Procedural Memory Distillation (PMD) advances reinforcement learning by transforming experiences across episodes into actionable intelligence, outperforming previous models like SDPO by 3.8-5.5% on SC…
Researchers discovered that reasoning models like o1 and R1 often overthink, computing answers at around 30% of their chain-of-thought but continuing for the remaining 70%. The termination decision is…
NVIDIA released Nemotron-Labs-Diffusion, a single model that eliminates the need for a separate draft model in speculative decoding, achieving 6.82 accepted tokens per forward pass in self-speculation…
Researchers introduced DominoTree, a training-free best-first draft tree method for speculative decoding that uses Domino's conditional correction to achieve up to 6.6x speedup over autoregressive dec…
Researchers propose a three-phase pipeline to improve small language models for low-resource programming languages like Julia and Ballerina, achieving up to +14.2 points on pass@1 metrics while reduci…
A developer benchmarked four Chinese LLM families—DeepSeek, Qwen, Kimi, and GLM—over 30 days using 1,247 prompts across code generation, reasoning, Chinese language, English chat, and vision tasks. De…
A backend engineer at an unnamed company slashed their team's LLM API costs from $11,400 to $1,830 per month by switching to cheaper models for most tasks and implementing tiered routing. The team rep…
A developer cut their company's AI API bill by 95% from $11,000 to under $400 per month by implementing per-request model routing and tiered escalation. The team replaced expensive GPT-4o calls with c…
New research from arXiv reveals that larger language models consistently outperform smaller ones on reasoning benchmarks by an average of 6-7%, with the key advantage being constraint-guided reasoning…
UC San Diego researchers developed DFlash, a speculative decoding method that uses a lightweight block diffusion model to draft entire token blocks in parallel, achieving up to 6.08x speedup on Qwen3-…
Researchers at SPAR argue that language models can report on their own internal activations, a method called self-report, which is underexploited compared to external analysis tools. They test five op…
JetSpec, a new speculative decoding method, trains a causal parallel draft head over fused hidden states from a frozen target model, enabling lossless verification of candidate trees in one forward pa…
Researchers introduced JetSpec, a speculative decoding method that trains a causal parallel draft head on a frozen target model to draft entire speculative trees in one pass while preserving autoregre…
Researchers trained Qwen3-8B natural language autoencoders with varying length penalties and found that explanation length can be significantly reduced without harming reconstruction fidelity, suggest…
Anthropic developed a method to read a model's internal state as plain English. Groq's founder argued that cheaper AI only increases compute demand. GLM-5.2 released an MIT-licensed model with near-fr…
Researchers introduced PragReST, a self-supervised framework that uses counterfactual reasoning to improve large language models' pragmatic language understanding. Across four benchmarks, PragReST boo…
Researchers introduced Regressive Plasticity Schedule (RPS), a two-stage post-training schedule that couples a learning-rate drop with a curriculum boundary between easier and harder data. Testing on …
Amazon Web Services announced container image caching for Amazon SageMaker AI inference, reducing end-to-end latency by up to 2x for generative AI models during scale-out events by eliminating contain…
Researchers propose ASAG, a training-free method that monitors attention distributions to detect when a reasoning model has reached a conclusion, stopping generation early. Applied to DeepSeek-R1-Dist…
Researchers introduced PERIA, a tool-augmented visual agent designed to improve spatial reasoning in vision-language models by actively acquiring evidence and performing multi-step visual interactions…