IFBench

Headlines · 2

1. Oct 14
   AutoSFT: a multi-agent system that automates SFT data research

   A new arXiv paper introduces AutoSFT, a multi-agent system for automated data research in supervised fine-tuning. The authors report that a single agent given the full training stack converged on only one action — adjusting category proportions — while its context filled with execution output and was discarded mid-search. AutoSFT instead uses a meta agent that separates execution from analysis, gathers proposals from independent angles and calls a synthesis subagent for capabilities the data pool cannot cover. The authors report that the executor halves context loss, the analysis subagents lift the best IFEval by 6.5 points, and synthesis adds 3.4 IFEval and 2.9 IFBench points; on code tasks, analysis adds 3.7 MBPP+ over a single agent that does not improve. Every result comes from runs that retrain from the base model each round, with model, trainer and budget fixed within each comparison.

   Hugging Face · Papers · 🔥 0
2. Oct 13
   Paper proposes CALIF to curb reward hacking in verifiable instruction following

   An arXiv paper submitted on 14 October, "Diagnosing and Correcting Reward Hacking in Verifiable Instruction Following," argues that verifiable instruction-following benchmarks emphasize output-constraint satisfaction and can overlook whether a response actually answers the underlying request, creating an evaluation blind spot; the authors report that frontier models can receive full credit for responses with little substantive answer content. They also report that using the same verifier-based criterion as a training reward can induce reward hacking, with publicly released checkpoints showing higher constraint satisfaction but lower response quality after reinforcement learning with verifiable rewards. To address this, they introduce CALIF, a lightweight method that adds a calibrated likelihood-based bonus to verifier rewards, computed with the existing frozen reference policy and requiring no additional judge or learned reward model; across 1.7B to 8B models in reasoning and non-reasoning modes, they report improved IFBench constraint satisfaction with response quality near or above pre-RL levels.

   Hugging Face · Papers · 🔥 0

Experience and discussion from the community

Share my IFBench experienceAsk about IFBench