Act on this
LGWM world model flags hijacked GUI agent transitions
Test the released LGWM model and dataset on Hugging Face to detect hijacked screen transitions in GUI agents.
- Signal
- 71
- Facts
- 89
- Hype
- 7
Hugging Face Papers
Source, web_page
38 cards
Act on this
PrismaCoder releases open 9B and 35B multimodal code models
Download PrismaCoder-9B or 35B and its 2M-sample dataset to test visual coding tasks.
- Signal
- 61
- Facts
- 89
- Hype
- 32
Brief
Agent4Sci tests coding agents on 20 scientific ML benchmarks
Claude Opus gains 10.4% mean relative improvement, rising to 20.1% with targeted plug-in skills. The code and data are public.
- Signal
- 55
- Facts
- 95
- Hype
- 13
Brief
PAPAV framework finds embodied agent benchmarks skip anticipation and verification
An arXiv paper proposes five capabilities for multimodal embodied agents. Across 64 benchmarks only 5 assess Anticipate and 3 assess Verify.
- Signal
- 42
- Facts
- 80
- Hype
- 1
Brief
TraceForge distills reasoning traces from a teacher model's answers
A paper that lifts HumanEval+ by 3.7 and MBPP+ by 7.1 points. No code or model release is linked.
- Signal
- 55
- Facts
- 86
- Hype
- 10
Act on this
WildAWM synthesizes 341 computer-use environments with public code
Use the open-source WildAWM pipeline and its Qwen checkpoints on Hugging Face to build computer-use training environments.
- Signal
- 76
- Facts
- 95
- Hype
- 13
Brief
DeLoop uses repeated actions as self-supervision for long-horizon agents
The method cuts repeated tool-call loops, lowering Qwen3-8B loop share from 57.3% to 6.1% on CodeGym under compression. The paper was published October 14.
- Signal
- 69
- Facts
- 88
- Hype
- 3
Brief
Agentic World Action Models paper jointly predicts code, video and robot actions
The arXiv paper AW-0 reports gains over VLA and WAM baselines on RoboTwin 2.0. No models or code are linked yet.
- Signal
- 45
- Facts
- 67
- Hype
- 6
Brief
Socratic Skill Learning interleaves teaching and practice for agent skills
The method reaches 0.801 utility on SkillTransferBench versus 0.723 with no skill. A variant uses 61.7% fewer educator tokens, and the dataset is on Hugging Face.
- Signal
- 52
- Facts
- 86
- Hype
- 3
Brief
AgenticRoboRL lifts robot planner success from 18% to 71%
A 9B planner post-trained with SFT and multi-turn RL reaches 71.0% on LIBERO-PRO, near a Codex GPT-5.5 setup with memory. One Qwen3.5-9B checkpoint is linked and the rest is promised.
- Signal
- 69
- Facts
- 87
- Hype
- 10
Brief
Harness Scaling paper trains agents across many execution harnesses
The paper reports better generalization to unseen white-box and external black-box harnesses. It is a research abstract with no linked code release.
- Signal
- 44
- Facts
- 67
- Hype
- 3
Brief
EAU co-evolves users, environments and agents for training
An arXiv paper where a meta-agent co-evolves a user simulator, environment and RL setup. It reports 14.7 to 15.0 percent overall gains and up to 36.7 percent on ColBench.
- Signal
- 53
- Facts
- 83
- Hype
- 4
Brief
TEAL evolves reusable analytics tools for coding agents over data lakes
An arXiv paper reports 9.3 and 21.7 point gains for Sonnet-5 at a 15-step budget. No code release is linked yet.
- Signal
- 52
- Facts
- 85
- Hype
- 2
Brief
ASAP evaluates agentic behavior in pretrained base models
An arXiv paper proposing a lightweight evaluation on end-to-end software and research tasks. It is background research with no engineering action yet.
- Signal
- 42
- Facts
- 68
- Hype
- 3
Brief
FlowState uses execution state as memory for long-horizon agents
An arXiv preprint proposing structured, revisitable execution state as agent memory. It reports gains over a full-context baseline on DeepSeek-V4-Flash, with no code linked.
- Signal
- 58
- Facts
- 83
- Hype
- 5
Brief
Épi: a persistent research agent reports 120 record-setting results
An arXiv preprint whose coding variant beats Codex by 5.4% on 30 FrontierOR tasks at 64% of the cost. No code release is confirmed.
- Signal
- 42
- Facts
- 90
- Hype
- 18
Brief
A decision-model controller raised agent cost 19% without better results
A pre-specified study on Terminal-Bench 2.1 found pass rates stayed at 73.0% in both arms. The controller was poorly calibrated and never chose to abandon a run.
- Signal
- 59
- Facts
- 89
- Hype
- 1
Brief
DecisionPivot shows aggregate accuracy hides decision-model failures
A diagnostic benchmark of 48 tasks and 40,033 items. It is a research paper only, with no code release confirmed.
- Signal
- 54
- Facts
- 89
- Hype
- 1
Brief
Anti-Persistence: discard agent guidance after each run
An arXiv paper proposing to drop task-specific guidance after execution. It reports higher held-out success and the lowest LLM-call cost among compared memory designs.
- Signal
- 57
- Facts
- 39
- Hype
- 11
Brief
RiseHarness improves terminal-agent harnesses with model weights fixed
An arXiv paper that iteratively refines the harness using successful and failed runs as references.
- Signal
- 53
- Facts
- 67
- Hype
- 5
Brief
S-Agent 2.0 adds spatial reasoning to a general coding agent
A command-line 3D perception toolkit with persistent scene memory. It reports 66.5 on MMSI-Bench and 82.3 on ReVSI with open Qwen3.8-27B.
- Signal
- 72
- Facts
- 91
- Hype
- 6
Brief
Sage-A1.5: a 35B agentic model for scientific exploration
A mixture-of-experts model with subagent delegation and verification. It was a Hugging Face paper of the day on October 20, with no linked models yet.
- Signal
- 51
- Facts
- 79
- Hype
- 17
Brief
HybridResearcher: a 4B deep research agent beats a 397B model
Trained on Qwen3.5-4B with synthetic cross-source trajectories. It gains 30.0 pass@8 points on HybridDeepResearch and beats the 397B model by 8.3.
- Signal
- 60
- Facts
- 83
- Hype
- 11
Brief
MMEnterprise-Bench tests enterprise agents on evolving multimodal records
An arXiv benchmark of 1,613 questions across 10 MLLMs, with code and data released. It covers evidence reasoning, conflicts and unanswerable cases.
- Signal
- 54
- Facts
- 92
- Hype
- 1
Brief
FocusGUI picks omni-modal evidence per step for GUI agents
An arXiv adapter reporting average Type Match gains of +33.2 across four backbones, with no fine-tuning of the backbone.
- Signal
- 65
- Facts
- 88
- Hype
- 8
Brief
Sauron verifies agent rubric claims against environment state
Reports 0.96 F1 versus 0.92 for Gandalf at 23% lower cost. It detects 97% of injected failures with 2% false flags.
- Signal
- 62
- Facts
- 87
- Hype
- 8
Brief
HarnessPro self-improves agent harnesses with Pareto cost trade-offs
An arXiv paper with lite, medium and heavy variants. Heavy mode reportedly gains 2.0 points and cuts per-task cost 26.1% versus RHO, with code on GitHub.
- Signal
- 72
- Facts
- 94
- Hype
- 10
Brief
Interlock trains long-horizon agents on typed completion graphs
Fine-tunes Qwen3.6 on accepted trajectories, improving benchmark pass rates by about 3 to 6 points.
- Signal
- 55
- Facts
- 82
- Hype
- 3
Brief
BizBench shows agents break constraints even when goals are met
The 50-task benchmark uses 789 docs and 601 tools. Frontier agents often break intermediate constraints on the way to a goal.
- Signal
- 62
- Facts
- 89
- Hype
- 4
Brief
LiveMACE tests five frontier agents over 30 days of live markets
The benchmark finds a gap between realized returns and capability diagnostics drawn from decision traces.
- Signal
- 46
- Facts
- 81
- Hype
- 4