Machine learning news — deep learning, reinforcement learning, neural architecture search, diffusion models, and new ML frameworks and libraries.
Researchers proposed SAGA (Self-evolving Agents through Experience-Grounded Abstraction), a framework that transforms LLM agent interaction trajectories into episodic descriptions, reusable procedures, and principles wit…
A new arXiv paper (2610.07023v1) introduces SSRFT (Supervised Safe-Role Fine-Tuning), a framework that reformulates LLM safety alignment as internalization of a predefined safe role rather than explicit refusal patterns.…
Researchers posted arXiv paper 2610.07018v1, "When to Rethink: Learning Multi-Perspective Self-Verification for Vision-Language Models," introducing MOTIVE, a Multi-View Self-Verification with Reliability-Guided Selectiv…
Two closed-form projection operators applied to a diffusion model's denoised clean-coordinate estimate recover perfect physical validity in AlphaFold 3-style biomolecular cofolding models while preserving structural-accu…
CuratorMAS, a multi-agent collaboration framework described in arXiv paper 2610.07075v1, automates dataset curation by decomposing the process into five programmable execution stages in a parallelizable workflow. The fra…
A new arXiv paper (2610.07093v1) benchmarks lightweight open-source language models for classifying smart data models (SDMs) on resource-constrained IoT edge devices, comparing general-purpose, reasoning-specialized, and…
A study of four open-weight large language models found that refusal behavior concentrates in sparse component mechanisms comprising 28–48% of upstream attention and MLP components, which retain 88–101% of full steering …
WavePrune, a method that restricts each RoPE channel to its first rotation period, raises the HELMET long-context score on four of five tested models without extra tuning, including 35.7 to 40.0 on Qwen3-8B, according to…
A new arXiv paper (2610.07070v1) introduces Turnslide, a fully automated multi-turn data synthesis framework that models each API as a finite-state machine to generate state-valid tool-call sequences, then translates the…
A new arXiv paper (arXiv:2610.06902v1) introduces NavTree, a leaves-only hierarchical retriever that builds a deterministic balanced segment tree over document chunks with zero language-model calls at indexing and uses t…
Condition-anchored generative distillation (CAGD) reduced four-task final held-out loss in a 219M masked diffusion language model from 2.927 to 1.114 in one task order and from 2.168 to 0.891 in exact reverse, according …
A jointly distilled and quantized student model for French-to-English biomedical translation achieved a 69% reduction in size, a 98.21% increase in inference speed, and a 98.46% reduction in CO2 emissions versus the orig…
JudgeMoE, a lightweight aggregator that assigns example-specific weights to cached LLM judge score distributions before fusing them, improved mean Spearman correlation over uniform log pooling by +0.079 on the original 1…
Researchers released CroissantMiner, the first benchmark for end-to-end evaluation of Croissant metadata extraction from ML dataset documentation, comprising 602 papers — 102 with human-validated gold annotations and 500…
Researchers proposed CARAT, a multimodal test-time adaptation method that decouples model reliance from runtime corruption detection, achieving the highest overall macro-F1 and a best mean rank of 2.42 across four wearab…
A new arXiv paper (2610.07348v1) introduces Stepped MoE, a framework combining elastic structures with sparsely gated architectures so a single model can operate at 1, 2, 3, or 4 billion parameters at inference time. The…
Researchers proposed CausalBind, a causal-modeling method for protein-molecule virtual screening that formalizes binding as sparse cross-modality interactions via a V-structure causal model under Heckman-style selection,…
Stock-JEPA, a joint-embedding predictive framework that learns predictable incremental revisions relative to a point-in-time financial prior, outperformed 13 strong baselines across large-scale China and U.S. equity univ…
Researchers propose DDT-RFE, a Decoupled Diffusion Transformer variant that removes residual connections around the Self-Attention and MLP operations in each encoder block while fusing the input patch embedding with inte…
A new method called DecisionSteer trains a System-1 decision operator via behavior cloning using 330K parameters, matching a 1.33M-parameter reinforcement-learning-trained operator while collapsing 3,685-token deliberati…