This week’s news from Zed, Anthropic, and OpenRouter shows why better harnesses matter more than better models
thenewstack.io·13h ago
AI Research
arXiv CS.AI·May 25·8 min read
EDGE-OPD: Internalizing Privileged Context with Evidence Guided On-Policy Distillation
TL;DR
Training large language models with privileged information (extra context unavailable at inference) causes unintended side effects like altered reasoning and degraded general capabilities. EDGE-OPD (Evidence Guided On-Policy Distillation) addresses this by using guided rollouts to ensure target behavior appears in training data and applying an evidence mask to update only tokens supported by privileged context.
✦ Why It Matters
Engineers can use EDGE-OPD to train models with privileged information without degrading general performance or introducing unintended behavioral changes.
Key Takeaways
- Utilize EDGE-OPD for improved model training with privileged context.
- Incorporate guided rollouts to enhance behavior representation.
- Apply evidence masks to focus updates on relevant tokens.
- Explore localized persona signals for efficient knowledge transfer.
How It Works
EDGE-OPD modifies the traditional OPSD approach by integrating guided rollouts, which inject privileged context into the training data at sampling time. This ensures that the rare target behaviors are present when the model learns.
Additionally, the evidence mask selectively updates the model only at token positions where the privileged context supports the sampled token, preventing unwanted side effects on model behavior.