2026
10,780 articles · 136 days · 10 categories · avg 79 per day
AVA improves multimodal ICL accuracy by 1.7 points without any training
A preprint proposes AVA, a three-stage pipeline that cleans up the examples in a frozen VLM's prompt. It gains +1.7 points on average across five benchmarks, with…
Coral-KV Compresses Video Model KV-Cache by 37–48% and Lifts PSNR by Up to 4.95 dB
Researchers propose Coral-KV, a KV-cache compression method for autoregressive video generation. On MAGI-1 it gains 4.95 dB PSNR over QVG-4 while using 37% less…
Black Lake: Chinese Factories Skip Digitization and Go Straight to AI Agents
Black Lake CEO Zhou Yuxiang likens factories' shift to AI agents to China's leap from bank cards to mobile payments, and plans to export the approach to global…
A Physical Transformer Built Entirely From Transistors Learns In-Context Learning in Simulation
Theory Shows Masked Diffusion on Transformers Learns from Õ(LK/ε) Samples
SAZO speeds up LLM fine-tuning without backprop: up to 6× fewer queries to reach the same loss
Holomorphic Neural Networks With Singular Functions Beat All Methods on Laplace Benchmarks
Cox Regression Learns From Truncated Data Orders of Magnitude Faster Than EM
An arXiv paper proves that Cox regression can be trained efficiently on truncated and censored data with gradient methods. The speedup over EM is orders of…
CacheScope speeds up diffusion models by up to 3.57x and cuts LPIPS by 70%
CacheScope needs no fine-tuning. It finds the steps where caching is risky and computes them in full. On FLUX.1-dev, Wan2.1 and Open-Sora 1.2, it gives up to 3.57x…
CircuitTree assembles 39 HuBERT phoneme circuits into a syllable transcriber with 64.8% accuracy
ResonatorLM Taught to Read Memory by Token: Perplexity at 1024 Drops from 12.81 to 3.76
Kairos: a byte-level model for all modalities converges on 440 bytes
Open LLMs Explain Satire More Cheaply Than Frontier Models, but Handle Australian Satire Worse Than American
perturber: a service with 68 text perturbations tests LLM robustness
Researchers have released perturber, a service and Python library that apply one meaning-preserving perturbation at a time from a catalog of 68 variants across 11…
GeopoliticalQA: bge-m3 Beats Commercial Embeddings on 5 525 Report-Based Questions
A new benchmark tests RAG on 415 geopolitical reports and 5 525 questions: compact bge-m3 outperforms commercial APIs, while generation struggles with inference and…
Qwen Picks What to Remember: Replay From a Frozen Model Reduces Forgetting
Researchers tested whether a language model can protect its own skills during fine-tuning. Generated replay reduces forgetting, and task selection works at 4B.
RGR Protects LLMs from Skill Loss in Self-Distillation by Preserving Representation Geometry
Researchers Propose a 3-Criterion Formula for Creativity and Test It on Chess and Picasso
Neural Networks Are More Modular Than Their Neurons Suggest: Three Levels of Testing
New Method Aligns Encoders With 100–500 Pairs Instead of Millions
LLM Pipeline Turns Datasheets into Knowledge Graphs with 98.8% Accuracy
A new pipeline builds hierarchical graphs from datasheets: 98.8% faithful properties versus 76.1% for a single-pass LLM query. Tested on 20 documents.
EvoSyn: Multi-Agent Data Synthesis Lifts Task Success by 13.3 Points
Researchers propose EvoSyn, a framework that reshapes the synthetic-data generation space based on student feedback and beats the strongest baseline by 13.3 points.
MCTS Loses to Random Restarts by 2.88% in DAG Scheduling at Equal Budget
arXiv paper finds five flaws in XAI faithfulness metrics
DBS CEO Tan Su Shan: Market Is Repricing the Risks of Financing the AI Buildout
Bain Capital's Bridge Data Centres Seeks to Raise Its Loan to $6.6 Billion
STAPLE adds billions of parameters to OLMo-2 and decodes twice as fast as MoE
STAPLE embeds static sparse memory into a pretrained LLM. On OLMo-2 1B it gains +4.27 points on HumanEval, while training and generation slow by no more than 3%.
MisVisAgentBench: Misleading Charts Cut Web Agent Success by 34 Percentage Points
The new MisVisAgentBench tested 13 visual web agents on 280 tasks. Swapping an honest chart for a distorted one cuts the success rate by more than 34 percentage…
Conversational Model Benchmark: GPT Realtime and Gemini Respond at Almost Every Pause
Researchers tested whether voice models can stay silent: on 427 pauses, APIs from OpenAI, Google and xAI almost always respond. Only Moshi and PersonaPlex hold back.
GEPA Lifts a Deep-Research Agent's Reward from 0.702 to 0.781 on Unseen Topics
ELR teaches one model to loop: Koi-4B scores 18 and 29
DCOCO speeds up LLM inference by up to 6.6x and beats rivals by 6–23%
