#Research
Curated coverage on «Research»: releases, research, benchmarks, open-source.
NEWThe DecoderGoogle taught its StreamVibe video agent to decide when to step in on its own: +10.7%
arXiv cs.AINew test catches hidden AI agent collusion by measuring how a partner reacts to a paraphrase
arXiv cs.AIIllinois compiled 12 151 open problems and tested whether AI can solve them
TG: @gonzo_ML (gonzo-обзоры ML статей)Layer Duplication During Training Lifts GSM8K from 12% to 19% on a 3B Model
arXiv cs.AILRMoE trims the MoE backward pass and speeds up training by 21–45%
arXiv cs.AIOntology as a checker for an AI agent: on UniProt, its error flags are only slightly more accurate than random
arXiv cs.AIIBM: Asking the Client Questions Before Generation Raised Smart Contract Accuracy From 54% to 73%
TG: @cgevent (Метаверсище и ИИще)T-Bank Fine-Tunes LLMs With Gradient-Free Evolution: 7.8x Faster on 128 GPUs
LRMoE trims the MoE backward pass and speeds up training by 21–45%
A new MoE training system drops the least useful token–expert pairs in the backward pass. On Qwen3-30B-A3B and GLM4.5-106B-A12B, it is 21.2–44.8% faster than the…
Ontology as a checker for an AI agent: on UniProt, its error flags are only slightly more accurate than random
Sano Centre researchers tested an ontology, a formal description of a knowledge graph, as a judge of AI agent answers. On UniProt, the expert ontology flags more…
IBM: Asking the Client Questions Before Generation Raised Smart Contract Accuracy From 54% to 73%
IBM researchers tested where AI agents lose the client's intent when writing smart contracts. Questions before generation beat any post-hoc checks: 72.9% vs 53.9%.
Knowledge Conflict in RAG: Why You Need Probability of Context Correctness, Not a Rigid Algorithm
Researchers Find AI Agents Confuse Old and New Facts in Memory Even After Correct Retrieval
Skills for AI Agents Aren't Universal: Weak Models Need Their Own Version of the Prompt
Study: A Single Random Failure Makes AI Agents Abandon a Working Tool 11.6% of the Time
Pretrained AI Image Generators Learn to Fill In Any Region at a Fraction of the Cost
Researchers at Stony Brook and ÉTS Montréal converted pretrained autoregressive image generators into any-order models. Adaptation costs over an order of magnitude…
Plato's Hypothesis Proven for JEPA: A 109M Text Model Beats Qwen2.5-72B
Researchers proved when networks trained on different modalities learn the same world representation. A 109M text model aligned with a vision model better than a…
8B-Parameter Agent Operates iOS Apps at Gemini 3.1 Pro Level
ByteDance Seed proposes AIR learning rate schedule: final loss 0.01 lower than with cosine
METR Experiment Recalculated: On Complex Tasks and Reviews, Developers Using AI Lose 24% of Their Hours
Google Taught Language Models to "Sleep" and Move Knowledge from Memory into Weights
Can You Teach a Language Model While It's Reading? Researchers Tested Every Architecture
Researchers made a model fine-tune itself on the fly as it reads text — and found the trick only works for classic transformers, while hybrid and diffusion models…
Who Wins the Instruction Conflict: Position, Chat Role, or Just a Label?
Researchers tested 14 AI models across 1,080 conflicting-command scenarios and found that models almost always obey whoever spoke first — not whoever formally…