Paper
Hugging Face
DRACO redistributes a single per-trajectory rubric score into differentiated per-step advantages for long-horizon agent training, gaining 15.9 points over a base model on AppWorld without using any verifier.
AppWorld gain over base model +15.9 ptsGain over sparse-reward GRPO +5.3 pts
Why it matters
- Addresses a real gap: most long-horizon agent tasks lack programmatic success checkers, so credit assignment from sparse rubric scores has been unreliable.
- Gains transfer out-of-domain to Tau-Bench even without a frontier judge, suggesting the method isn't overfit to one benchmark.
Paper
Hugging Face
Analyzing RLVR on the Countdown task, researchers find reasoning diversity is lost almost entirely at the first token of a solution rather than during downstream reasoning, and a targeted checkpoint-interpolation fix recovers 37% more solution coverage at no cost to accuracy.
Solution coverage drop observed up to 67%Coverage recovered by fix +37%
Why it matters
- Explains a known side effect of RL fine-tuning (reduced test-time-scaling returns) with a specific, actionable mechanism rather than a vague diagnosis.
Paper
Hugging Face
VeriPhy compiles a video prompt into typed physical obligations and checks generated clips against them using frozen perception experts, catching more real generation failures than a published question-decomposition evaluator on the same clips.
Failures caught (of 304) 228 vs. 164 baseline
Paper
Hugging Face
PACE is a new benchmark testing whether assistants can spot user requests that look reasonable but conflict with facts the assistant would need to retrieve from a knowledge base, plus a multi-agent retrieval method that outperforms existing approaches on it.
Paper
Hugging Face
AutoTraceGT automates grounded-theory coding on agent trajectories, recovering 73-91% of human-annotated failure modes across six trajectory corpora and surfacing additional patterns human taxonomies missed.
Failure-mode recovery 73-91%
Paper
Hugging Face
BCIT decides which past post-training update evidence is still safe to reuse after a model has changed, authorizing fewer harmful updates and reaching higher equal-budget quality than alternatives in a 4B-model case study.