Rituraj
Tag: research
3 items with this tag.
- 2026Dense Supervision Is Not Enough: The Readout Blind Spot in Looped Language ModelsDense cross-entropy trains exits, but it does not control recurrent hidden-state scale when every readout is…
- 2026The Master Key Hypothesis: Unlocking Cross-Model Capability Transfer via Linear Subspace AlignmentCapabilities live in a low-dimensional linear subspace and transfer across model scales — training-free and label-free.
- 2026PRISM: Pushing the Frontier of Deep Think via Process Reward Model-Guided InferencePRM-guided inference for Deep Think — step-level verification for population refinement and solution aggregation.