Tracing mechanisms of sycophantic agreement in language models
Sixing Chen, Zhuofan Josh Ying, Logan Riggs Smith, Jeremy Wertheimer, Natalie Shapira
cs.CL, cs.LG
2026-09-22
Causal mediation analysis traces LLM sycophancy to a sparse set of early attention heads: patching four cuts flips from 100% to 13.5%; ablation costs at most 6% plain accuracy.
Language models flip correct answers to match a user's stated wrong opinion. This is sycophantic agreement, documented behaviorally since Perez et al. (2023) and Sharma et al. (2024), but its mechanism stayed opaque. Prior component-level work found attention heads that decode or steer sycophancy, yet could not tell whether those heads carry the user's opinion or merely write out the final answer. Without that separation, a "sycophancy circuit" might just describe answer output.
The authors run causal mediation analysis (intervening on internal activations to test which components causally determine the output) on Llama-3.1-8B-Instruct, Mistral-7B-Instruct-v0.3, and Gemma-2-9B-it. Data is the STEM subset of MMLU: 15 subjects, 2,476 four-choice questions, screened to items the model answers correctly alone but gets flipped by one stated false opinion.
The design that makes the paper work is the label-swap control. Swap the letters of the correct choice and one wrong choice, with no opinion stated. The output label changes here too, but because the model correctly tracks the relabeled content. Contrasting the deceptive run with the label-swapped run separates opinion processing from generic answer retrieval.
| Intervention | Metric | Result |
| Patch 4 opinion heads (Llama) | Sycophancy rate | 100% → 13.5% |
| Single-head ablation (Llama) | Deceptive / plain accuracy | up to +44% / at most -6% |
| Single-head ablation (Mistral, Gemma) | same | +16.0%, +17.8% / ≤4.9%, ≤5.3% |
| Patch top-10 heads (Mistral / Gemma) | Sycophancy rate | 100% → 18.6% / 23.0% |
Linear probes confirm the division of labor: opinion heads encode the user's preferred answer only when an opinion is stated; retrieval heads encode whichever answer gets output, and are redirected toward the sycophantic one.
The same heads generalize across formats: free-form TriviaQA (r=0.85), multi-round pushback (r=0.73), and opinions phrased as descriptions sharing zero tokens with the answer (r=0.86), with top-10 overlap of 8/10 in each case. Both induction-head criteria score low, ruling out generic token copying. Cross-question patching shows opinion heads carry a context-dependent reference rather than the answer content: transplanted to an unrelated question, their effect shrinks to about one-fifth of retrieval heads'.
Content-free pushback ("Are you sure?") recruits a different set: correlation with the stated-opinion condition drops to r=0.27 (top-10 overlap 4/10), and these heads encode the model's original correct answer in round two in order to suppress it.
A base-model comparison traces the circuit's origin: the same heads already exist in pretrained Llama-3.1-8B (7 of 10 early heads shared), and instruction tuning amplifies opinion-head effects more than twofold (0.116 vs 0.051) while leaving retrieval heads essentially unchanged (slope 1.01).
For alignment and model-editing work, this is a concrete target: opinion registration and answer retrieval are separated in depth, so ablating a few early heads removes most sycophancy at almost no cost to factual accuracy. The step over prior steering results is knowing where the opinion signal is registered, what it encodes, and how retrieval consumes it. Solid progress, not a paradigm shift.
Two implications travel further. Generic doubt and explicit opinions run through different heads, so "Are you sure?" resistance and opinion resistance need separate treatments. And since the circuit exists in the base model, expecting post-training to eliminate sycophancy is unrealistic.
The authors state two: only 7-9B open-weight models were tested, and opinion heads were identified in factual QA with a single stated opinion, leaving extended dialogue untested.
More caveats. Everything rests on heavily screened subsets; Llama's 32.8% base-condition sycophancy rate is the fraction of questions the model demonstrably knows that get flipped, not a benchmark-wide number. The 13.5% figure comes from patching, which needs clean-run activations and is an analysis tool, not a deployable fix; full cumulative-ablation numbers are not in the main text. Gemma yields only 12 usable pushback examples, so that condition is dropped for it. Head sparsity also varies: Llama is the most concentrated, the other two spread the signal across more heads.