ReFormBench: Benchmarking the Stability of Response-Conditioned Format Correction in Large Language Models
We introduce ReFormBench, a benchmark that uses each model's unconstrained response to the same question as its formatting baseline.
ProofPaper ↗
Key points
- Reliable use of large language models requires responses to satisfy formatting requirements across changing instructions.
- Standard compliance scores do not distinguish preserving a model's observed format from correcting it, or reveal whether that correction holds across variants.
- A prompting experiment on Qwen3.5-9B improves both metrics, and the catalog covers over 95% of extracted instances on four additional models.
- ReFormBench makes response-conditioned correction across instruction variants a measurable dimension of format following.
Sources (1)
- [1]ReFormBench: Benchmarking the Stability of Response-Conditioned Format Correction in Large Language ModelsarXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 14, 07:16 AM
We introduce ReFormBench, a benchmark that uses each model's unconstrained response to the same question as its formatting baseline.
Reliable use of large language models requires responses to satisfy formatting requirements across changing instructions.
Extractive summary: sentences quoted from the sources.
Before this
- Oct 14, 2026Who's Behind the Text? Robust Model Authorship Attribution
- Oct 14, 2026Beyond Corrected Actions: Training LLM Agents on Textual Critiques
- Oct 14, 2026ReaDy: Recurrent States as Dynamic Context for Speculative Drafting in Hybrid Linear Attention Large Language Models
- Oct 14, 2026Sparse Autoencoders Are a Shared Coordinate System for Symbolic Matching
- Oct 13, 2026ListQA: A Benchmark for Evaluating List-Formatted Factual Knowledge Retrieval in Large Language Models
- Oct 13, 2026GeoReader: Learning Latent Geometry for Spatial Reasoning in Vision-Language Models
