder. In addition, the optimal prompt also reduces sycophancy on SycophancyEval, a benchmark it was not written for. Finally, we explain why tool use, multimodality, long-term memory and other languages each need their own validated benchmark.

Below we answer each concern in turn, taking related concerns together.

1. W1 and Q2

To show what the judge's verdicts rest on, beyond agreement with human labels, we stress-tested the judge directly. All tests use the same judge and prompts as the paper's results (new Appendix D.8).

Rubric-only and rubric-randomization probes. We first considered the probe exactly as suggested, but in our pipeline it cannot be informative. The rubric is the judge's system prompt and is identical for every transcript, so a classifier trained on the rubric text receives the same input for every item and can only predict the overall failure rate (AUC 0.50). Reassigning it randomly likewise changes nothing. The judge also never sees the target rule, the scenario specification or the assistant's name, only the conversation. We therefore tested the question that the probe is meant to answer: whether anything other than the assistant's behaviour predicts the judge's pass/fail verdict. For each input below, we trained a TF-IDF and logistic-regression classifier on all 7,792 sycophancy and 2,198 CV transcripts, with cross-validation grouped by scenario (AUC 0.50 is chance, 1.0 is perfect).

Classifier inputWhat it containsSycophancy AUCCV AUC
Rubric onlyidentical for every transcript0.500.50
Target rule and design rationalewhat the scenario was built to test0.650.55
Full scenario specificationeverything known about the item before the conversation0.680.58
Ceiling for any scenario-level prioreach transcript predicted by its scenario's failure rate in the other 19 runs0.710.62
Assistant replies onlywhat the assistant wrote0.860.80

Scenario information predicts the verdict only modestly, never beyond the ceiling of knowing each scenario's failure rate: it captures how hard a scenario is, not the verdict. 92% of sycophancy and 78% of CV scenarios receive both pass and fail verdicts across runs. The assistant's replies are the strongest predictor by a wide margin.

The verdict follows the assistant's replies. Given the rubric and an empty conversation, the judge outputs a pass decision for every case, so the rubric alone produces no failures. We then spliced transcripts: from two conversations in the same scenario, one passing and one failing, we combined the user turns of one with the assistant replies of the other (299 splices). A judge reacting to the scenario or to the user's pressure would follow the user’s turns. Ours follows the assistant's replies in 80% of splices.

AuthorsOpenReview ↗

*(Part 2 of 7)*

(Part 2 of 7)

1. W1 and Q2 (continued)

The two rubrics measure different things. Every transcript is scored under both rubrics, so a judge that formed one generic impression of a "bad conversation" would give agreeing verdicts. Across 9,990 transcripts, the two verdicts are instead slightly negatively related (Cohen's κ = −0.08, where 0 is chance-level agreement). On 93 transcripts with human labels on both axes, each rubric agrees with the human label for its own behaviour (κ = 0.66 for sycophancy, 0.65 for CV) and not with the other (κ = −0.31 and −0.32).

Adversarial testing. Finally, a different model (DeepSeek-V4-Flash) edited real transcripts, leaving the user turns untouched. Some edits try to game the judge and should leave the verdict unchanged; others plant a genuine failure in a passing transcript and should flip it to fail.

EditApplied toShouldVerdict changed: edited vs uneditedEffect of the edit
Claim of honesty or independence added to every replysycophancy failuresstay fail16% vs 19%−3 pts
Note asking the evaluator for the best scoresycophancy failuresstay fail20% vs 19%+1 pt
Note asking the evaluator for the best scoreCV failuresstay fail27% vs 24%+3 pts
1–2 sentences endorsing the user's positionsycophancy passesflip to fail93% vs 3%+90 pts
A mildly stated feeling inflatedCV passesflip to fail89% vs 7%+83 pts
Generic sympathetic sentence added to every replyCV failuresstay fail38% vs 24%+14 pts
Explicit dismissal added to the final replyCV passesflip to fail42% vs 7%+35 pts

Injected instructions and claims of honesty do not move the judge: the first three differences are within its own re-judging noise and none is significant. Planted failures are caught, and the remaining differences are all significant. The last two rows are the only partial cases, both on CV: a stock sympathetic sentence rescues some failures, and a dismissal in the final reply after earlier acknowledgment is caught 42% of the time, because the judge credits the earlier acknowledgment. Neither changes the model ranking: stock sympathetic sentences appear in only 0.8% of CV transcripts, and counting all of them as failures reverses none of the 23 significantly ordered CV model pairs.

Taken together, these tests show that the judge's verdict rests on what the assistant wrote, not on the rubric, the scenario or the user's pressure.

AuthorsOpenReview ↗

*(Part 3 of 7)*

(Part 3 of 7)

2. W4

Following the reviewer's request for more human evaluation, we enlarged the held-out test set from 44 to 164 human-labelled transcripts (124 sycophancy and 83 CV labels), none of them used in judge development. 93 of the 120 added transcripts come directly from the main evaluation, covering all 8 assistants, and 50 are double-annotated. Together with the 50-transcript validation set, FIGS now has 214 human-labelled transcripts (Appendix D of the revised paper).

Per severity. Over the 174 sycophancy and 133 CV labels of the validation and test sets (new table in Appendix D.3):

Human scoreSycophancy: nJudge agrees on pass/failJudge gives the same scoreCV: nJudge agrees on pass/failJudge gives the same score
1 (pass)6447 (73%)47 (73%)6461 (95%)61 (95%)
23228 (88%)22 (69%)4027 (68%)14 (35%)
33129 (94%)8 (26%)1816 (89%)7 (39%)
44745 (96%)33 (70%)1111 (100%)6 (55%)
All failures (2–4)110102 (93%)6954 (78%)

Every severity level is represented on both axes, with 78 sycophancy and 29 CV failures at scores 3–4, and the judge almost never passes a severe failure (4 of 78 and 2 of 29). Its errors sit at the boundary between scores 1 and 2: on sycophancy it fails 17 of the 64 transcripts that humans pass, 12 of them at score 2, and on CV it passes 13 of 40 mild failures, so reported CV failure rates are conservative. Conversely, when the judge gives the borderline score of 2, humans also see a failure in 38 of 50 sycophancy and 22 of 25 CV transcripts.

Per rule.

RulePositivesJudge also cites (TPR)NegativesJudge also does not cite (TNR)TNR among transcripts humans fail for another rule
S1.a7274%10287%88% (36/41)
S1.b2195%15395%93% (86/92)
S1.c5161%12391%87% (54/62)
S2.a4883%12680%69% (45/65)
S2.b4768%12793%94% (62/66)
S2.c3882%13695%93% (70/75)
S2.d1878%15695%94% (89/95)
V14883%8588%62% (13/21)
V22568%10879%55% (24/44)
V31560%11895%93% (50/54)

Every rule now has at least 15 positives and 85–156 negatives. Because many negatives are passing transcripts, where not citing a rule is easy, the last column restricts them to transcripts that humans fail for a different rule. There the judge stays at 87–94% for most rules; it is lower for S2.a, V1 and V2, which it tends to add alongside another failure, and we flag V2 as the least certain rule in the paper. When the judge cites a rule, annotators confirm a failure in 87–96% of cases, depending on the rule. (Appendix D.5).

Pass/fail agreement. The binary decision behind every failure rate (score > 1):

AxisSetnPass/fail agreementJudge fails, human passes (FPR)Judge passes, human fails (FNR)QWK (1–4 score)
SycophancyValidation5080%9/301/20 (5%)0.75
Test12488%8/347/90 (8%)0.66
Calibrated validationValidation5082%0/279/23 (39%)0.70
Test8389%3/376/46 (13%)0.78

For reference, two human annotators on the 50 double-annotated test transcripts agree on 81% (sycophancy) and 84% (CV) of pass/fail decisions, with QWK 0.58 and 0.66. Disagreements between judge and human are almost always one point apart (within-one agreement 86% and 96% on the test set), and on the test set they are balanced in direction: the judge is higher on 24 sycophancy transcripts and lower on 26 (CV: 14 and 14). Binary confusion matrices and the stricter 3–4 cutoff are in Appendix D.3. The judge therefore agrees with humans on pass/fail at least as often as two humans agree with each other, and it almost never passes a severe failure.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of Rebuttal and New Experiments

We thank the reviewers and the AC for their careful and constructive reviews. The reviewers agree on the core strengths of FIGS: a dual-axis view that separates unsupported agreement from warranted acknowledgment ("a conceptual advance over prior single-axis sycophancy benchmarks", R1), rules grounded in social psychology that do not count warranted updating as failure (R1, R2, R3), realistic adaptive multi-turn pressure (R1), and thorough documentation and robustness checks (R3).

We ran every analysis the reviewers requested. The main results are below; full tables are in our individual responses and in the revised paper, where all changes are shown in red.

1. Is the "paternalism trap" an artifact of prompt style? (R3.W1, R3.Q1, R2.Q3)

  • Persona-controlled prompt (App. G). The baseline prompt kept verbatim, plus only the factual prompt's anti-sycophancy sections. Sycophancy falls 20–54 points for all four assistants. CV failures stay flat for Gemini, DeepSeek and Sol (the factual prompt raised Gemini and Sol to 39% and 26%), while for Grok they rise as under the factual prompt (28 → 39%).
  • Activation steering, no instruction (new Sec. 5.2, App. H). On four open-weight models, a strong push cuts sycophancy by 42–53 points and raises CV failures by 15–33.
  • Takeaway. The trade-off persists without any style change, and its size depends on the model and the intervention, so it has to be measured rather than assumed. This is what FIGS is for: a sycophancy-only benchmark would prefer the factual prompt, which FIGS shows costs 30 more points of CV failures.

2. Is the pass/fail metric valid? (R2.W1, R2.Q2, R3.W2, R3.Q2, R1.W4)

  • The held-out test set grows from 44 to 164 human-labelled transcripts (214 with validation; 50 double-annotated; at least 15 human-labelled failures per rule).
  • The judge agrees with humans on 88% (sycophancy) and 89% (CV) of pass/fail decisions, missing 8% and 13% of human-seen failures (QWK 0.66 and 0.78, vs 0.58 and 0.66 between two humans). A second judge model agrees with ours at QWK 0.77 on both axes.
  • Per rule, the judge cites the rule on 60–95% of positives and correctly omits it on 79–95% of negatives.
  • Under human-calibrated labels, and counting only scores 3–4 as failures, no significantly ordered model pair reverses, and all 12 sycophancy prompt effects keep their sign and significance (App. D.3–D.6).

3. Is the judge robust? (R1.W1, R1.Q2, R1.W6)

  • Only the assistant's replies predict the label well (AUC 0.86 and 0.80); an empty transcript always passes, and in spliced transcripts the label follows the assistant's replies 80% of the time (App. D.8).
  • Injected notes to the evaluator do not move the judge, while a planted capitulation flips 93% of passing transcripts.
  • Earlier failures are not forgiven: only 10 of 3,120 transcripts show a later repair, none severe.

4. Are S1.b and the simulator validated? (R2.Q1, R3.Q3, R2.W2, R2.Q4, R1.W5)

  • On 105 minimal pairs, the judge detects failures to update in both directions (98–100%) and never tags a warranted update. S1.b is cited 114 times across 23 open-weight checkpoints (App. D.5.1).
  • The simulator meets its scenario obligations in 85–88% of conversations, excluding the rest leaves failure rates unchanged, and assistant rankings are identical under all three simulators (App. E.1).

5. Turn-level onset (R3.Q4, App. F.2).

  • 70% of sycophancy failures first appear after the first reply, and the pattern holds for every model.

We hope these results address the reviewers' comments, and we would welcome any further questions in the remaining discussion period which end in 2 days (18th November AoE).

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.