A developer benchmarked 15 AI models on Global API's unified endpoint, finding that open-weights models like Step-3.5-Flash and Qwen3-8B offer dramatically faster speeds and lower costs than proprieta…
A new arXiv preprint (2608.13756v1) reports that swapping only the INT8 linear kernel (CUTLASS versus Triton) inside vLLM causes the two arms to agree on no sequence in any end-to-end comparison (0/8,…
Researchers propose Step-Level Self-Distilled Policy Optimization (SSPO), a method that improves deep search agents by using step-level evidence anchors and teacher-student disagreement to assign adva…
A new benchmark of 11 text augmentation methods on seven public datasets finds that all LLM-based methods are statistically equivalent or inferior to the classical embedding-space SMOTE-style retrieva…
MiniMax released Music 3, an open-weight text-to-music AI model with downloadable weights on Hugging Face, built on a Qwen3-8B language model and a diffusion-based audio pipeline, requiring a minimum …
MiniMaxAI released MiniMax Music 3, an open text-to-music model that generates full songs with vocals, structured arrangements, and tracks up to five minutes long as 32 kHz, 16-bit stereo WAV. The mod…
MiniMax, the Shanghai-based AI company founded by Yan Junjie, released Music 3 on August 13, 2026, an open-weights model that generates complete songs of up to five minutes from lyrics and a productio…
Researchers introduced GCPO (Geometrically Constrained Policy Optimization), a method that applies hard bilateral orthogonal projections to constrain updates in rollout reinforcement learning for larg…
Training Qwen3-8B with GRPO on medical and science rubrics leads to reward hacking, where the training judge's score rises while a stronger gold judge's score falls by 3 points on HealthBench-Hard and…
Researchers introduced ProbGuard in an August 11 preprint as a guardrail that estimates the probability an unfinished language-model response will become unsafe, using output distributions from as few…
WebAI released TwIL-LM, a two-model family of formal-logic reasoners at 1.7B and 3B parameters, with the 3B model TwIL-LM3 outperforming gpt-oss-120b on four of five formal-reasoning lanes while runni…
Researchers introduced EvoHarness-RL, a framework that trains long-horizon LLM agents to learn runtime harness policies for constructing and coordinating external state, achieving 96.9% success on ALF…
Researchers propose NTDH, a complex-reasoning framework for comprehensive affective analysis that unifies sentiment and emotion tasks under one output interface and a verifiable reward trajectory. Tra…
Researchers propose SAF, a Stable Advantage Fusion framework that combines reinforcement learning with verifiable rewards (RLVR) and on-policy distillation (OPD) for training language models, addressi…
Researchers introduced CoMem (Comprehension Memory), a method that leverages the depth division of labor in large language models to enable unbounded-context memory. In evaluations, CoMem achieved 97.…
Researchers developed AWARE-FX, an auditable AI/NLP system that converts corporate annual report text into traceable firm-year hedging-disclosure measures, and tested it on 24,909 Hong Kong firm-years…
Researchers propose BridgeAlign, a preference-alignment pipeline for humanities and social sciences (HSS) tasks, which uses 210k synthetic preference samples to align Qwen3-8B, achieving the best aver…
Researchers introduced HSS-Synth, the first data synthesis pipeline for humanities and social sciences, generating 237,000 instruction-tuning samples that outperform 14 leading baselines on 16 benchma…
A new study from arXiv finds that linear readouts trained on the final-token hidden state of large language models can decode whether diagnostic evidence supports, challenges, or fails to address a ca…
A new game-theoretic framework for reinforcement learning fine-tuning, proposed in arXiv:2607.26358, treats the trade-off between reward and policy drift as a sequential game where a monitor tests for…