Machine learning news — deep learning, reinforcement learning, neural architecture search, diffusion models, and new ML frameworks and libraries.
A new arXiv paper (2609.36172v1) reports that VGG16 achieved 89.55% validation accuracy in recognizing topological relationships between objects in images, outperforming traditional models including Naive Bayes, KNN, Ran…
PreviewDiff, a training-free test-time search method described in arXiv paper 2609.36199v1, turns diffusion sampling from scalar search into a multimodal critic-guided search over intermediate latents, decoding partial p…
A new arXiv paper (2609.37175v1) presents a general-purpose vision-language solver that augments textual instance descriptions with input-derived visual representations for end-to-end combinatorial optimization, trained …
Researchers proposed TagPR, a framework that adds semantic tags to a large language model's reasoning process for step-by-step guidance on personalization tasks, according to an arXiv paper (2509.23140v2). TagPR generate…
A new arXiv paper (2609.35868v1) proposes Desired-Update-Aligned Synthetic Data (DASA), a method that fine-tunes large language models using continuous synthetic input embeddings guided by activation-gradient feedback fr…
A new arXiv paper (2609.35897v1) presents the first causal mechanistic audit of a self-discovered reinforcement learning rule, Disco103, which previously surpassed PPO to reach state-of-the-art benchmark performance. By …
A controlled evaluation of eight generated LLM harnesses plus a baseline and nine byte-identical baseline copies on 386 MATH-500 tasks found that identical programs alone yield 2.16 percentage points of repeat-averaged o…
A controlled study of GPT-2 Small found that representational simplicity and causal circuit size dissociate in a threshold-dependent way, with adversarially trained models needing substantially fewer edges only at high f…
Researchers introduced COFFEE, a plug-and-play framework that guides discrete diffusion models with sequence-level objectives without retraining the diffusion model, according to the arXiv paper 2609.35924v1. COFFEE sepa…
Researchers released GeoOutageBench, a benchmark for evaluating LLM-based geospatiotemporal knowledge-graph question answering (KGQA) on multimodal power outage and resilience analysis, with code, data, and results avail…
A sweep of eight agent orchestration architectures across five instruction-tuned 7-9B models and six benchmarks found that scaling a team from three to thirty calls raises accuracy by up to 17 points on the arithmetic wo…
Sage, an agentic formalization framework described in arXiv paper 2609.35790v1, suppresses answer leakage to 2.7% while reaching 73.3% pass@4 joint compilation and semantic fidelity on Omni-MATH without proofs, versus 42…
A new arXiv paper (2609.35797v1) finds that collapsing large language model judge rewards to pass/fail {0, 1} hides proportional changes in how well a response meets each criterion, and recommends keeping at least three …
HeadGuard, a composable head-protection method for low-bit vision-language model KV-cache quantization, recovers a substantial fraction of lost accuracy by keeping roughly 1/8 of physical KV heads in bfloat16, according …
A new arXiv paper (arXiv:2609.35856v1) proposes using generative large language models with weak supervision to classify Economic Policy Uncertainty (EPU) in news articles, replacing keyword-based methods that produce hi…
A calibrated hybrid ensemble combining five deep learning models and three classical machine learning models via a logistic regression stacking meta-learner achieved 74.2% seizure-level sensitivity at 1.24 false alarms p…
A new arXiv paper (2609.35852v1) analyzes Transformer weights at a mesoscopic level using row and column scale fields — median-centred log-RMS profiles over a matrix's channels — which, with a global scale and a full bal…
Mara Chain, a refinement procedure that retains and iteratively refines rejected candidate configurations instead of discarding them, outperformed GEPA, ACE, and SkillOpt-Lite by up to 20.5% in relative performance on Ap…
A new arXiv paper (2609.35884v1) introduces a minimal representation that encodes a real D-component divergence-free vector field on a D-dimensional domain as a real (D-1)-component field on the same domain, using a Hous…
A new arXiv paper (2609.35910v1) introduces DEEP (Decision Evaluation for Escalated Perception), a benchmark that scores pre-escalation compute allocators against a budget-constrained decision oracle, finding that 34–54%…