Machine learning news — deep learning, reinforcement learning, neural architecture search, diffusion models, and new ML frameworks and libraries.
An AI agent rediscovered a degree-four Blaschke-curve invariant from numerical data, fitting a homogeneous cubic whose frozen coefficients predicted 480 lines from 80 unseen parameter values with a recorded RMS scale-fre…
Researchers proposed MoFlow, a method that generates agentic workflows optimized across multiple objectives at once by formulating workflow generation as a multi-objective Markov decision process and solving it with Conv…
AREX-2, an LLM agent built on Qwen3.8-27B and trained on synthesized long-horizon improvement trajectories from machine learning and algorithmic programming tasks, scored 81.8 on MLE-bench Lite and 70.7 on Frontier-CS, a…
Researchers introduced EVOKE, a post-training method that elicits pretrained world knowledge in LLM agents by holding environment state and interaction history fixed while ranking the same candidate actions under alterna…
DEdit, a diffusion-based speculative decoding drafter from an arXiv paper (arXiv:2609.38510v1), achieves macro-average speedups of 5.72x on Qwen3-4B and 5.97x on Qwen3-8B over autoregressive generation under greedy decod…
LoopVL, a recurrent vision-language model that combines Module-Loop and Model-Loop computation to iteratively update a unified vision-language state through shared modules, outperforms similarly sized and larger non-recu…
A new arXiv paper (2609.38817v1) introduces RADAR (Reasoning-state Analysis via Dynamic Attention Responses), a method that identifies a large reasoning model's current reasoning state in real time and detects abnormal a…
The CARAT benchmark, detailed in arXiv paper 2609.38340v1, finds that materials LLMs often recite structural relations rather than reason from them: a frozen model quoted a link yet answered identically when the link was…
GoldiMask, a fine-tuning method for discrete diffusion language models introduced in arXiv paper 2609.38385v1, selects which response tokens to reveal as context by approximately maximizing a submodular objective and the…
Researchers introduced Decode-Latency Feedback Prefill (DLFP), a model-free controller implemented in vLLM that resizes prefill chunks based on observed decode intervals, cutting P99 inter-token latency by a mean 27.7% (…
A new arXiv paper (2609.38448v1) introduces CHOIR (Collective Hierarchically-Ordered Inquiry Responses), a framework that adapts free-list elicitation from cognitive anthropology to LLM ensembles to distinguish genuine m…
The Agentic Idea Manager (AIM), a fully autonomous framework for idea-driven automated research from a team publishing on arXiv as 2609.38445v1, surpassed the strongest baseline by 1.6 percentage points on System Optimiz…
A new arXiv paper (2609.38181v1) introduces Survprompt, a framework that converts structured patient covariates into free-text clinical vignettes to prompt pre-trained LLMs for zero-shot survival prediction, benchmarked …
A study of 17 instruction-tuned models spanning 8 architecture families and 1.5B to 72B parameters found that safety system prompts barely alter internal transformer representations, with changes statistically indistingu…
A study posted to arXiv (2609.38261v1) introduced NinaXander, a method that connects layers of frozen language models from different architecture families through a single trained shared-latent adapter, composing the rec…
A new arXiv paper (2609.38274v1) proposes a heterogeneity-driven team selection framework for multi-LLM systems that profiles each model's capability plus two complementary signals: error-pattern decorrelation to reduce …
A new arXiv paper (2609.38406v1) introduces a ~9.2K-instance multi-domain benchmark for narrative text infilling and reports that model scale does not reliably predict infilling quality, with Gemma-2-2B scoring highest q…
A new arXiv paper, arXiv:2609.38374v1, proposes Doc2LoRA, a method that represents each scientific paper as a LoRA adapter generated by a Doc-to-LoRA hypernetwork, so that every point in the paper space — including mixtu…
A training-free framework called Inductive Visual Logic (IVL) achieves the highest aggregate accuracy across multiple distant out-of-distribution (OOD) benchmarks under two vision-language model backbones, according to a…
Researchers introduced MOBA-VL, a 9B-parameter vision-language model trained with event-localized multi-turn reinforcement learning on game telemetry to deliver real-time commentary for Multiplayer Online Battle Arena es…