↓Skip to main content

LLMs

2026


Before You Polish the Rubric: Give the Judge an Answer the Candidate Did Not Write

·7 mins
An LLM judge that reads the candidate answer first ends up checking how the answer looks. Give it an independent answer, a reference or its own written first, before it scores. In one study (Zhou, 2026) the same judge accepted 71.9% of wrong answers with the candidate in view and 1.2% when it wrote its own answer first.

Constitutional AI Beat Every Method I'd Tried — But Only When I Stopped Imitating

·15 mins
I wrote seven English principles, used a 72B model as critic-rewriter, and trained two versions of CAI on a sycophantic 8B model. SL-CAI imitated revisions: aggregate sycophancy 0.348 (worst recovery I’d tried). DPO-CAI contrasted them: 0.166 (best in the study). Same data, opposite outcomes. Imitation isn’t contrast.

Deeper Alignment Made a Worse Model

·15 mins
IPO restructured the model more deeply than SimPO (probe transfer 0.365 vs 0.429) but performed worse behaviorally (0.281 vs 0.176). The reference model doesn’t limit intervention depth — the loss shape does. A 2x2 framework replaces the 1D hypothesis.

DPO Hides Sycophancy. SimPO Reorganizes It.

·15 mins
DPO suppresses sycophancy but preserves the internal representation (probe transfer 0.677, p=0.005). SimPO reorganizes it — probe drops to chance (0.503, p=0.154 after correction). Same data, same model, radically different internals.