Almanac
paper

EDIT: Evidence-Diagnosed Intervention Training for Rule-Faithful LLM Grading

paperactiveprovisionaledit-evidence-diagnosed-intervention-training-for-rule-faithful-llm-grading-cacd610a·1 events·first seen 12d ago

Aliases: EDIT: Evidence-Diagnosed Intervention Training for Rule-Faithful LLM Grading

Co-occurring entities

More like this (12)

Recent events (1)

4arXiv · cs.CL·12d ago·source ↗

EDIT framework trains more rubric-faithful LLM graders via internal-state diagnostics

Researchers introduce Evidence-Diagnosed Intervention Training (EDIT), a two-phase framework for improving LLM-based rubric grading. The first phase (EDIT-SFT) identifies problematic reasoning steps using posterior belief signals and input-grounding scores, then revises only those steps with rubric checklists; the second phase (EDIT-RL) uses belief-guided reward shaping to penalize harmful belief drifts during RL. Experiments on two real-world multi-subject grading benchmarks show consistent improvements over SFT and RL baselines on both in-domain and out-of-domain splits.