Large language model (LLM) news — GPT-4, Claude, Gemini, Llama, Mistral and the latest research on training, fine-tuning, RLHF, and deployment of LLMs.
Researchers found that unlearning knowledge from large language models causes collateral damage that decays with semantic distance but persists across domains. They developed a method to audit this damage before unlearni…
Researchers propose a unified framework for customizing and deploying LLM-based multi-agent systems in enterprise settings, combining model adaptation techniques with inference optimizations that achieve a 4.48x throughp…
Researchers introduced Steerable Cultural Preference Optimization (SCPO), a reward model training algorithm that balances diverse cultural preferences in large language models. SCPO improved minority reward model perform…
Researchers introduced PragReST, a self-supervised framework that uses counterfactual reasoning to improve large language models' pragmatic language understanding. Across four benchmarks, PragReST boosted accuracy by up …
Researchers introduced Gaussian Mixture Attention (GMA), a probabilistic attention mechanism that replaces pairwise query-key comparisons with routing through K learned Gaussian mixture components, achieving linear-time …
Researchers introduced PROPEL, a framework that trains task generators for reinforcement learning by using a lightweight activation probe to predict solver pass rates, avoiding costly solver rollouts. In tests on math, c…
Researchers propose a structural pruning framework for Mixture-of-Experts (MoE) models that reformulates prune-ratio allocation as a channel-score coverage maximization problem, solved via attribution-based approximation…
Researchers propose SAGE, a post-hoc method to sanitize the final update vector in LLM unlearning, reducing the trade-off between forgetting and retention without rerunning the original pipeline. SAGE uses spectral activ…
Researchers introduced RegimeVGGT, a training-free accelerator for Visual Geometry Grounded Transformer (VGGT) that achieves a 6.7x speedup over VGGT* at matched reconstruction quality by applying layer-wise U-shaped com…
Researchers introduced Consensus Frame GRPO (CF-GRPO), a process-level reward framework for video multimodal large language models that uses consensus frame alignment to improve evidence-aware reasoning without temporal …
Researchers at arXiv published a retrospective on SWAVE, a complex-valued recurrent language model, revealing that its Resonance Head suffered from a failure mode called cos-domination collapse, which was resolved by ado…
Researchers introduced Self-Consistency Training with Reinforcement Learning (Self-CTRL), a method that aligns language models' self-explanations with their actual behavior. In tests, the approach improved correlation be…
Researchers propose Data-Forcing Distillation (DFD), a post-training framework that restores diversity and fidelity in few-step video generation models by addressing mode collapse and over-saturation artifacts. DFD achie…
Researchers introduced CHIEF, a human-AI co-creation framework for iterative video generation that places creators at the center, using persona-conditioned multimodal LLMs to provide subjective audience feedback. The sys…
A new open-source benchmark platform called AIStupidLevel continuously monitors 21 production AI models from 7 providers for performance degradation. The platform runs 24/7 tests including coding, deep reasoning, tool ca…
Parcle, a shared memory layer for AI agents, reduces token consumption by over 60% on agentic tasks by eliminating repeated context retrieval. The system indexes operational context and allows agents to retrieve only rel…
A developer detailed a 7-layer prompt-injection challenge, explaining how each defense layer works and where it breaks. The core issue is that LLMs have no hard boundary between instructions and data, making prompt injec…
The Shanghai Stock Exchange clarified IPO rules for unprofitable AI model developers, allowing listings on the Star Market for firms with an anticipated market cap of at least 4 billion yuan. The move aims to support cas…
QuantaMind built the 'Quant Audit' feature to measure how quantization degrades AI agent reasoning, warning that leaderboard scores misrepresent real-world performance. The tool systematically evaluates performance drop-…
Microsoft has sold OpenAI models to major Chinese technology companies including ByteDance, Ant Group, Meituan, and Tencent Holdings, with ByteDance on track to spend over $1 billion annually on Microsoft AI and cloud se…