EMO alternates two steps per question.
Alternating context edits and logit steering boosts frozen LLM accuracy
EMO closes the loop between feedback-driven context revision and decoding-time distribution correction, improving zero-shot performance across eight models and six benchmarks.
Big Tech
Wei Xia · Yehui Tang · Zhi-Hong Deng
Peking University · Samsung Research
Research Digest··2 min read
Thread:Hidden-State Steering
The authors propose EMO, a training-free framework that alternates question-specific context optimization (E-step) with confidence-constrained logits steering (M-step) while model weights stay frozen.
Why this paper
From Samsung Research and Peking University · Part of Hidden-State Steering, now 20 papers
In one line
EMO improves frozen language-model reasoning by alternating judge-guided context revision with confidence-constrained steering derived from context-induced logit changes.
What we could check
- ·No code link found
- ·No weights link found
- ·No dataset link found
- ·No compute details found
- ✓Limitations stated by the authors (4 noted)
- ·No benchmark numbers found
Observed from the paper text and links we have. Absence here means we did not find it, not that it does not exist.
§