Research Command Center

EnvHarness and EnvRigger dynamically reshape static environments via programmable plugins to target agent weaknesses and improve reinforcement learning co-evolution.

Read the abstract and check whether the method changes tool, benchmark, or model coverage.

StateM is a runtime system that improves long-horizon agent execution through durable states, recoverable runbooks, and enforceable procedural controls without altering model weights.

Read the abstract and check whether the method changes tool, benchmark, or model coverage.

A new benchmark for AI-generated video detection reveals that current detectors fail to generalize across realistic crisis-related videos and become less reliable as content spreads socially.

Read the abstract and check whether the method changes tool, benchmark, or model coverage.

Spark-to-Paper is a lightweight, composable workflow inside coding assistants that generates research papers by separating planning from reporting, enforcing evidence-based claim revision, and using integrity checks to reduce fabrication.

Read the abstract and check whether the method changes tool, benchmark, or model coverage.

Combodied Agents integrate digital and embodied tools into a closed-loop framework that models individual human-state trajectories over time to provide proportionate, consent-aware support.

Read the abstract and check whether the method changes tool, benchmark, or model coverage.

Latent-to-4D enables reusable direct 4D generation from video diffusion latents via alignment with a pretrained decoder and spatiotemporal attention, transferring across generators without retraining.

Read the abstract and check whether the method changes tool, benchmark, or model coverage.

A 150M-parameter reasoning model using recurrent latent reasoning and in-context learning achieves a new cost-accuracy frontier on ARC-AGI-1.

Read the abstract and check whether the method changes tool, benchmark, or model coverage.

Macaron-V1 is an open agent-model family that uses a Mixture-of-LoRA architecture and recursive self-improvement to enable continual learning and collaboration across specialized tasks.

Read the abstract and check whether the method changes tool, benchmark, or model coverage.

Unsupervised on-policy self-distillation improves large language models by using internal consistency and majority-vote pseudo-solutions to correct confident errors without external supervision.

Read the abstract and check whether the method changes tool, benchmark, or model coverage.

This paper introduces a method to automatically generate large volumes of training data for AI agents that perform long, multi-step tasks in terminal environments (like running commands in a shell). The key problem is that such data is normally hand-crafted and costs hundreds to thousands of dollars per task because the instruction, environment, and solution must all be consistent. The new approach, called Recursive Synthetic Terminal Tasks (RST), starts with a few verified tasks and repeatedly extends their solutions, then automatically updates the instructions and checks to match, validating each new task in a sandbox. Over 15 rounds, it produced 37,484 tasks at about $0.05 each, with difficulty rising sharply—median solution length grew from 67 to 374 lines, and the best model's success rate dropped from 90% to 2.5%. Training models on these synthetic tasks improved their performance on standard terminal-agent benchmarks by up to 41% relative, and the process shows no signs of plateauing, suggesting it could scale further.

Read the abstract and check whether the method changes tool, benchmark, or model coverage.

OpenART introduces a scalable red-teaming arena with evolving stateful environments to evaluate long-horizon AI agent safety, using the EMHA attack policy to expose increasing failure rates as task complexity grows.

Read the abstract and check whether the method changes tool, benchmark, or model coverage.

This paper introduces Qwen-UI-Agent, a foundation model designed to act as a general-purpose executor across mobile, desktop, web, and search interfaces. Unlike typical GUI agents that operate in isolated sandboxes, it runs on real devices and unifies GUI clicks with command-line actions, generating multiple steps per turn to handle long, complex workflows. A key innovation is an automated data flywheel where the agent itself builds tasks, diagnoses its own failures, and plans improvements, plus online reinforcement learning on trajectories over 100 steps using 10,000+ parallel environments. The headline results show state-of-the-art mobile performance (82.1% on MobileWorld, 92.2% on MobileWorld-Real) and competitive computer/browser scores (79.5% on OSWorld-Verified, 73.6% on WebArena) against frontier models like GPT-5.6 and Gemini 3.1 Pro. The work matters because it pushes agents from scripted demos toward reliable, self-improving operation on actual devices, a prerequisite for practical digital assistants.

Read the abstract and check whether the method changes tool, benchmark, or model coverage.

AskChem changes chemistry literature search from returning ranked papers to returning atomic, provenance-carrying claims—each tied to a source DOI and a verbatim quote or evidence locator. It indexes 2.4M claims from 147K papers and offers multiple access paths: a faceted taxonomy for browsing, an evidence graph linking claims, and a living taxonomy organized by scientific principles, plus REST, SDK, and MCP interfaces for AI agents. In their benchmark, grounding a GPT-5.5 reader in AskChem achieved 100% resolvable DOIs versus 88.3% without retrieval, and the highest citation density among five systems tested. This matters because it directly addresses the manual assembly and verification burden in cross-paper synthesis, for both humans and AI agents.

Read the abstract and check whether the method changes tool, benchmark, or model coverage.

This paper introduces OpenMLE, an open-source framework for studying recursive self-improvement (RSI) in machine learning engineering, where AI systems help build better AI. The framework includes a verifiable task environment, reinforcement learning for operator learning, and long-horizon evolutionary search. Using this stack, the authors post-train a 35B-parameter model called Frontis-MA1 to act as a meta-evolution agent, composing four atomic program-editing operations—Draft, Improve, Debug, and Crossover—into an iterative loop. On the MLE-Bench Lite benchmark, the model improves its Medal Average from 39.39% to 60.61% over its base model, and reaches 71.21% with an enhanced search variant, outperforming GPT-5.5 + Codex and approaching much larger models like GPT-5.6 Sol and Kimi K3. The results demonstrate that both the trained model and the search framework independently contribute to performance gains, and the authors release the model weights and code to enable reproducible RSI research.

Read the abstract and check whether the method changes tool, benchmark, or model coverage.

Kimi K3 is a 2.8-trillion-parameter open-source AI model that uses a Mixture-of-Experts design to activate only 104 billion parameters per token, making it efficient despite its size. It introduces new attention mechanisms and a stable expert-routing method to improve performance across long sequences and deep networks, achieving roughly 2.5x better scaling efficiency than its predecessor. The model excels at long-horizon coding, agentic tasks, reasoning, and vision, outperforming all other open models and most proprietary ones, though it still trails the top closed models like Claude Fable 5 and GPT-5.6 Sol. Its release as open weights aims to accelerate research and deployment of frontier-level AI.

Read the abstract and check whether the method changes tool, benchmark, or model coverage.

ABot-World-0 is a video world model that runs real-time, interactive simulations on a single desktop GPU. It learns to generate coherent video sequences in response to keyboard actions by training on a diverse mix of AAA games, simulations, and internet videos. The key innovation is a training method called LongForcing that aligns the model's long self-generated sequences with a more accurate teacher model, preventing the drift that typically plagues long video generation. On a single RTX 5090, it streams 720P video at up to 16 FPS with only 1.2 seconds of latency, enabling interactive scene roaming and character control.

Read the abstract and check whether the method changes tool, benchmark, or model coverage.

RynnBrain 1.1 is a family of embodied AI models (2B, 9B, and a 122B-A10B mixture-of-experts variant) trained to handle perception, spatial reasoning, localization, and planning for robots. The key upgrades over version 1.0 are contact-point prediction (where a robot should touch an object) and native 3D grounding for the smaller models, which makes their outputs more directly usable for manipulation. The authors also built a separate VLA (vision-language-action) model with a shared action space across different robot bodies, using masking to handle their differing morphologies, and tested it on three real robots. On benchmarks like VSI-Bench and RefSpatial-Bench, the largest model beats all proprietary and open-source competitors, and real-robot tests show that policies initialized with RynnBrain outperform those based on Qwen or other generalist VLAs. Notably, training on multiple tasks and robot types simultaneously improved both process scores and success rates compared to training each task separately.

Read the abstract and check whether the method changes tool, benchmark, or model coverage.

LongStraw is a system that enables reinforcement learning (RL) post-training on contexts exceeding 2 million tokens using a fixed GPU budget, addressing the gap where inference handles million-token contexts but RL training typically tops out at 256K tokens. It works by evaluating the shared prompt without autograd, keeping only essential model state for later tokens, and replaying short response branches one at a time—trading extra replay time for a smaller live training graph. Implemented on Qwen3.6-27B and GLM-5.2 models, LongStraw completed grouped scoring and backward passes at 2.1 million positions on eight H20 GPUs, with a stress test reaching 4.46 million positions. This matters for AI agents that accumulate long trajectories of observations and tool outputs, though the paper notes these experiments establish execution capacity rather than full training correctness.

Read the abstract and check whether the method changes tool, benchmark, or model coverage.

Modern AI agents rely on a 'harness'—the code that builds prompts, manages state, calls tools, and coordinates execution—which must be constantly updated as models and APIs change. The main bottleneck is finding exactly which lines of code implement a desired behavior, because harness code is large, tangled, and spread across many files. This paper introduces the Harness Handbook, an automatically generated document that maps each high-level behavior to its source code locations, and a method called Behavior-Guided Progressive Disclosure (BGPD) that helps both human developers and coding agents navigate from a behavior description down to the relevant code. In tests on two real open-source harnesses, the handbook approach improved the accuracy of locating code and planning edits while using fewer tokens, especially for behaviors scattered across modules or on rarely executed paths.

Read the abstract and check whether the method changes tool, benchmark, or model coverage.

Orca establishes a unified world latent space through next-state-prediction modeling using multimodal data and demonstrates superior performance in downstream tasks compared to specialized baselines.

Read the abstract and check whether the method changes tool, benchmark, or model coverage.