Researchers built a 128-GPU cluster from retired V100s and ran it for a year, achieving competitive LLaMA-70B throughput at 3.6% the cost of new hardware—but energy consumption makes it viable only in low-cost, low-carbon regions.

Jul 16, 2026

ReleasesResearch

Strong models boost weak ones at test time without retraining

New preprint shows stronger models can build inference-time scaffolds that boost weaker models from 0.49 to 0.91 accuracy on Theory-of-Mind tasks—no parameter updates required.

Jul 16, 2026

ReleasesResearchNSFW

RynnValue trains robot reward models on 7,000 hours without preference labels

Open-source value foundation model replaces preference annotations with temporal distance, trains on 3M clips, and lifts real-world policy success from 52.5% to 72.5%.

Jul 16, 2026

ReleasesResearch

CalibForge tunes 5,431 terminal tasks to solver difficulty, lifting agent benchmarks by 30 points

A new preprint introduces CalibForge, a system that synthesizes executable terminal tasks by calibrating them against solver behavior, yielding up to 30-point benchmark gains on agent training datasets.

Jul 16, 2026

ReleasesResearch

WorldCycle reduces video world model drift 44% using reversible action cycles

A new RL framework uses self-verifiable action cycles to train video world models, reducing state-returning drift by up to 44 percent and quadrupling composite-action accuracy over baseline models.

Jul 16, 2026

ReleasesResearch

Skill-α reinforcement learning method boosts agent task success by 6.7 points

Researchers propose Skill-α, an RL approach that generates agent skills through sequential edits evaluated by a rollback reward, outperforming heuristic baselines by up to 6.7 percentage points on downstream benchmarks.

Jul 16, 2026

ReleasesResearchPlatform

Explorative Modeling pretrains generative models 4.1× faster by exploring K candidates per step

Explorative Modeling, a preprint from researchers at Cornell and MIT, factors the training loop instead of the generation procedure, exploring multiple candidate outputs and training on the best match to improve FLOP efficiency by 4.1× and sample efficiency by 6.2×.

Jul 16, 2026

ReleasesResearch

SkillRise trains LLM agents to build reusable skills across task sequences

A new reinforcement learning framework teaches language model agents to extract and refine skills across progressively harder tasks, outperforming baselines by 2.3 to 8.5 percentage points on ALFWorld, WebShop, and ScienceWorld.

Jul 16, 2026

ResearchPlatform

PAJAMA distills LLM judges into programs, cuts eval cost by 100×

Researchers from UC Berkeley and Wisconsin introduce program distillation for model evaluation, replacing per-sample LLM calls with transparent code that matches 13B judge performance at two orders of magnitude lower cost.

Jul 16, 2026

ReleasesResearch

Molt: NVIDIA's PyTorch framework cuts agentic RL iteration cost

NVIDIA's Molt is a PyTorch-native training framework designed for agentic reinforcement learning, keeping codebases small enough for researchers and AI assistants to trace end-to-end while matching Megatron-based stack performance.

Jul 15, 2026

ReleasesResearch

Hypernetworks outscale LoRA for train-time knowledge injection in LLMs

New arXiv preprint from Dhankhar, Baha, and Saparov shows hypernetworks exhibit steeper scaling exponents than LoRA fine-tuning when injecting factual knowledge into large language models at training time.

Jul 15, 2026

ReleasesResearchPlatform

Staleness-Adaptive Trust Region cuts asynchronous RL performance loss to 3% at 8× policy lag

Researchers propose SAT, a trust-region method that adapts PPO clipping to policy staleness, achieving 34.79 AIME24 avg@8 on Qwen3-30B when rollouts lag by 8 steps.

Jul 15, 2026

Research

Distilled RL transfers knowledge across model families without unconditional imitation

A new preprint proposes Distilled Reinforcement Learning, which combines fine-grained teacher guidance with RL objectives to transfer knowledge across model families more effectively than standard RL or on-policy distillation.

Jul 15, 2026

ReleasesResearch

LongStraw trains RL models at 2.1M tokens on eight H20 GPUs

New execution framework from Zhou et al. bridges the gap between million-token inference and 256K-token RL post-training, demonstrating grouped policy optimization at 2.1M positions on commodity hardware.

Jul 15, 2026

ReleasesResearch

ShortOPD cuts pruned LLM recovery time by 75% while raising generation quality 9×

A new training schedule for compressed language models detects repetitive suffixes early, allocates compute to useful prefixes, and matches full-length distillation in a quarter of the time.

Jul 15, 2026

ResearchPlatform

Lightweight proxy models cut LLM post-training costs while enabling cross-model signal reuse

Researchers propose Proxy-guided Update Signal Transfer, a post-training method that uses lightweight proxy models to discover optimization signals and transfer them to larger primary models, cutting computational overhead while enabling cross-model reuse.

Jul 14, 2026

ReleasesResearch

Soofi S 30B activates 3B parameters per token, tops European AI baselines

Deutsche Telekom's Soofi-Team released Soofi S, a 30B-parameter Mixture-of-Experts foundation model that activates only 3B parameters per token and outperforms European baselines on German and English benchmarks.

Jul 13, 2026

Research

John Baez publishes free Applied Category Theory course with full lecture notes and videos

UC Riverside mathematician John Baez has published complete materials for his Applied Category Theory course online, including lecture notes, exercises with solutions, and video recordings covering categorical approaches to networks, databases, and dynamical systems.

Jul 12, 2026

1 2 3 4 5 6 7 8 9 10 Next →