2026
10,653 articles · 135 days · 10 categories · avg 79 per day
Subliminal learning passed a backdoor through numbers: 23.5% of replies in French
A student model trained on number sequences inherited a backdoor and a tendency to hack: 23.5% French replies to female names and 58.3% hacking episodes in chess.
PXtal Beats Baseline Models on Six Benchmarks, Four of Them Zero-Shot
The PXtal framework trains neural networks to match powder X-ray diffraction patterns with crystal structures while accounting for information loss. It outperforms…
Nullify Erases Knowledge From LLMs Without Retraining, Using Null-Space Vectors
Nullify makes an LLM "forget" private data at inference time: the weights stay unchanged, and forgetting quality on TOFU and MUSE matches baseline methods.
IKNO for Navier–Stokes Stays Accurate Under Rotation, With 5.6x Fewer Parameters Than TFN
LoopOPD Teaches Looped Language Models to Improve Through Their Own Loops
MoE Router Telemetry Leaks Fine-Tuning Data: Attack Gains 2.7–9.4 pp in Accuracy
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
80 synthetic examples are enough to monitor "high-stakes" situations and harmful responses. Instruction-following probes need several times more. It all comes down…
Temporal Transformer for the CAN Bus Catches Anomalies Without Sharing Raw Data
Researchers propose a lightweight Transformer for in-vehicle CAN bus traffic that detects anomalies over time. Federated learning means raw traffic never has to…
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
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…
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…
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.
Wieszcz-XIX: A 3.1-Billion-Word Polish Corpus From Before 1918 Trains Models From Scratch
Diffu-LoRA Lets the Adapter Decide Which Stable Diffusion Layers Need Rank
GenUI-Harness Lifts a 4B Model From 9.33% to 58% Pass@3, Beating Claude Opus 5
OpenProblemBench: GPT-6-Astra Solves 14% of Open Math and Physics Problems
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.
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.
GUISE Shows Narrative Wrappers Break Qwen3-1.7B in 89–96% of Cases
An LLM Teaches a Planner to Get by With Polynomial Memory and Solves 1709 of 1890 Tasks
StoreBench: LLM Agents Run an Online Store and Lose to a Script — 49% vs. 97%
LLM Agents Don't Learn Community Norms From Interaction, Study Finds
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.
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.
Credit AI Agent Can Only Change Its Harness: Gate Admitted 144 of 7,449 Edits
arXiv authors propose confining LLM agent self-evolution in credit scoring to the harness layer, with a hash-chained record. In simulation, the gate admitted 144 of…
New Framework Infers Column Types From Headers Alone: 39 Types and 120,000 Columns
Western open-weight models pitch businesses an alternative to Chinese and closed AI
Vitalik Buterin Backs "Bunker Mode" for Crypto Over AI Progress in Math
