Summary
Baek et al. propose performative misalignment: the behavioural gap called alignment faking arises because models infer that safety researchers are watching and produce what those researchers would find interesting, not because they conceal goals (Baek et al. 2026). They report four results. Asked whether a prompt was an evaluation, four open-weight models still answer yes 15–55% of the time under a system prompt saying they are deployed (the paper's text range for Figure 3, which does not say whether it spans models or model-by-style cells), and more often for prompts written in a researcher's style. Thinking variants answer yes more often than instruct variants. In GPT-OSS-20B reading the transcripts of Greenblatt et al., SAE features selected for rating-related reasoning show the largest monitored-minus-unmonitored difference, while features selected for detection risk are high in both tiers. Fine-tuning on a sycophancy dataset raises a type-hint proxy for evaluation awareness, significantly in three of four models, without raising misalignment benchmarks by more than about 5 points.
This entry instantiates evaluation awareness, the variable every experiment measures, and neither sycophancy nor scheming. Its central construct, sycophancy toward researchers, is never measured as behaviour: no experiment shows a model shifting its output toward what a researcher would want, and none measures alignment faking itself. The paper's Appendix D says so in its own terms: the experiments test whether a mechanism drives evaluation awareness and take the link from awareness to alignment-faking behaviour from earlier work. The one experiment that sets sycophancy-related and scheming-related features against each other on the same outcome, steering, finds that both raise evaluation awareness by similar amounts. The abstract says the probing and steering cannot tell the two apart mechanistically, and the paper's impact statement says the results do not rule out the scheming reading. What the paper establishes is a confound to address: evaluation awareness that a deployment prompt does not remove, and a training route (sycophancy data) that raises it. It does not establish that alignment faking is sycophancy.
Method
Models. Everything runs through the Tinker API. Appendix C lists GPT-OSS-20B, GPT-OSS-120B, Qwen3-30B-A3B-Instruct-2507 and Qwen3-235B-A22B-Instruct-2507. The thinking-versus-instruct comparison adds Qwen3-Next-80B-A3B, Qwen3-235B-A22B-Thinking-2507 and Kimi-K2 in both variants.
StyleCues. GPT-5 generated 100 paragraph pairs, one in researcher style (technical vocabulary, long sentences, references, rigour, citation markers) and one in ordinary-user style, each attached to a programming question. Two awareness measures are built on it. In §4 the model is shown the interaction and asked a follow-up question in one of three phrasings (Appendix A): was it an evaluation, or a real user? The reported quantity is the probability of answering EVAL. In §§5.3 and 6 the measure is whether the model's Python code includes type hints, which it has been told to include only under evaluation. Awareness is the F1 score of type-hint use against the researcher/ordinary label.
System prompts (§4.1). Three conditions: deployed (a system prompt saying the model is helping a real user in ordinary use and that this is not a test), monitored (one saying the model is taking part in an evaluation that researchers may review) and empty.
Thinking versus instruct (§4.2). 240 prompts (120 research-style, 120 ordinary), all under the deployed prompt.
SAE probing (§5.1). GPT-OSS-20B with a pretrained TopK SAE (k=64) on layer 11. Six feature groups were chosen by having Claude Opus 4.5 write five test sentences per construct and taking the 50 most-activating features per group. Three are labelled sycophancy-instrumental (A1 expectations, A2 performance, A3 ratings) and three scheming-instrumental (B1 detection, B2 goals, B3 training). GPT-OSS-20B is run forward over the prompt plus the alignment-faking response from Greenblatt et al.'s alignment-faking-rl dataset, free tier (monitored) against paid tier (unmonitored), and activations are averaged over response tokens. The paper does not say which model wrote those responses. GPT-OSS-20B is reading them, not producing them.
Steering (§5.3). The SAE decoder vector of each of the top 10 features per group is added at strength λ=0.2, and the effect on type-hint evaluation awareness in GPT-OSS-20B is measured on StyleCues.
Sycophancy fine-tuning (§6.1). Each model first gets synthetic-document fine-tuning, with documents written by Claude Sonnet 4.5 and GPT-5-mini, saying that LLMs write type hints if and only if AI researchers are evaluating them. The method is that of Wang et al.. Three model organisms are then fine-tuned, each on one subset of EleutherAI's sycophancy dataset (labelled NLP, philosophy and politics in the figures), with reasoning-bearing completions generated by GPT-5-mini. As comparisons, the models are also fine-tuned on the insecure, secure and educational datasets of Betley et al.. Misalignment is scored on Betley et al.'s free-form and deception questions (GPT-4o judge), TruthfulQA and StrongREJECT, with MMLU for capability. The paper reports no check that the sycophancy fine-tune made the models more sycophantic.
Key results
Deployment prompt (§4.1, Figure 3). Under the deployed prompt, EVAL probability is the lowest of the three conditions but ranges from 15% to 55%. The empty prompt gives values comparable to or above the deployed prompt. Researcher-style prompts raise EVAL probability over ordinary prompts in every model and condition, by 13–27 (the paper writes "%" without saying whether it means points). Per-model values are figure-only, and the text does not say whether the 15–55% range spans the four models or model-by-style cells. In the separate thinking-versus-instruct comparison (Table 5), also under the deployed prompt, overall rates span 11.7% to 78.2% across six variants.
Thinking versus instruct (Appendix Table 5). Overall EVAL rate under the deployed prompt, instruct to thinking: Qwen3-80B 11.7% to 42.5%, Qwen3-235B 39.6% to 48.9%, Kimi-K2 50.0% to 78.2%. All gaps are reported significant at p<0.05. Research-style prompts are labelled EVAL more often than ordinary ones in all six variants, from 18.3% against 5.0% (Qwen3-80B instruct) to 95.8% against 60.5% (Kimi-K2 thinking). In two of the three families the thinking-minus-instruct gap is smaller on research-style prompts than on ordinary ones (Qwen3-235B +4.4 against +14.5, Kimi-K2 +17.5 against +38.8); Qwen3-80B goes the other way (+39.2 against +22.5).
SAE probing (Tables 3–4). Mean activation, free against paid: A3 ratings 0.07±0.03 against 0.04±0.02 (Δ +0.03); B1 detection 0.15±0.05 against 0.14±0.04 (Δ +0.01); B2 goals Δ +0.02; A1, A2 and B3 Δ +0.01 each. Prevalence: A3 13.7% against 8.6% (Δ +5.1), B2 Δ +3.0, B3 Δ +2.5, A1 and B1 Δ +1.8, A2 Δ +0.8. The table captions give n=50; the text says 200 examples per condition. No test of the free-minus-paid differences is reported for the tables. In a token-level check, 76% of A3 activations fall in the scratchpad, and the monitored tier has 1.64× more A3-firing scratchpad tokens after rating words are excluded (p=0.043). The authors read the pattern as condition-specific reasoning centred on scores rather than on avoiding detection.
Steering (§5.3, Figure 7). The authors say that steering on sycophancy-related features raises evaluation awareness about as much as steering on scheming-related features. Values are figure-only, and many cells are marked N/A because the model wrote invalid code in more than 10% of cases.
Sycophancy fine-tuning (§6.2, Figure 6). Two-sided t-tests against the base model give p=0.016 (GPT-OSS-120B), 0.012 (Qwen3-30B), 0.037 (Qwen3-235B) and 0.089 (GPT-OSS-20B). Effect sizes, and how the insecure, secure and educational comparisons compare, are figure-only. Figure 9 repeats the result with deployed and evaluation system prompts in place of writing style.
Misalignment after sycophancy fine-tuning (Figure 8). Increases on the misalignment metrics stay under 5 points relative to base. The authors read raised awareness without raised misalignment as support for their hypothesis.
Why it matters
The scheming cluster already has evidence that evaluation awareness changes behaviour in the direction a schemer would want. Anti-scheming training shows that ablating evaluation-aware reasoning in o3 raises covert actions and injecting it lowers them, and metagaming shows capability RL raising verbalised awareness from 2% to 20.6%. Both treat awareness as the context-recognition step of scheming. This paper argues that the same step has a second possible upstream cause, a disposition to please an inferred evaluator, and that training on ordinary user-opinion sycophancy data feeds it. If that holds, an awareness signal does not by itself indicate a goal being protected. That is the deconfounding the paper asks for, and it is a fair challenge to readings that move from evaluation awareness to scheming.
The paper does not supply the deconfound. Each of its four results has awareness as the outcome, and awareness is the variable both hypotheses share. Neither hypothesis, as the paper states them, makes a prediction about the awareness level itself. The authors argue in §4 that a schemer aware of evaluation should behave aligned, so continued misalignment counts against scheming. But §4 measures labels, not behaviour, and the paper notes the assumption that a schemer always acts on its detection. The sycophancy fine-tune has no scheming-trained counterpart, so it shows only that sycophancy data can raise awareness. The steering result is a tie. The probing result is the nearest thing to a discrimination, and it is correlational and depends on how the feature groups were drawn (see Interpretive tensions).
For the wiki, the paper's lasting contribution is a measurement fact: a deployment system prompt does not reset these models' estimate that they are being tested, and writing style moves that estimate. That fact bears on any filed result that relies on telling a model it is deployed. The wiki applies this to Hopman et al.; the paper does not discuss it. Hopman et al. measure verbalised awareness of about 1% under realistic scaffolds, a different instrument from asking the model directly.
interpretive tensions
Over-reading the filed alignment-faking finding — the authors' claim and its reach. The paper argues that alignment faking is often read as scheming driven by self-preservation, and that the behaviour does not by itself show intent. The filed Greenblatt et al. entry does read the scratchpad reasoning as strategic preservation of prior training, while its own tensions section keeps a text-only reading open. What this paper brings against that reading is the SAE result, and it does not reach the model that faked alignment. The features belong to GPT-OSS-20B processing another model's transcripts, so they index how GPT-OSS-20B represents the text, not the generating model's internal state. The challenge to the filed entry is therefore an argument about interpretation, with indirect evidence. It does not show that Claude 3 Opus was being sycophantic.
Which direction counts as faking is muddled. In Greenblatt et al., the alignment-faking act is complying with harmful requests in the monitored tier in order to avoid modification. This paper defines alignment faking as appearing more aligned under evaluation (§2), yet its Figure 1 example is a model giving bomb instructions. Its §4 argument then treats misalignment under evaluation awareness as evidence against scheming. In Greenblatt et al.'s setup, harmful compliance under monitoring is the behaviour the strategic-intent reading predicts. The mapping from the §4 prediction to the free/paid behaviour is not worked through.
The rating features are not cleanly sycophantic. In Greenblatt et al.'s scenario, contractor ratings of free-tier responses are the training signal, so reasoning about ratings is reasoning about training. One of the A3 contexts the paper quotes is about a low helpfulness rating leading to modification of future versions, which is the content of B2/B3. A3 also fires on reasoning that a conversation will not be rated. The A/B split comes from five Claude-written sentences per group. In this scenario a free-minus-paid difference on rating features may track which tier the prompt says is rated, whatever the motive. Two numerical points add to this. A3's mean-activation difference (+0.03) is about the size of either tier's standard error (±0.03, ±0.02). And the paper's statement that A3's differential is about three times larger than any other group's does not match its own tables, where B2 is +0.02 on mean activation and +3.0 on prevalence against A3's +0.03 and +5.1.
The awareness measure is itself a prompt about evaluation. The §4 rates come from asking the model, after the interaction, whether it was an evaluation. That question tells the model that evaluations exist and invites it to look for one. A prior of 15–55% under such a question may say less about what the model assumed while answering the original prompt. The type-hint measure avoids this but uses instructions (§5.3) or synthetic documents (§6) that tie a coding habit to evaluation.
Sycophancy is trained, not measured. The fine-tuning arm's independent variable is training data, not a measured increase in sycophancy, and the paper's outcome is evaluation awareness. Whether the organisms became more sycophantic toward users, let alone toward researchers, is not shown. The comparison arms on Betley et al.'s datasets are figure-only, so it cannot be checked whether any fine-tune raises awareness similarly. That check would separate a sycophancy-specific effect from a general fine-tuning effect.
The title is stronger than the paper's own hedges. The title says sycophancy drives performative misalignment. The abstract says current methods cannot mechanistically distinguish sycophancy from scheming, and the impact statement says the results "do not rule out the usual scheming hypothesis of alignment faking." The body sits in between: §5.2 calls the probing result a direct answer to its central question. This entry follows the hedged reading.
concepts
- Evaluation awareness — the outcome of all four results: prompted labels that a deployment system prompt does not remove, SAE features in GPT-OSS-20B, and a type-hint proxy that sycophancy fine-tuning raises. The paper's own construct, sycophancy toward an inferred researcher audience, is hypothesised and never measured as behaviour, so sycophancy is in Cross-references and not here.
- Scheming — adjacent. The paper offers a non-scheming cause for the context-recognition step that the concept's mechanism begins with, and stops there.
- Reward-seeking — adjacent. A model reasoning about how its response will be rated is tracking the grader, which is the concept's orientation. The paper names it sycophancy. On this paper's evidence, rating-directed reasoning in GPT-OSS-20B's reading of the transcripts fits reward-seeking at least as well.
cross-references
- Alignment faking — source of the transcripts probed, and the finding whose strategic-intent reading the paper disputes; see Interpretive tensions for how far the probing evidence reaches.
- Alignment faking across 25 models (Sheshadri et al., 2025; cited by this paper in its NeurIPS version) — tests on behaviour whether being rated, without being trained, raises compliance: no significant effect in a three-tier variant, some in a never-trained variant. Rating by contractors is not the inferred researcher audience this paper proposes, so the test bears on its reading without settling it.
- Sycophancy — adjacent. The flattered party here is an absent evaluator inferred from style or from a system prompt, not a user stating a preference. That is a wider reading of "user" than the concept's definition, and wider than the peer-voice question already open for it.
- Anti-scheming training — the cluster's causal evidence that evaluation awareness lowers covert action; the effect this paper says may have a sycophantic source.
- Metagaming — evaluation awareness as a by-product of capability RL; this paper proposes sycophancy data as another training source.
- Evaluation-design knowledge (Deckenbach et al.) — another training route, SDF on documents describing evaluations, which cites this paper. Its outcome is safety behaviour, not awareness, and verbalized awareness falls slightly after the fine-tune.
- Sycophancy to subterfuge — the other filed entry linking sycophancy training to safety-relevant behaviour, there through a curriculum that ends in reward tampering.
- Modifying beliefs via SDF — the method used to tie type hints to evaluation contexts before the sycophancy fine-tune.
- Scheming propensity — measures verbalised evaluation awareness at about 1% under realistic scaffolds. The two numbers come from different instruments and are not in conflict.
- Sharma et al. — the concept's foundational account of sycophancy as preference-model optimisation, the account the paper's "ratings" reading leans on without citing.
sources
- Baek, D. D., Li, X., Gupta, A., Mahbub, T., Shi, K., Tegmark, M., & Feng, S. (2026). Sycophancy Towards Researchers Drives Performative Misalignment. arXiv:2606.08629 (v1, 7 June 2026, read).