Learning Lean Feedback for Efficient Proof-Candidate Evaluation on GPUs achieves 87.8% accuracy in predicting Lean acceptance. The two-stage evaluator processes 3.5 candidates per second on a single H100, 4.2× faster than a 64-worker baseline.
Key points
Trained neural evaluators on 4.84M labelled programs.
Combined evaluator processes 3.5 candidates per second on one H100, 4.2× the throughput of the 64-worker Lean baseline.
Sources
MLExpert: 35B model lifts MLE-Bench Lite rate to 72.7%
MLExpert is a 35B Mixture-of-Experts agentic model for autonomous machine learning engineering. It raises the average medal rate on MLE-Bench Lite from 37.9% to 72.7%, matching performance of larger frontier models such as Kimi-K3 and GLM-5.3.
Key points
Model size: 35B Mixture-of-Experts agentic model
Medal rate improvement: 37.9% → 72.7% on MLE-Bench Lite
Comparable performance to Kimi-K3 and GLM-5.3 models
Sources
- Original post
- MLExpert: Towards a Generalist Agentic Model for Autonomous Machine Learning Eng - Linked page
- github.com - Linked in the post
Immune changes could signal inflammatory bowel disease up to 10 years before diagnosis
Immune changes could signal inflammatory bowel disease up to 10 years before diagnosis presents a large-scale study that identified distinct antibody patterns a decade before IBD onset. Machine-learning models using these signatures achieved AUC 0.90 for Crohn’s disease and 0.84 for ulcerative colitis, and 0.85 for predicting complicated Crohn’s disease.
Key points
Analyzed 2,000 serum samples from 500 individuals, including 200 Crohn’s, 200 ulcerative colitis and 100 controls.
Identified 123,980 unique antibodies; antibody diversity was lower in future IBD cases versus controls.
Sources
threejs Rapier game fps increase 20× on https
threejs Rapier game ran at 2–3 fps on an iPhone and now reaches a locked 60 fps. Switching from http to https on iOS Safari re-enables JIT, making physics 20× faster.
Key points
FPS: 2–3 fps on http, 60 fps on https.
Performance gain: physics 20× faster on https.
Sources
Evaluation Awareness Beyond English: LLMs show higher refusal in Indic languages
Evaluation Awareness Beyond English: Do LLMs Act Differently When Tested in Indic Languages? finds refusal of harmful requests rises 14.4 pp on average in Indic languages. The study tests four instruction-tuned models in five Indic languages and English, using cues about monitoring and safety testing.
Key points
Models refuse harmful requests more often when told they are being tested; the increase appears in all four models and all six languages.
In Indic languages, refusal of harmful requests rises by 14.4 percentage points on average; refusal of benign Indic requests rises from 6.9% to 11.3%.
Sources
GS-PRM: Gaussian Visual Roadmap for Image-Goal Navigation announced with performance gains
GS-PRM: Gaussian Visual Roadmap for Image-Goal Navigation introduces a roadmap built directly from 3D Gaussian Splatting renderings without occupancy grids. The paper reports up to twice higher average success rate and up to 57.2% reduction in travel time compared to a state-of-the-art 3DGS-based system.
Key points
Introduces render-based collision-free (RC-free) configurations and paths using panoramic depth and edge-aligned orthographic rendering.
Demonstrates deterministic query completeness and probabilistic construction completeness under defined conditions.
Sources
Provably Efficient Risk-Sensitive Offline RL via Optimized Certainty Equivalents
Provably Efficient Risk-Sensitive Offline Reinforcement Learning via Optimized Certainty Equivalents introduces pessimistic value-iteration algorithms for dynamic-OCE and static-OCE formulations and proves sublinear suboptimality bounds. It also provides minimax lower bounds and extends the approach to linear function approximation with utility-dependent guarantees that scale with feature dimension d.
Key points
Pessimistic value-iteration algorithms achieve tight sublinear suboptimality bounds for both dynamic-OCE and static-OCE RL.
Linear function approximation version yields utility-dependent sublinear guarantees with explicit dependence on feature dimension d.
Sources
Qwen Go personal assistant launch this month
Qwen Go personal assistant launch is coming this month. It builds on Qwen's agent work to tackle real errands instead of just chat.
Key points
Launch: this month
Builds on Qwen's agent work to tackle real errands instead of just chat
