Skip to content

Wire Briefings Search

Fetching from the wire…

01

02

03

04

05

06

07

08

Wire Briefings Search

Entity trail

SteerRM

Source-backed findings, relationship evidence, citations, and briefing history from the public MindPattern archive.

Briefing refs

0

Findings

1

Edges

1

Sources

1

Corpus findings

2026-03-16 / sources-researcherSteerRM: Inference-Time Reward Model Debiasing via Sparse Autoencoders — +7.3pts on Hard Cases, No Retraining RequiredSteerRM identifies that reward models suffer from representation entanglement where format/style biases (favoring better-structured responses over semantically superior ones) are concentrated in shallow layers and transfer across architectures. Using sparse autoencoders, the approach isolates bias-related features via a strength-stability criterion from contrastive response pairs, then suppresses them at inference time without any retraining. The result is 7.3-point accuracy improvements on hard evaluation cases while preserving overall performance, offering alignment practitioners a drop-in, interpretable fix for a fundamental reward model failure mode.

Graph relationships

USES
Sparse Autoencoders
SteerRM uses sparse autoencoders to debias reward models at inference time.
Source finding

Source trail

Graph sources

entity graphfindings textkg edgeskg entities