R0057/2026-04-01/C005

Research	R0057 — RLHF Yes-Men Claims v3
Run	2026-04-01
Claim	C005

Claim: Synthetic non-sycophantic training data reduces sycophancy by 4.7-10%.

BLUF: Confirmed. Wei et al. (2023) report reductions between 4.7% (Flan-PaLM-62B) and 10.0% (Flan-cont-PaLM-62B) across PaLM model variants.

Probability: Very likely (80-95%) | Confidence: High

---

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	The 4.7-10% range is accurate	Supported
H2	The range is accurate but may not generalize to all models	Not supported
H3	The reported percentages are wrong	Eliminated

Searches

ID	Target	Results	Selected
S01	Synthetic data reduces sycophancy 4.7 10 percent	10	1

Sources

Source	Description	Reliability	Relevance
SRC01	Wei et al. (2023) — Simple synthetic data reduces sycophancy	High	High

Revisit Triggers

• If replication studies show different reduction ranges