Read this because Builders can simplify training infrastructure by replacing complex multi-stage pipelines with a single unified trainer.
Researchers demonstrate that foundation model training stages can be unified under a single likelihood-based algorithm.
Read this because This work addresses the critical gap in high-quality training data for multimodal coding, offering open models that rival proprietary systems through decomposed rubric synthesis.
PrismaCoder-9B and 35B models leverage a novel rubric-based data pipeline to lead open multimodal code intelligence benchmarks.
Read this because This demonstrates that specialized fine-tuning combined with iterative inference loops can push open-weight models to elite competitive levels without external tools.
NVIDIA fine-tunes Nemotron models to reach gold-level results in both the International Olympiad in Informatics and International Mathematical Olympiad.
Read this because Developers can improve model alignment and reward scores without expensive fine-tuning or parameter updates, preserving original model capabilities.
New MISVO method steers frozen language models via pre-logit vectors, balancing reward optimization with generation quality using local KL geometry.
Read this because Separating test generation from code repair prevents error agreement and false confidence, significantly improving agent reliability.
ExecCritic uses separate test and repair agents with reinforcement learning to boost coding agent success rates on SWE-bench Verified.
Read this because Engineers can improve LLM reasoning capabilities at test time without relying on expensive or unavailable ground-truth labels.
TTPO uses majority-vote pseudo-labels to match supervised performance on competition-level math benchmarks without ground-truth data.