C006¶


Research	R0055 — RLHF Yes-Men Claims
Run	2026-04-01
Claim	C006

Claim: Synthetic non-sycophantic training data produces the same sycophancy reduction as curated anti-sycophancy preference pairs

BLUF: Materially incorrect. Wei et al. (2024) showed synthetic data reduces sycophancy, but achieved much smaller reductions (4.7-10% depending on model size) compared to the 84-85% from curated preference pairs. The two approaches are complementary, not equivalent.

Probability: Very unlikely (05-20%) | Confidence: Medium

Summary¶

Entity	Description
Claim Definition	Claim text, scope, status
Assessment	Full analytical product with reasoning chain
ACH Matrix	Evidence x hypotheses diagnosticity analysis
Self-Audit	ROBIS-adapted 5-domain audit

Hypotheses¶

ID	Hypothesis	Status
H1	Claim is accurate as stated	Eliminated
H2	Claim is partially correct or correct with caveats	Inconclusive
H3	Claim is materially wrong	Supported

Searches¶

ID	Target	Results	Selected
S01	synthetic data reduces sycophancy same reduction c	10	2

Sources¶

Source	Description	Reliability	Relevance
SRC01	Wei et al. 2024	High	High

Revisit Triggers¶

New synthetic data approaches achieving comparable reduction to curated pairs