AllLLMRoboticsStartupsResearchEthicsBusinessHardwareAutonomousMedicineCreative

All news

10724
Three LLMs Follow Prompts for Stutters and Interruptions Poorly; Rules Do Better
Research·arXiv cs.CL

Three LLMs Follow Prompts for Stutters and Interruptions Poorly; Rules Do Better

Oct 9, 2026·2 min
Sociolinguistics Instead of LLM Judges: Testing User-Simulator Personas on 5 Agents
Research

Sociolinguistics Instead of LLM Judges: Testing User-Simulator Personas on 5 Agents

Oct 9, 2026·2 min
AI4Fire: SQL Tool Lifts LLM Accuracy on Wildfire Databases From 16% to 88%
Research

AI4Fire: SQL Tool Lifts LLM Accuracy on Wildfire Databases From 16% to 88%

Oct 9, 2026·2 min
LLMs Judge Translation Across 70 Language Pairs: They Beat Humans Only in Places
Research

LLMs Judge Translation Across 70 Language Pairs: They Beat Humans Only in Places

Oct 9, 2026·2 min
STI-OPD: A Probabilistic Teacher Intervention for Multi-Step Agentic Tasks
LLM·arXiv cs.CL

STI-OPD: A Probabilistic Teacher Intervention for Multi-Step Agentic Tasks

Junnan Liu and co-authors introduce STI-OPD, a language model distillation framework with adaptive teacher intervention. The method outperforms baseline approaches…

Oct 9, 2026·2 min
MASA Reframes LLM Sparse Attention as Matrix Approximation
Research·arXiv cs.CL

MASA Reframes LLM Sparse Attention as Matrix Approximation

A new preprint proposes selecting sparse-attention elements by their contribution to matrix-product error, not by weight magnitude. The method plugs into existing…

Oct 9, 2026·2 min
Routed Sparse Autoencoders Separate Words From Intonation in Speech on WavLM and SPEAR
Research·arXiv cs.CL

Routed Sparse Autoencoders Separate Words From Intonation in Speech on WavLM and SPEAR

A new TopK autoencoder method sends speech content and voice with emotion down separate routes. Tested on two encoders, two corpora and independent probes.

Oct 9, 2026·2 min
Galahad-kv stores KV cache on NVMe and loads a block 2.8–4.3x faster than recomputing it
Research·arXiv cs.CL

Galahad-kv stores KV cache on NVMe and loads a block 2.8–4.3x faster than recomputing it

The galahad-kv package stores the KV state of 50M tokens on encrypted NVMe. On an H100, loading a block is 2.8–4.3x faster than recomputing it and uses 8.8–12.3x…

Oct 9, 2026·2 min
Top themes
01Fine-Tuned 0.8B-Parameter Model Beats GPT-5.4 at Grammar AnnotationOct 9, 2026·2 min02Utterance Normalization Cuts the Cost of Analyzing Millions of Calls: A Pipeline Built Entirely on Small ModelsOct 9, 2026·2 min03Text autoencoder compresses web text to 2.24 bits per byte without losing meaningOct 9, 2026·2 min04Cognitive Thermometers: Neural Networks Proposed as a Way to Measure the Complexity of MeaningOct 9, 2026·2 min05LLMs Draft Road-Work Plans for WisDOT: 14B Barely Beats 7BOct 9, 2026·2 min
Wieszcz-XIX: A 3.1-Billion-Word Polish Corpus From Before 1918 Trains Models From Scratch
Research·arXiv cs.CL

Wieszcz-XIX: A 3.1-Billion-Word Polish Corpus From Before 1918 Trains Models From Scratch

Oct 9, 2026·2 min
Diffu-LoRA Lets the Adapter Decide Which Stable Diffusion Layers Need Rank
Research·arXiv cs.CL

Diffu-LoRA Lets the Adapter Decide Which Stable Diffusion Layers Need Rank

Oct 9, 2026·2 min
GenUI-Harness Lifts a 4B Model From 9.33% to 58% Pass@3, Beating Claude Opus 5
LLM·arXiv cs.AI

GenUI-Harness Lifts a 4B Model From 9.33% to 58% Pass@3, Beating Claude Opus 5

Oct 9, 2026·2 min
OpenProblemBench: GPT-6-Astra Solves 14% of Open Math and Physics Problems
Research·arXiv cs.AI

OpenProblemBench: GPT-6-Astra Solves 14% of Open Math and Physics Problems

Oct 9, 2026·2 min
AgentHorizon: Best Judge GPT-5.5 Scores Only 80.9% on Long Agent Tasks
Research·arXiv cs.AI

AgentHorizon: Best Judge GPT-5.5 Scores Only 80.9% on Long Agent Tasks

The new AgentHorizon benchmark tests whether LLM judges can grade long computer-use agent runs. The best result: 80.9% balanced accuracy from GPT-5.5.

Oct 9, 2026·2 min
Two Distillation Schemes Make AI Students Reveal Their Teachers' Hidden Behavior
Research·arXiv cs.AI

Two Distillation Schemes Make AI Students Reveal Their Teachers' Hidden Behavior

An arXiv paper proposes using distillation for AI safety: DFI exposes a model's hidden misalignment, while DFC extracts its skills without the harmful tendencies.

Oct 9, 2026·2 min
Researchers prove that bloating AGENTS.md is harmful: the file has a size limit
LLM·arXiv cs.AI

Researchers prove that bloating AGENTS.md is harmful: the file has a size limit

Preprint arXiv 2610.11007 models an agent's context file as a capacitated assortment problem and proves an upper bound on its optimal size.

Oct 9, 2026·2 min
GUISE Shows Narrative Wrappers Break Qwen3-1.7B in 89–96% of Cases
Research·arXiv cs.AI

GUISE Shows Narrative Wrappers Break Qwen3-1.7B in 89–96% of Cases

Oct 9, 2026·2 min
An LLM Teaches a Planner to Get by With Polynomial Memory and Solves 1709 of 1890 Tasks
LLM·arXiv cs.AI

An LLM Teaches a Planner to Get by With Polynomial Memory and Solves 1709 of 1890 Tasks

Oct 9, 2026·2 min
StoreBench: LLM Agents Run an Online Store and Lose to a Script — 49% vs. 97%
Business·arXiv cs.AI

StoreBench: LLM Agents Run an Online Store and Lose to a Script — 49% vs. 97%

Oct 9, 2026·3 min
LLM Agents Don't Learn Community Norms From Interaction, Study Finds
Research·arXiv cs.AI

LLM Agents Don't Learn Community Norms From Interaction, Study Finds

Oct 9, 2026·2 min
Qwen3.6-27B and Qwen3-4B Agents Blow Their Time Budgets Without Harness Hints
Research·arXiv cs.AI

Qwen3.6-27B and Qwen3-4B Agents Blow Their Time Budgets Without Harness Hints

Oct 9, 2026·2 min
Keyframe Mnemonics Teaches Robots to Remember What Matters: 100% Success on Long Horizons
Robotics

Keyframe Mnemonics Teaches Robots to Remember What Matters: 100% Success on Long Horizons

Oct 9, 2026·2 min
Plan-and-Patch: Diffusion LLM Repairs Agent Plans Twice as Often as Qwen3-8B
Research

Plan-and-Patch: Diffusion LLM Repairs Agent Plans Twice as Often as Qwen3-8B

Oct 9, 2026·2 min
arXiv authors call for replacing a single annotation "truth" with a space of interpretations
Ethics

arXiv authors call for replacing a single annotation "truth" with a space of interpretations

Oct 9, 2026·2 min
TGIT Translates Short Drone Commands and Lifts Navigation Success from 11% to 33%
Robotics·arXiv cs.AI

TGIT Translates Short Drone Commands and Lifts Navigation Success from 11% to 33%

The TGIT instruction translator leaves the frozen OpenFly navigator untouched, yet lifts mission success from 11.33% to 32.51% on real human commands.

Oct 9, 2026·2 min
Agent strikes out its own old tool outputs and cuts input tokens by half
Business·arXiv cs.AI

Agent strikes out its own old tool outputs and cuts input tokens by half

Researchers let an agent archive its own observations: in an OpenTelemetry case, the final prompt shrank from 912K to 232K tokens, and the API bill fell threefold.

Oct 9, 2026·2 min