Large Language Model News

Researchers introduced MA-ProofBench, the first formal theorem-proving benchmark for mathematical analysis, containing 200 problems across two difficulty levels. Evaluations of leading LLMs showed poor performance, with …

Researchers propose using hyperdimensional computing (HDC) for structured querying on tabular data embeddings, addressing the lack of interpretable similarity scores in current methods. The HDC approach, based on Hologra…

Researchers introduced Risk-Aware Causal Gating (RACG), a framework that gates decisions from LLM agents based on causal risk rather than predictive confidence, reducing high-cost errors while preserving utility. The met…

Researchers at arXiv have developed a method using LLM representational geometry to predict compositional errors, finding that near-orthogonal concept encodings lead to successful composition while close encodings cause …

Researchers propose MINIM, a trusted local broker that performs privacy-aware minimization on client-side UI observations before transmission to remote LLM inference servers. The system learns dual scores for each UI ele…

The QIAS 2026 shared task, co-located with LREC 2026, evaluated large language models on Islamic inheritance reasoning using the MAWARITH benchmark of 12,500 Arabic cases. Sixteen teams submitted systems employing prompt…

Researchers at CohereLabs argue that large language models suffer from a 'cultural data funnel,' where cultural signals decline sharply during post-training while geographically concentrated data dominates. They release …

Researchers introduced DLawBench, a benchmark for evaluating large language models on multi-turn legal consultation tasks. The benchmark includes 461 cases from Chinese and U.S. law and tests models on four client types;…

A new study finds that post-training techniques, including supervised fine-tuning and reinforcement learning, can transform large language models into effective medical coders for ICD coding, challenging prior assumption…

A new study finds that large language models' preferences and values shift significantly depending on the deployment context, such as whether the model is writing a Reddit post or a news article. Across five LLMs and ove…

Researchers introduced MedLatentDx, a latent multi-agent communication framework that enables cross-hospital rare-disease diagnosis by exchanging compact latent KV blocks instead of identifiable clinical text, improving …

Researchers applied likelihood ratio-based forensic text comparison to Japanese digital texts for the first time, fusing stylometric and embedding systems to estimate authorship likelihood ratios. The fused system improv…

Researchers introduced Compliance Asymmetry (A = BCR/HCR), a bidirectional diagnostic for LLM compliance, and found that models exhibit direction-blind moral compliance—following helpful and harmful nudges at nearly iden…

Researchers introduced llada.cpp, the first NPU-aware inference framework for accelerating diffusion large language models on smartphones, achieving 17x-42x latency reduction over CPU baselines while preserving generatio…

Researchers at arXiv have causally demonstrated that the weight norm sets the grokking timescale in neural networks, settling a dispute over whether weight norm causes the delayed generalization. By intervening on the no…

Researchers identified the double-drift phenomenon causing instability in diffusion policy-gradient methods and proposed DiPOD, a framework that interleaves self-distillation with policy-improving gradient updates to mai…

A new thesis investigates how Bayesian principles can improve understanding of modern deep learning systems, introducing the Deep Variational Implicit Process (DVIP) and post-hoc methods VaLLA and FMGP for uncertainty es…

Researchers propose SuperThoughts, a method that compresses pairs of consecutive Chain-of-Thought tokens into single latent representations and decodes two tokens per step, doubling inference throughput while reducing Co…

Researchers propose Gefen, a memory-efficient optimizer that reduces AdamW's memory footprint by ~8x while maintaining performance, enabling larger microbatches and improved throughput in deep learning training. Gefen au…

Researchers propose the Hybrid Open-Ended Tri-Evolution (HOTE) framework, which uses hybrid-mode reinforcement learning to evolve a proposer, solver, and judge collaboratively for deep research tasks. An 8B model trained…