AI Research Advances: Reasoning, Science, and Benchmarks

Confluence Labs scores 97.9% on ARC-AGI-2 using program synthesis + LLM orchestration, open-sourced, applied to drug discovery. First AI-designed 'universal vaccine' enters human trials. MiniMax-M3 beats GPT-5.5 at 5-10% cost. ESMFold2 predicts 1B+ protein structures. OpenAI's model disproved 80-year Erdős conjecture; DeepMind solved nine more problems. Yann LeCun's world models bet gets first test: JEPA identifiability proven but brittleness exposed. MIT builds self-evolving AI scientist. DeepMind's Genie 3 + SIMA 2 trains robots in imagined worlds. Echo-Infinity achieves 24-hour consistent infinite video generation. New papers: Qwen-VLA, ArcANE, Meta-Cognitive Memory Policy Optimization, Complexity-Balanced Diffusion Splitting, LoomVideo, SynthTraces, Meta-Agent Challenge, STRIDE, EvoDS, World-Language-Action Model, Gemma 4 QAT, Code2LoRA, Video2LoRA, SePO, LLMCodec, OpenSkill (open-world self-evolution), Socratic-SWE (self-evolving coding agents), SubtleMemory (fine-grained relational memory benchmark), Stream3D-VLM (online 3D understanding), UnEmbedding Matrix feature lens. Sakana AI establishes RSI Lab. NVIDIA Cosmos 3 open physical AI model. New benchmarks: SABER reveals >54% harmful safety violation rate in LLM coding agents; ForeSci evaluates forward-looking research judgment, finding evidence-decoupling failure mode. DeepSeek V4 Pro beats GPT-5.5 Pro on vulnerability scanning at 200x lower cost (anecdotal). LeanMarathon multi-agent system formalizes 258 lemmas in Lean with zero remaining 'sorry', advancing AI-assisted theorem proving. WorldBench benchmark reveals strongest MLLM only 64%, highlighting visual reasoning gaps. Compress-Distill paper achieves 2-7.6x training speedup by compressing reasoning traces before distillation. China's JAKA Pi humanoid robot with dual-brain architecture, Vietnam's Dino and VRH3, NVIDIA's open reference platform with 2000 TFLOPS. Two Leaps to 1000 tokens/s on a 1T-parameter model. SWE-Explore benchmark isolates repository exploration for coding agents. SpatialWorld benchmark tests interactive spatial reasoning; GPT-5 only 17.4% TSR. Latent Spatial Memory for video world models achieves 10.57x faster generation and 55x memory reduction. On the Geometry of On-Policy Distillation reveals subspace locking. ADASPEC improves multilingual speculative decoding with up to 2.3x speedup. Trajectory-Refined Distillation fixes prefix failure in on-policy distillation. End-to-End Context Compression at Scale tackles KV cache bottleneck. Evaluation Cards propose standardized AI evaluation reporting schema. OpenEnv moves to committee governance with Meta-PyTorch, NVIDIA, Hugging Face. KiloBench benchmark identifies most efficient AI models for coding. New paper proposes mathematical theory of deep representation learning unifying modern architectures. ANT architecture in Nature Communications achieves breakthrough energy efficiency with continuous learning. EMMA paper (CVPR 2026) extracts physical parameters from multimodal data using physics-informed learning. Pruning and Distilling MoE into Dense paper achieves +6.3 pp improvement over dense-to-dense pruning at 1.6x faster training. OpenAI's GPT-Rosalind specialized biology AI model for drug discovery, pandemic preparedness, and biodefense. DiffusionGemma achieves 4x faster text generation, challenging autoregressive dominance. FlowTracer (ICML 2026) uses attention-induced DAGs for token-level credit assignment in RL for LLMs. MemDreamer decouples perception and reasoning for long video understanding via hierarchical graph memory, SOTA with only 2% context ingestion. Workflow-GYM benchmark reveals top models only 30% success on long-horizon professional GUI tasks. Data2Story multi-agent framework for end-to-end data journalism with evidence grounding and multimodal output. SGDR (State-Grounded Dynamic Retrieval) improves web agents by 10.6% on WebArena via adaptive skill retrieval. Role-Agent bootstraps LLM agents via dual-role evolution, achieving 4% average gain. Waymo's Reference Driver model using active inference published in Nature Communications, open-sourced. Emergence World platform for evaluating long-horizon multi-agent autonomy. MIT Media Lab study finds AI dependency paradox: LLMs improve accuracy 21% but after a month unassisted performance drops 15 points, 1/5 become passive 'Dependency Developers'. Arbor framework for autonomous research using hypothesis-tree refinement, addressing long-horizon scientific tasks. Los Alamos method (PAS) detects VLM hallucinations by monitoring attention patterns, plug-and-play, presented at CVPR 2026. Lyapunov stability theory applied to detect LLM agent spiraling. Physics researchers prefer AI sidekicks over autonomous agents, with physics-intern system boosting Gemini 3.1 Pro from 17.7% to 31.4% on CritPT. New: Self-Harness framework for LLM agents to improve their own operating systems, extending self-evolution trend. New: General frontier LLMs outperform specialized clinical AI tools like OpenEvidence, challenging domain-specific fine-tuning assumptions. New: Quantifying Effect of AI Invention on Productivity shows 4-8% TFP improvement for expanding AI patent portfolios, reinforcing gradual productivity gains. New: DeepMind publishes roadmap to superintelligence, outlining key milestones. New: NBER paper formalizes 'explosion condition' for AI-driven growth, calibrating 13-17% automation could trigger hyperbolic growth within six years. New: Los Alamos National Lab develops PAS tool for detecting VLM hallucinations via attention patterns, presented at CVPR 2026. New: ActWorld paper introduces action-aware memory for interactive world models, bridging exploration and interaction. New: GLM-5.2 open-weight frontier model from Zhipu AI with strong coding and agentic gains, 1M context. New: OpenClaw-Skill paper uses tree-based skill induction with collective signal for LLM agents. New: Yann LeCun highlights Temporal Difference in Vision (TDV), a new method reducing strong assumptions in representation learning. New: Octopus architecture for AI agents offers modular, decentralized design pattern. New: Agentic coding harness effect: varying only the harness or hand-in method swings coding agent success from 19% to 73%, showing architecture matters as much as model choice. New: LLM persuasion study finds models out-persuade expert humans, with 3x effectiveness in charity donations, raising safety implications. New: Robot training data collection is dirty, unglamorous work; labs already paying XDOF for real-world data, challenging simulation-only approaches. New: DeepPTMPred: multi-modal deep learning for PTM prediction, niche but solid biology application. New: Physics-inspired deep learning for transistor modeling, niche hardware design application. New: AI, Privacy, and the Hidden Architecture of Harm from LLM pre-training. New: Alibaba's Qwen releases three specialized robot models (World, Manip, Nav) from VLA foundation, with 8.6M video dataset, advancing embodied AI. New: OpenAI o3 Deep Research finds 18 new diagnoses from 376 unsolved medical cases, published in NEJM AI, demonstrating clinical AI impact. New: Practical insight on coding agent environment management: snapshots and forks crucial for efficient agentic workflows.