#Research

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

A training-free method has a base model use its SFT version's answer as context. Across 19 model pairs, general-skill losses average no more than 2.2 percentage…

Oct 9, 2026·3 min

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

Lapras, a new post-training method, teaches time series models to reason in latent space: up to +10.79% F1 over CoT and 23.9x fewer generated tokens.

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

Analysis of 127 833 clinician queries to an AI assistant: diagnosis is only 3.7%, and a typical benchmark matches the real workload by just 31%.

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…

01Three LLMs Follow Prompts for Stutters and Interruptions Poorly; Rules Do BetterOct 9, 2026·2 min 02Sociolinguistics Instead of LLM Judges: Testing User-Simulator Personas on 5 AgentsOct 9, 2026·2 min 03AI4Fire: SQL Tool Lifts LLM Accuracy on Wildfire Databases From 16% to 88%Oct 9, 2026·2 min 04LLMs Judge Translation Across 70 Language Pairs: They Beat Humans Only in PlacesOct 9, 2026·2 min 05MASA Reframes LLM Sparse Attention as Matrix ApproximationOct 9, 2026·2 min

Top themes

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

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

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

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.

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

A new benchmark of 82 unsolved math and theoretical physics problems: the top model, GPT-6-Astra, scored 14.0%, while full-size open models scored 5.5–6.7%.

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

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

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

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

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

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.

New Framework Infers Column Types From Headers Alone: 39 Types and 120,000 Columns

An arXiv paper proposes an explainable framework that infers data types and rates source quality from column headers alone. It was tested on about 120,000 columns.