Chroma's July 2025 study of 18 frontier models found that performance degrades as input length grows, even before hitting hard context limits, and that models perform better on shuffled documents than…
SPARK, a new approach to diagnosing hidden-state failures in AI reasoning, boosted accuracy on the MATH-500 benchmark for Qwen3-4B from 82.0% to 84.6% and for Qwen3-8B from 82.4% to 85.6%, according t…
An independent developer built a pipeline where an AI agent uses reinforcement learning (RL) to train smaller models via RL, achieving a peak reward of ~0.63 over 54 training steps and transferring to…
New experiments with Qwen3 models reveal that continual learning in language models struggles to retain earlier knowledge, with accuracy dropping to 1% for bare-statement facts after 20 updates, while…
Nova, a home AI running locally in France, reduced its first-reply latency from 242 seconds to 11 seconds by disabling unused tool schemas, caching the prompt per session, and moving cold starts to a …
A 5-condition ablation study (3,175 total runs) validates that growth-ratio normalized Lyapunov energy functions serve as precise leading indicators of agent task failure, achieving zero stability vio…
A new signed symmetric quantization method eliminates clipping errors in low-bit AI models without the runtime penalty of asymmetric quantization, achieving up to 2.45 times faster inference on AMD EP…
Researchers introduced TREK (Teacher-Routed Exploration via Forward KL), a new AI training method that enhances learning through unconventional exploration strategies. TREK significantly improved perf…
Researchers introduced TREK, a method that improves AI problem-solving by using verified output trajectories to extend model learning. TREK boosted Qwen3-8B's performance on AIME 2024 from 36.9 to 40.…
Researchers propose Jet-Long, a tuning-free zero-shot method for extending LLM context windows that dynamically adjusts a rescaling factor based on sequence length, preserving short-context fidelity w…
An independent benchmark of on-device LLMs on iPhone A17 Pro found Apple's system model achieves ~149 tok/s with only 12MB peak app memory, while 4B-class open models like Qwen3 4B and Llama 3.2 3B tr…
Researchers released the official Multi-Teacher On-Policy Distillation (MOPD) paper, which composes multiple capabilities into a single large language model by training domain-expert teachers independ…
Researchers propose DeLS-Spec, a speculative decoding method that decouples long- and short-context experts to accelerate LLM inference. By adding a lightweight local head to the existing DFlash model…
Hugging Face announced that its Transformers modeling backend for vLLM now matches or exceeds native vLLM speed for compatible architectures, reducing the need for separate hand-optimized serving port…
Hugging Face announced that the transformers vLLM modeling backend now matches or exceeds native vLLM throughput for many LLM architectures, allowing model authors to run their transformers implementa…
A developer released TurboQuant, a pip-installable adapter for mlx-lm on Apple Silicon that uses a randomized Hadamard transform for data-oblivious, calibration-free quantization of weights and KV cac…
Researchers introduced LearnStop, a cost-aware early-exit method for reasoning language models that learns when to stop computation based on online features like answer confidence and entropy. Across …
A non-programmer named Tim is setting up a local LLM on a MacBook M5 Max with 128GB unified memory, using Docker Desktop with Model Runner, Open WebUI, and models like Gemma 4 and Qwen3 30B-A3B-Q4_k_m…
A developer identified a structural cause for coding agents ignoring system-prompt rules mid-session: attention dilution as context grows. The rule becomes an old, low-weight token among thousands of …
Researchers at the University of Washington released TheoremGraph, a unified dependency graph spanning 18 million+ mathematical statements from arXiv papers and the Lean formal proof assistant. The gr…