CriticGen: Generation-Aware Evaluation as Actionable Feedback

arXiv:2609.05439v1 Announce Type: new Abstract: Current evaluation methods for large language models are coarse-grained and decoupled from generation, producing generic explanations that fail to provide actionable feedback for model improvement. We propose CriticGen, a fine-grained, generation-aware evaluation framework that turns evaluation into actionable control for answer improvement. CriticGen first generates sample-specific evaluation dimensions and scoring criteria under high-level categories such as subjective, objective, and self-derived constraints. These criteria then serve as a dynamic rubric for jointly producing a score, a reason, an executable refinement suggestion, and a refined answer. This rubric-conditioned refinement process enables models to diagnose flaws and perform targeted answer improvement. Experimental results show that fine-grained evaluation should be both instance-specific and actionable. CriticGen induces higher-quality rubrics, improving relevance/coverage from 3.33/4.03 to 3.97/4.24. CriticGen also achieves the best score correlations, with 0.9556 Pearson and 0.9560 Spearman, and raises the F1 of criterion-grounded reasons and executable suggestions from 0.6369/0.5994 to 0.7554/0.7900. Crucially, its feedback translates into reliable answer improvement, improving 73.17% of answers with a 93.28% non-degradation rate.

CriticGen's evaluation framework improves model performance and answer quality.

CriticGen offers actionable, fine-grained evaluation feedback that improves model answer quality, which can enhance AI utility and reliability. However, it indirectly raises concerns about over-reliance on automated self-improvement and evaluation, possibly embedding biases or errors without human oversight. Evidence is solid but effects are incremental rather than transformative.

This is an incremental technical improvement to LLM self-evaluation and refinement methodology, with measurable benchmark gains. It has modest positive implications for AI reliability and quality control but no broad societal claim beyond model performance metrics.

SOURCE     arXiv cs.AI (tier 1)
   ↓
DOCUMENT   165589c9-94e4-419e-b378-182d1507ae01
           https://arxiv.org/abs/2609.05439
   ↓
EVIDENCE   2 extracted excerpts
   ↓
MODEL RUN  3 runs, methodology 1.0.1
   ↓
SCORE      +34  (Favourable)
   ↓
CONFIDENCE 78%