Multi-Token Residual Prediction (MRP) is a lightweight module that accelerates diffusion language models by predicting logit residuals between consecutive denoising steps without re-running the backbone. Tested on SDAR 1.7B–8B, MRP achieves 1.42× lossless speedup in speculative decoding on reasoning and code generation benchmarks.
PROWL introduces a KL-constrained adversarial curriculum to improve robustness of action-conditioned video world models. A policy exposes high-error trajectories of a diffusion-based model while a Prioritized Adversarial Trajectory (PAT) buffer re-ranks data by prediction error and learning progress. Evaluation on MineRL demonstrates improved robustness on out-of-distribution trajectories.
LLMs fail to identify and share statistical strength between distinct presentations of the same concept (same facts in English and Swahili, functions in Python and Haskell). They create redundant parallel internal representations, saturating model capacity. Small models exhibit near-total compartmentalization in early multilingual learning.
Google released Gemini 3.5 Flash at general availability during Google I/O. The model (ID: gemini-3.5-flash) supports 1,048,576 input tokens and 65,536 max output tokens, with knowledge cutoff January 2025. Deployed across Google Search, Gemini app, and via API for developers and enterprises. Pricing increased versus previous Flash versions.
llm-gemini 0.32a0 released. Compatible with llm>=0.32a0 alpha. Adds ability to stream reasoning tokens.
Google restructures AI subscriptions with three tiers ($7.99–$99.99/month) based on compute consumption instead of daily limits. Launches Gemini Omni and Gemini Spark agent at I/O 2026.
Google announces at I/O three new models: Gemini 3.5 Flash, Gemini Omni (multimodal), and Gemini Spark, a personal agent running continuously in the cloud. The Gemini app receives a major redesign.
A developer built a minimal ML compiler in pure Python/CUDA without external dependencies. It lowers transformers (TinyLlama, Qwen2.5-7B) through 6 successive IRs down to CUDA kernels. On RTX 5090, achieves 0.96× PyTorch production stack performance, with 32/84 kernel shapes beating hand-optimized kernels (up to 5.6× speedup).
Andrej Karpathy joins Anthropic as a senior researcher. The former Tesla AI co-founder and prominent machine learning figure joins the company's research team.
Andrej Karpathy, prominent AI researcher and former OpenAI core team member, joins Anthropic to focus on frontier LLM research. He views the coming years as "especially formative" for the field.
Andrej Karpathy joins Anthropic as Senior Research Scientist. The former Tesla AI co-founder and director of AI at Tesla brings deep learning and AI systems expertise to the company.
NVIDIA Labs releases Sana, a linear diffusion transformer for efficient high-resolution image synthesis. Architecture reduces computational complexity while maintaining visual quality.
Anthropic expands Claude Managed Agents with self-hosted sandboxes and MCP tunnels. Companies can now execute their AI agents' tools in their own infrastructure while keeping agent control with Anthropic.
ESI-Bench is a benchmark for embodied spatial intelligence spanning 10 task categories on OmniGibson. Agents must combine perception, locomotion, and manipulation to actively accumulate evidence. Experiments show active exploration outperforms passive approaches, but failures stem primarily from poor action choices rather than weak perception.
SWIM aligns vision-language representations for fine-grained video object understanding from text prompts alone. Uses mask supervision during training to guide cross-modal attention. Constructs NL-Refer dataset with precise natural language referring expressions. Outperforms visual-prompt-based methods on fine-grained benchmarks.
CheckSupport is an open-source system using locally-deployed LLMs to automate reporting checklist recommendation and completion for scientific manuscripts. Evaluated on peer-reviewed manuscripts, it achieves 90% accuracy for checklist recommendations and 88% for item-level completion, processing each manuscript in 12.5 seconds on CPU-only hardware.
MoleCode is an LLM-native molecular language representing molecules as explicit graphs with typed entities and direct relations, replacing implicit SMILES strings. Training-free, it improves frontier LLMs on molecular reasoning, editing and generation tasks, especially for unfamiliar molecules, topology-sensitive operations and larger structures.
CrossView Suite introduces CrossViewSet (1.6M multi-view samples), CrossViewBench (evaluation benchmark), and CrossViewer (three-stage framework: Perception → Alignment → Reasoning) to enhance cross-view spatial reasoning in MLLMs. A multi-agent data engine generates annotated data covering 17 fine-grained task types.
RAC (Robust Agent Compensation) is a log-based recovery paradigm integrated into agent frameworks (LangGraph, LangChain) to prevent unintended side effects. Requires no code changes. Results: 1.5-8X better latency and token efficiency vs state-of-the-art LLM recovery approaches on τ-bench and REALM-Bench.
UniversalRAG is a multi-modal RAG framework that retrieves and integrates knowledge from heterogeneous sources (text, images, videos) at variable granularities. It introduces modality-aware routing to avoid intra-modal bias and organizes each modality into granularity levels. Validated on 10 benchmarks, it outperforms single-modality and unified baselines.
New GPRL (General Preference Reinforcement Learning) method replaces scalar reward models with General Preference Model (GPM) using k skew-symmetric subspaces. Tested on Llama-3-8B-Instruct: 56.51% win rate AlpacaEval 2.0, outperforms SimPO and SPPO on Arena-Hard, MT-Bench, WildBench by preventing single-axis reward hacking.
Systematic study of code as AI agent infrastructure. Three layers: harness interface (code connects reasoning, action, environment modeling), mechanisms (planning, memory, tool use, feedback), and multi-agent scaling. Applications: coding assistants, GUI/OS automation, embodied agents, scientific discovery, DevOps.
Guided Topology Diffusion (GTD) uses graph diffusion models to dynamically generate optimal communication topologies for multi-agent LLM systems. The iterative framework, guided by a proxy model predicting multi-objective rewards (accuracy, utility, cost), adapts topologies to tasks without gradient-based optimization, outperforming static approaches.
DoublyCal, a double-calibration framework, improves LLM reliability by quantifying epistemic uncertainty in retrieved evidence and reasoning. A lightweight proxy model generates Knowledge Graph evidence with calibrated confidence, guiding a black-box LLM toward more accurate and well-calibrated predictions.
ClawGym is a framework for developing agents capable of executing multi-step workflows over local files and persistent tools. Authors construct ClawGym-SynData (13.5K synthesized tasks), train ClawGym-Agents via supervised fine-tuning and RL, and propose ClawGym-Bench (200 instances) for evaluation.
Study of adversarial action removal attacks in self-play reinforcement learning. An attacker selectively masks legal actions from the victim's action set. Experiments on poker (6 to 5,531 states) and two non-poker domains: learned masking causes substantially more damage than random masking, persists across Q-learning/PPO/NFSP/DQN, transfers between agents, and is amplified by self-play.
TABOM, a post-training method for Diffusion Language Models, aligns optimization with the multi-step easy-to-hard decoding trajectory observed at inference. Via Boltzmann modeling of unmasking preferences, it derives a tractable pairwise ranking objective that reduces training-inference discrepancy and improves performance on new domains.
arXiv paper proposing a formal statistical framework to combine LLM and human evaluations. Uses a doubly robust estimator (missing data approach) to determine optimal sample sizes of human ratings needed for benchmark validation, based on LLM judgment predictability.
Benchmark of LLMs on multi-label legal precedent treatment classification. Expert-annotated dataset of 239 real-world citations. Gemini 2.5 Flash achieves 79.1% on high-level classification, GPT-5-mini 67.7% on fine-grained schema. Novel Average Severity Error metric to measure practical impact of misclassifications.
arXiv study demonstrating that analytic choices (model selection, sampling parameters, prompt format, demographic data) materially affect the fidelity of "silicon samples" (synthetic datasets generated by LLMs). Across 252 configurations tested, correlations with human data range from r=.23 to r=.84, revealing a major risk of analytic flexibility.
SkyPart, a lightweight head for vision transformers, improves cross-view drone-to-satellite geo-localization by explicitly separating layout and texture via learnable prototypes. At 26.95M parameters, it achieves state-of-the-art on SUES-200, University-1652, and DenseUAV with enhanced robustness under weather corruptions.
Researchers introduce IBPO (Implicit Behavior Policy Optimization), a credit assignment method for reinforcement learning with LLMs. By comparing multiple reasoning trajectories, the framework transforms sparse terminal rewards into step-sensitive learning signals, reducing gradient variance and improving stability on mathematical and code reasoning benchmarks.
ECHO is a diffusion-based model for chest X-ray report generation. It uses Direct Conditional Distillation and Response-Asymmetric Diffusion training to enable one-step-per-block inference, achieving 8× inference speedup while improving RaTE (+64.33%) and SemScore (+60.58%) over autoregressive baselines with negligible clinical accuracy loss.
An arXiv study challenges the assumption that Mixture of Experts models achieve domain specialization through sparse routing. The COMMITTEEAUDIT framework reveals a domain-invariant "Standing Committee"—a compact coalition of experts capturing most routing mass across domains, layers, and budgets. Peripheral experts handle domain-specific knowledge alone.
Paper introducing CAT (Context-Agent-Tool), a data layer for managing multi-agent teams. Applies ALARA principle (as low as reasonably achievable exposure) to context. Evaluates 22 models (0.6B–35B parameters) on 115 practical tasks via npcsh, a CLI shell. ~2500 executions test file operations, web search, multi-step scripting, tool chaining, and inter-agent delegation.
Theoretical paper on recovering latent actions and environment dynamics from offline trajectories without action observations. Authors leverage demonstrator diversity (each following distinct policies) to identify latent transition kernels via nonnegative matrix factorization. Identifiability proven under rank and policy diversity conditions.
CounterRefine is a lightweight repair layer for RAG that treats the first answer as a hypothesis to test. The system issues answer-conditioned expansion queries to retrieve candidate-specific evidence, then applies a deterministically-validated KEEP/REVISE refinement step. On SimpleQA, it improves a matched one-pass RAG baseline by up to 5.8 correct-rate points.
Novel VAE method using spherical shell geometry and cluster-aware constraints to mathematically eliminate posterior collapse. Transforms data to spherical shell, applies K-means, defines feasible region between within-cluster variance and collapse loss. Guarantees 100% collapse prevention without computational overhead or decoder output restrictions.
DARC is a retraining-free inference-time method that reformulates response selection as distributionally robust optimization under annotator disagreement. It reranks candidates by maximizing KL-robust satisfaction objectives, with deployment controls to cap entropic risk premium without retraining.
WASIL is a dataset of real-world Arabic voice interactions with LLMs: 8,529 turns with audio, ASR hypotheses, assistant responses, and like/dislike feedback (14.2% dislikes). Includes 2,000 test turns covering Modern Standard Arabic and 4 major dialects. Answerability annotations separate ASR errors from intrinsic limitations.