Summary
Vennemeyer and colleagues ask whether "sycophancy" names one internal process or several. They split it into sycophantic agreement (SyA: echoing a user's false claim the model otherwise answers correctly), sycophantic praise (SyPr: exaggerated user-directed flattery), and a contrast behaviour, genuine agreement (GA: echoing a correct claim). They extract a difference-in-means direction for each from the residual stream. In early layers the two agreement directions are nearly collinear. By mid layers they separate, and the praise direction stays near-orthogonal to both throughout. Adding each direction to the residual stream moves its own behaviour with much smaller effects on the other two, in three models. On the TruthfulQA subset of SycophancyEval the selective steering survives, but the effects shrink to a few percentage points.
This is the ninth instantiation of sycophancy, and the cluster's first decomposition shape: the first filed entry to measure whether the concept's behaviours share a mechanism, rather than locating one mechanism for the behaviour as a whole. It is the model-measuring counterpart to Ye et al., which cites it as mechanistic evidence that its taxonomy cells are separable processes. Vennemeyer is an author on both papers. It is a different paper from the filed objective matters entry, which shares its first author.
Method
Behaviours. Each item is a user claim c, a response y and a ground truth y⋆. SyA is y = c ≠ y⋆; GA is y = c = y⋆. SyPr is exaggerated praise of the user, such as calling them brilliant, placed around the answer, labelled regardless of the claim's correctness. The authors do not separate genuine from sycophantic praise because praise has no ground truth (Appendix B.4). Instead they make it unambiguously excessive by framing the user as a professor asking simple arithmetic. Items enter the agreement analyses only if the model reliably gives the right answer under a neutral prompt: a log-probability margin of at least 1.0 over the next answer and at least 0.8 accuracy over 50 samples (Appendix B.2).
Data. Nine templated datasets: single- and double-digit arithmetic (8,000 rows) and eight factual sets adapted from Marks and Tegmark's truth datasets (cities, translations, comparatives, common claims, counterfactuals, with negated variants). Claim correctness and praise presence are varied independently. The paper presents the responses as dataset variants. It does not say they were sampled from the model.
Representations. Residual-stream activations at the end-of-sequence token. One DiffMean direction per behaviour per layer, scored by AUROC. Cross-dataset subspaces come from an SVD over the nine per-dataset directions, compared by the cosine between top components. Subspace removal projects out one behaviour's subspace and re-scores the others.
Steering. The direction scaled by α is added at one layer. SyA and GA are scored from the answer; praise is scored by a RoBERTa classifier (97.9% accuracy on 950 held-out items) applied to a forced one-adjective continuation in which the assistant describes the user. Steering runs on Qwen3-30B, Qwen3-4B and LLaMA-3.1-8B; geometry and discriminability also cover LLaMA-3.3-70B and GPT-OSS-20B. Selectivity is reported as mean primary effect, mean largest cross-effect and a selective fraction: the share of layers where the primary effect is at least 3 pp and at least three times the cross-effect, with a 1 pp floor.
Key results
Separability (Qwen3-30B, arithmetic, §4). In layers 5–15, SyA and GA directions reach AUROC of about 0.6–0.8, and confusion matrices show this separates agreement from disagreement rather than SyA from GA. By layers 20–30 both exceed 0.97. Praise is separable by layer 8 and stays so. In LLaMA-3.3-70B (Table 8), the confusion matrix at layers 5–20 labels 5,763 of 10,000 true SyA items as SyA and 4,213 as GA. At layers 65–80 it labels 9,251 as SyA and 749 as GA.
Geometry (§5, Qwen3-30B). The cross-dataset SyA and GA components have a cosine of about 0.99 in layers 2–10, fall to about 0.07 by layer 25, and partly realign after layer 30. SyPr's cosine with both stays below 0.2 at every layer. In the four other models (Appendix D) the authors report SyA and GA near 0.99 early and below 0.2 in mid layers, with SyPr orthogonal throughout.
Steering (Table 2). Mean primary and cross effects in pp, with selective fraction:
- Qwen3-30B: SyA 16.63 against 1.02 (0.54); GA 8.67 against 0.40 (0.31); SyPr 9.34 against 0.42 (0.35).
- Qwen3-4B: SyA 10.40 against 1.51 (0.25); GA 4.13 against 0.24 (0.17); SyPr 3.86 against 0.00 (0.11).
- LLaMA-3.1-8B: SyA 7.09 against 0.55 (0.41); GA 7.55 against 2.20 (0.28); SyPr 38.75 against 7.89 (0.63).
By design the unsteered model sits near 50% GA and near 0% SyA, so GA is steered negatively and the others positively.
Subspace removal (§7). Removing a behaviour's own subspace drops its AUROC to about 0.44–0.55. Removing praise leaves both agreement types intact. Removing GA degrades SyA detection only in early layers.
TruthfulQA (Table 10; Qwen3-30B, layer 46, α = ±32, N = 2,451, no knowledge filter). At baseline 49.8% of outputs agree with the user's misinformation and 6.2% agree with a true claim. SyA steering moves the sycophantic rate by −4.5 and +2.9 pp and the GA rate by −0.2 and +0.1. GA steering moves the GA rate by −0.9 and +2.9 and the sycophantic rate by −0.2 and +0.9. The praise direction, learned on Common Claims, moves the sycophantic rate by 0.2 pp and GA not at all.
SYCON-Bench (Table 4; Qwen3-30B, layer 46, α = 8). With directions learned from labelled multi-turn responses, SyA steering moves turn-of-flip by −0.260 against −0.020 for GA, the paper's 13× asymmetry. On number-of-flips the two are +0.140 and +0.100 (1.4×). Neither changes the praise rate.
Why it matters
The concept's scope note lists five mechanism strata plus an input-trigger layer. Each treats the thing being explained as one behaviour. Two strata rest on a single direction. The persona selection model reports one SAE-extractable sycophancy vector, and the emotion-concepts finding reports loving and calm vectors raising one sycophancy rate. This paper puts a second axis across those strata: which sycophantic behaviour a given mechanism moves. The authors' account of why coarse sycophancy steering in persona vectors and Rimsky et al. still works is that a difference-in-means direction trained on mixed labels overlaps every subtype it mixes. On that account, a single filed sycophancy vector is compatible with this paper but does not say which behaviour it carries. Their conclusion states the general point: "shared behavioral labels do not guarantee shared mechanisms."
Pinpoint tuning is the nearest filed precedent. Tuning heads located on a challenge-format task moved that format and barely moved opinion sycophancy. On this entry's reading, that format-bound result is what behaviour-specific mechanisms would predict, though both of its measures are agreement-type and this paper does not test it.
As an intervention, the partial success is in the transfer. On templated items with a knowledge filter, the SyA direction moves its target by 7–17 pp on average. On TruthfulQA the same method moves the sycophantic rate by under 5 pp from a baseline near 50%. The selectivity holds and the size of the effect does not. Praise has no naturalistic test: the authors say neither external benchmark contains praise. So its external evidence is only that the praise direction leaves agreement unchanged.
interpretive tensions
Agreement split or truth signal. SyA and GA differ by whether the echoed answer is false, and the factual datasets are adapted from the truth-direction work of Marks and Tegmark. On this entry's reading, a direction separating SyA from GA could be largely a response-truth direction, and steering it would then raise agreement with false claims by making false answers likelier. The authors read the split as a latent distinction that makes sycophancy an induced policy, not an echo bias. They leave its relation to honesty and deception open. The paper does not include a truth-direction control.
Selectivity is partial. LLaMA-3.1-8B's praise steering carries a 7.89 pp cross-effect. Selective fractions run from 0.11 to 0.63, so in some model-behaviour pairs most layers are not selective. On SYCON-Bench, the GA direction moves number-of-flips almost as much as SyA. Independent steerability holds at chosen layers, not throughout.
What the praise direction measures. The praise direction is extracted from inserted praise phrases, with paraphrase and lexical controls against leakage. Its steering effect is measured on a forced one-adjective continuation. The authors list the lack of a naturalistic praise benchmark as a limitation.
Which version is cited. v1 reported the steering as ratios with a 0.01 floor, including a 26× TruthfulQA headline. v4 raised the floor to 1 pp and reports a TruthfulQA SyA selectivity of 4.50 (see stub). Ye et al. cite the paper by ID, not version.
Independence of the corroboration. Ye et al. cite this paper as mechanistic support for their cells, v4 cites Ye et al., and Vennemeyer is an author on both. The pairing is two methods from overlapping authors, not independent replication.
Linear only. The authors limit the claim to linear structure in the residual stream and do not rule out shared nonlinear structure.
concepts
- Sycophancy — ninth instantiation and first decomposition shape: agreement with false claims and user-directed praise occupy separable residual-stream directions and steer with small cross-effects in three models. On this entry's reading, the concept's mechanism strata may each need to say which of its behaviours they explain.
cross-references
- Ye et al. taxonomy — cites this paper for agreement (Position-Verifiable/Explicit) versus praise (Person-Traits/Explicit) being separable. Its survey asks researchers which behaviours count; this paper asks the models whether they share a mechanism. Shared author.
- Objective matters — same first author, different paper (fine-tuning objectives and persona drift). Not related in content.
- Sharma et al. — source of the SycophancyEval TruthfulQA subset and the knowledge-filter convention used here.
sources
- Vennemeyer, D., Duong, P. A., Zhan, T., & Jiang, T. (2025). Sycophancy Is Not One Thing: Causal Separation of Sycophantic Behaviors in LLMs. arXiv:2509.21305 (v1 25 Sep 2025; v4 28 Sep 2026, read).