Daily updates, analyzing cutting-edge AI research papers
ScienceBuddy is an open-source AI research agent that uses a recursive dual-loop architecture to continuously improve its harness and model weights from real user interactions, achieving 42.2% to 73.3% accuracy gains on scientific benchmarks across genomics, molecular biology, and pharmacology domains.
Community tests of OpenAI's GPT-6.1 Sol reveal it writes 159 lines for an SVG pelican versus Sonnet 5.5's 517, matches Opus 5.5 in Blender carrier building, but fails at Three.js Eiffel Tower, while costing $5.00 and taking 99 minutes for three racing tracks compared to GPT-6 Sol's $3.90 and 78 minutes.
Anthropic's new research reveals two agent paradigms: Nine Loops demonstrates a single agent completing a 96 CPU-week physics computation with only periodic check-ins, while ART orchestrates 949 agent sessions to autonomously discover a novel enzyme system, exposing critical insights on long-task reliability, multi-agent organization, tool-use pitfalls, and interpretability for attribution.
Peking University and Google's Harness-Zero introduces agent-as-harness, where a harnessing agent reviews and corrects a student agent's actions using an evolved reference harness, then distills the corrected trajectories into the student model via SFT, achieving 90% relative improvement on a 9B model without external scaffolding at deployment.
This article surveys the 2026 reinforcement learning landscape for large language models, detailing foundational algorithms (PPO, DPO, GRPO), advanced GRPO variants (DAPO, GSPO, GMPO, GFPO), emerging Agentic RL methods (ARPO, Tree-GRPO), and a practical scenario-based selection table.
The article introduces Jev, a fast, cheap AI model from TypeSafe that handles reflexive decisions while LLMs handle reasoning, showcasing five demos: context compression, ad analysis, real-time Mario gameplay, probability-based animations, and autonomous rocket landing — all at fractions of LLM cost and latency.
Laya, a 421M-parameter open-source non-autoregressive decision engine, outperforms the commercial Jev model with 7.8x faster inference, 3.9% higher accuracy, and better calibration, using a ModernBERT encoder with masked token scoring and a multilingual router, while honestly acknowledging limitations in zero-shot and high-cardinality tasks.
This article dissects the looped transformer architecture behind GPT-6 Astra, explaining recurrent depth, adaptive recursion mechanisms from Universal Transformer to Mixture-of-Recursions, and research showing recurrence improves multi-step reasoning without increasing knowledge capacity, while reducing training compute by 6.8–18% at scale.
SenseNova-U1.5 introduces an 8B MoT unified multimodal architecture that natively integrates understanding, generation, and editing via spatial joint reconstruction and a staged training strategy, achieving state-of-the-art results on Qwen-Image-Bench, GenEval, CVTG-2K, ImgEdit, and OpenING while maintaining strong language understanding.
This article analyzes five fundamental shifts in LLM post-training over the past six months: moving from monolithic RL to expert distillation (MOPD), online distillation as a 10x cheaper RL alternative, refined RLVR techniques addressing entropy collapse and exploration, SFT-RL distribution alignment, and data quality as an irrecoverable hard constraint.
