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 propose Visual-Seeker, a multimodal deep search agent that actively reasons over visual details to perform multi-hop, cross-modal search. The agent achieves state-of-the-art results on five benchmarks, surpas…
Researchers introduced Mask-Proof, an LLM-based automated data curation pipeline that converts mathematical proofs into checkable masked-step tasks, creating the Mask-ProofBench benchmark with 292 problems. Testing 17 mo…
Researchers at UC Riverside introduced the first robustness study for proof autoformalization in Lean 4, testing seven LLM-based models on perturbed informal proofs. All models showed sensitivity to global paraphrasing a…
Researchers at arXiv propose Telegraph English, a readable symbolic format for context compression that rewrites retrieved passages into structured entity-relation statements. In experiments on multi-hop QA datasets, it …
Researchers propose AC-GPT, a modification to causal Transformers that enables evaluating and sampling from arbitrary conditionals in a single forward pass without degrading standard left-to-right performance. The method…
Researchers introduced CoRA, a GRPO-based reinforcement learning framework that aligns LLM confidence with rationale quality in chain-of-thought reasoning. Across MedQA, MathQA, and OpenBookQA, CoRA reduced confidence-ra…
NVIDIA released Nemotron 3 Ultra, a 550B-parameter hybrid Mamba-Transformer model with 55B active parameters, achieving up to 6x higher inference throughput than state-of-the-art LLMs while maintaining accuracy. The mode…
A new study finds that online web agents augmented with memory, workflow, or skill modules often fail to outperform a token-matched vanilla baseline under a fixed inference budget. Testing across three models and multipl…
A new empirical study systematically compares tokenizers for multilingual large language models across 11 Southeast Asian languages, finding that Parity-aware BPE achieves the best balance of compression efficiency and c…
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-Distill and Qwen3 models…
Researchers introduced Ling-2.6 and Ring-2.6, a family of models designed for efficient and scalable agentic intelligence at trillion-parameter scale. Ling-2.6 optimizes for instant response generation, while Ring-2.6 fo…
Researchers introduced PanicCognitivePath (PCP), a framework combining cognitive graphs with large language models to predict panic emotional arousal timing before manifestation. PCP uses a Psychological Safety Distance …
Researchers introduced AgentViSS, a benchmark for evaluating visual social intelligence in multimodal agents, featuring 240 scenarios and four role-level tasks. Tests on seven multimodal large language models revealed th…
Researchers propose QPILOTS, a method that steers flow-matching and diffusion policies at inference time by projecting noisy intermediate states to clean action estimates for critic gradient computation. QPILOTS achieves…
Researchers propose GRASP, a method for multi-source transfer learning that achieves 93.5% mean accuracy across benchmarks while using constant memory, compared to 71.7% for ensemble methods. GRASP sequentially merges so…
Researchers reformulated the autoresearch pattern as a stateful ReAct agent using LangGraph, achieving 90% fewer tokens on hyperparameter tuning and 52% fewer on code optimization compared to stateless designs. The state…
Researchers introduced FASTMIX, a framework that automates data mixture optimization for training large models by reformulating mixture selection as a bilevel optimization problem and using gradient descent to jointly op…
Researchers trained a two-layer transformer encoder to classify rational elliptic curves by rank, achieving over 99% accuracy. The model learned the Mestre-Nagao heuristic from data alone, with input weights matching the…
Researchers introduced a temporal difference (TD) objective for diffusion models that enforces cross-time consistency along the denoising trajectory, improving sample quality especially with few sampling steps. The metho…
Researchers introduced AdaNAGED, a zero-order parameter-free optimization method for LMO-based fine-tuning of large language models, addressing memory overhead and sensitivity to hyperparameters. The method unifies gradi…