2026

Robotics

MESSENGER cuts 3D motion forecasting error by 71.6% on nuScenes

LinSlot Teaches a Neural Network to Break a Scene Into Objects and Their Attributes Without Labels

OcuFlow Removes Eyeglass Glare in One Step and Wins 67.32% of Votes in a Blind Test

Research·arXiv cs.CV

Researchers release VR180 dataset shot on Blackmagic cameras: 1211 stereo samples

First VR180 dataset captured directly on Blackmagic URSA Cine Immersive: 636 videos, 575 photos, factory calibration, and a finding that fisheye needs 38% more…

Research·arXiv cs.CV

SLVR Adds 14.2 Points on BLINK Relation to Qwen2.5-VL-7B Without Text CoT

Researchers propose SLVR: latent reasoning with typed stages. On Qwen2.5-VL-7B, the method added +9.4 on MMVP and +14.2 on BLINK Relation without generating…

Business·AI Business

Google Releases a Workplace AI Agent for Enterprises, but Pricing and Integrations Remain Open Questions

Google's agent uses company context for office tasks, but pricing and compatibility with third-party systems remain the key questions for businesses.

Researchers Prove Why Adversarial Training Makes Saliency Maps Sparse

A new arXiv paper proves that for two-layer ReLU networks, the gradient-norm penalty under ℓ∞ attacks is anisotropic and pushes saliency maps toward sparsity.

01Distilling ECG into PPG Memorized Identity: Fixed-Effects More Than Doubled the State SignalOct 9, 2026·3 min 02Subliminal learning passed a backdoor through numbers: 23.5% of replies in FrenchOct 9, 2026·2 min 03PXtal Beats Baseline Models on Six Benchmarks, Four of Them Zero-ShotOct 9, 2026·2 min 04Nullify Erases Knowledge From LLMs Without Retraining, Using Null-Space VectorsOct 9, 2026·2 min 05Multimodal Alzheimer's AI model reaches 0.935 AUROC on 2649 patientsOct 9, 2026·2 min

Top themes

CARing teaches LLMs to predict diagnoses: R@30 hits 46.52% on MIMIC-IV

LLM·arXiv cs.LG

Visible CoT Does Not Universally Improve SQL and pandas Code Generation, New Study Finds

D-SLR Replaces Truncated SVD With Three Decompositions and No Tuning

Hardware·arXiv cs.LG

Phase-HDC trains a model with no optimizer history and stores 16–23x less than Adam

Robotics·arXiv cs.LG

Kolmogorov-Arnold Networks Match MLP Performance With 40% Fewer Samples

In reinforcement learning, KANs reach the same results as MLPs with 40% fewer samples. Peak relative gain is up to 50%, and reward noise does not erase the…

IKNO for Navier–Stokes Stays Accurate Under Rotation, With 5.6x Fewer Parameters Than TFN

Researchers propose IKNO, a compact graph model whose predictions strictly follow coordinate rotations in 3D flows. It matches TFN accuracy with 5.6× fewer…

LoopOPD Teaches Looped Language Models to Improve Through Their Own Loops

LoopOPD makes an intermediate loop of a LoopLM learn from the final loop of the same model, with no external teacher. Tested on Ouro-Thinking, trained on math only.

MoE Router Telemetry Leaks Fine-Tuning Data: Attack Gains 2.7–9.4 pp in Accuracy

Medicine·arXiv cs.LG

SPERA Trained on 80,000 Hours of EEG From 29,048 People and Tested on 9 Tasks

80 Synthetic Dialogues Are Enough to Train Activation Probes for Gemma-3-27B, but Not Everywhere

Temporal Transformer for the CAN Bus Catches Anomalies Without Sharing Raw Data

Freezing the Decoder in KV-Cache Compression Saves 3x Memory but Doesn't Improve Accuracy

SFT-as-Context Restores General Skills in Fine-Tuned LLMs, Within 2.2 Points

Lapras replaces CoT with latent reasoning: up to +10.79% F1 with 23.9x fewer tokens

Doctors Don't Ask LLMs About Diagnoses: 127 833 Queries vs. 58 Benchmarks

FedAlphaEdit Merges Knowledge Edits from Different Organizations into One LLM Without Sharing Data

Researchers show that naively combining AlphaEdit and CollabEdit breaks down. Their FedAlphaEdit merges edits in a shared null space and comes close to centralized…

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

The tau-Voice authors tested how three LLMs imitate disfluencies, interruptions and backchanneling in user simulators. A deterministic algorithm proved more…

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

A pilot arXiv study defines simulator personas through measurable style parameters and checks them with stylometry on five customer-service agents.

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

The AI4Fire benchmark ran 35 LLMs on wildfire management tasks. With SQL access, accuracy rose to 88%, but no model beat a simple rule on staffing forecasts.

01LLMs Judge Translation Across 70 Language Pairs: They Beat Humans Only in PlacesOct 9, 2026·2 min 02STI-OPD: A Probabilistic Teacher Intervention for Multi-Step Agentic TasksOct 9, 2026·2 min 03MASA Reframes LLM Sparse Attention as Matrix ApproximationOct 9, 2026·2 min 04Routed Sparse Autoencoders Separate Words From Intonation in Speech on WavLM and SPEAROct 9, 2026·2 min 05Galahad-kv stores KV cache on NVMe and loads a block 2.8–4.3x faster than recomputing itOct 9, 2026·2 min

Top themes

Business·arXiv cs.CL

Fine-Tuned 0.8B-Parameter Model Beats GPT-5.4 at Grammar Annotation

Business·arXiv cs.CL

Utterance Normalization Cuts the Cost of Analyzing Millions of Calls: A Pipeline Built Entirely on Small Models

Text autoencoder compresses web text to 2.24 bits per byte without losing meaning

Cognitive Thermometers: Neural Networks Proposed as a Way to Measure the Complexity of Meaning

Business·arXiv cs.CL

LLMs Draft Road-Work Plans for WisDOT: 14B Barely Beats 7B

Researchers fine-tuned open models on the WisTMP archive and ran them locally. Quality improved, but 14B gained little over 7B/8B, and cost estimates remain a weak…

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

Researchers collected 294 369 Polish press documents from 1800–1918 and trained models from 47M to 349M parameters from scratch. The models know post-1918…

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

A new method for personalizing diffusion models allocates LoRA rank across layers using learnable gates and pruning. Tested on Stable Diffusion and DreamBooth sets.

LLM·arXiv cs.AI

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

LLM·arXiv cs.AI

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

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

LLM

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

Business

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

Research·arXiv cs.AI

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

An arXiv preprint introduces the notion of LLM normative competence. In debates with synthetic norms, baseline agents failed to infer the rules, even when doing so…

Research·arXiv cs.AI

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

Tests on MLE-Bench Lite and Zork I show small agents ignore a time limit stated in the prompt. RL with GRPO holds the budget almost perfectly but doesn't become…

Robotics·arXiv cs.AI

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

A self-supervised method picks keyframes for behavior cloning: +13.9 pp success on 23 manipulation tasks and 80% SR on a real robot at a horizon 20x longer.

Research·arXiv cs.AI

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

The diffusion-based DreamReasoner-8B repairs agent plans with a 53.7% success rate versus 27.0% for Qwen3-8B, and after fine-tuning it generates plans 39–46% faster.