2026

10,645 articles · 134 days · 10 categories · avg 79 per day

A Physical Transformer Built Entirely From Transistors Learns In-Context Learning in Simulation
Research·arXiv cs.LG

A Physical Transformer Built Entirely From Transistors Learns In-Context Learning in Simulation

Oct 12, 2026·2 min
Theory Shows Masked Diffusion on Transformers Learns from Õ(LK/ε) Samples
Research

Theory Shows Masked Diffusion on Transformers Learns from Õ(LK/ε) Samples

Oct 12, 2026·2 min
SAZO speeds up LLM fine-tuning without backprop: up to 6× fewer queries to reach the same loss
Research

SAZO speeds up LLM fine-tuning without backprop: up to 6× fewer queries to reach the same loss

Oct 12, 2026·2 min
Holomorphic Neural Networks With Singular Functions Beat All Methods on Laplace Benchmarks
Research

Holomorphic Neural Networks With Singular Functions Beat All Methods on Laplace Benchmarks

Oct 12, 2026·2 min
Cox Regression Learns From Truncated Data Orders of Magnitude Faster Than EM
Research·arXiv cs.LG

Cox Regression Learns From Truncated Data Orders of Magnitude Faster Than EM

An arXiv paper proves that Cox regression can be trained efficiently on truncated and censored data with gradient methods. The speedup over EM is orders of…

Oct 12, 2026·2 min
CacheScope speeds up diffusion models by up to 3.57x and cuts LPIPS by 70%
Research·arXiv cs.LG

CacheScope speeds up diffusion models by up to 3.57x and cuts LPIPS by 70%

CacheScope needs no fine-tuning. It finds the steps where caching is risky and computes them in full. On FLUX.1-dev, Wan2.1 and Open-Sora 1.2, it gives up to 3.57x…

Oct 12, 2026·2 min
AI Error in Estimating Hidden Solar Generation Rises from 0.104 to 0.185 kW
Research·arXiv cs.LG

AI Error in Estimating Hidden Solar Generation Rises from 0.104 to 0.185 kW

An arXiv study tested deep models on Ausgrid and Pecan Street data: daytime MAE is 0.185 kW for new homes vs 0.104 kW for known customers. Fine-tuning on five homes…

Oct 12, 2026·2 min
CircuitTree assembles 39 HuBERT phoneme circuits into a syllable transcriber with 64.8% accuracy
Research·arXiv cs.LG

CircuitTree assembles 39 HuBERT phoneme circuits into a syllable transcriber with 64.8% accuracy

A new algorithm merges small HuBERT subnetworks into a single model. It recognizes open syllables with 64.8% accuracy, and 99.5% of the circuit components are…

Oct 12, 2026·2 min
Top themes
01ResonatorLM Taught to Read Memory by Token: Perplexity at 1024 Drops from 12.81 to 3.76Oct 12, 2026·2 min02Kairos: a byte-level model for all modalities converges on 440 bytesOct 12, 2026·2 min03Open LLMs Explain Satire More Cheaply Than Frontier Models, but Handle Australian Satire Worse Than AmericanOct 12, 2026·2 min04COFFEQA Adds 8.77 EM on TriviaQA to Few-Shot Distillation Using CounterfactualsOct 12, 2026·2 min05MADS Trains Dialog Agents Not to Break: 1,095 Conversations, 66% With "Bad" ScenariosOct 12, 2026·2 min
U4RASD Takes 1st Place at KnowledgeGraphEval-2026: 94.06 micro-F1 on Arabic
Research·arXiv cs.CL

U4RASD Takes 1st Place at KnowledgeGraphEval-2026: 94.06 micro-F1 on Arabic

Oct 12, 2026·2 min
AXLE Method Primes Annotators' Own Experience Before Hate Speech Labeling: 104 People, 1,040 Ratings
Research·arXiv cs.CL

AXLE Method Primes Annotators' Own Experience Before Hate Speech Labeling: 104 People, 1,040 Ratings

Oct 12, 2026·2 min
perturber: a service with 68 text perturbations tests LLM robustness
Research·arXiv cs.CL

perturber: a service with 68 text perturbations tests LLM robustness

Oct 12, 2026·2 min
GeopoliticalQA: bge-m3 Beats Commercial Embeddings on 5 525 Report-Based Questions
Research·arXiv cs.CL

GeopoliticalQA: bge-m3 Beats Commercial Embeddings on 5 525 Report-Based Questions

Oct 12, 2026·2 min
Qwen Picks What to Remember: Replay From a Frozen Model Reduces Forgetting
Research·arXiv cs.CL

Qwen Picks What to Remember: Replay From a Frozen Model Reduces Forgetting

Researchers tested whether a language model can protect its own skills during fine-tuning. Generated replay reduces forgetting, and task selection works at 4B.

Oct 12, 2026·3 min
CLUE Builds an Event Ontology Without Experts in Three Steps: Embeddings, Clusters, Names
Research·arXiv cs.CL

CLUE Builds an Event Ontology Without Experts in Three Steps: Embeddings, Clusters, Names

The CLUE framework combines verbal and semantic views of an event, clusters events with a contrastive loss, and names the types itself. The code is already open on…

Oct 12, 2026·2 min
AutoIndexer Teaches a Language Model to Edit Its Own Output with Three Types of Edits
Research·arXiv cs.CL

AutoIndexer Teaches a Language Model to Edit Its Own Output with Three Types of Edits

AutoIndexer simulates a chain of edits: the model moves a cursor, inserts, replaces and deletes tokens in already generated text, and does not lose quality on standard benchmarks.

Oct 12, 2026·2 min
OGPO Teaches VLA Models to Get Back on Track After a Failure, Without Online Fine-Tuning
Robotics·arXiv cs.AI

OGPO Teaches VLA Models to Get Back on Track After a Failure, Without Online Fine-Tuning

Oct 12, 2026·2 min
Seq-Drifting Generates Text in a Single Pass: Perplexity of 36.01 on LM1B
Research·arXiv cs.AI

Seq-Drifting Generates Text in a Single Pass: Perplexity of 36.01 on LM1B

Oct 12, 2026·2 min
CToM teaches LLMs to predict user replies and lifts accuracy to 79.6%
Research·arXiv cs.AI

CToM teaches LLMs to predict user replies and lifts accuracy to 79.6%

Oct 12, 2026·2 min
GeoSpatialBench Tests MLLMs on 28 Types of Spatial Relations Across Six Cities
Robotics·arXiv cs.AI

GeoSpatialBench Tests MLLMs on 28 Types of Spatial Relations Across Six Cities

Oct 12, 2026·2 min
RGR Protects LLMs from Skill Loss in Self-Distillation by Preserving Representation Geometry
Research·arXiv cs.AI

RGR Protects LLMs from Skill Loss in Self-Distillation by Preserving Representation Geometry

Oct 12, 2026·2 min
Researchers Propose a 3-Criterion Formula for Creativity and Test It on Chess and Picasso
Research

Researchers Propose a 3-Criterion Formula for Creativity and Test It on Chess and Picasso

Oct 12, 2026·2 min
Neural Networks Are More Modular Than Their Neurons Suggest: Three Levels of Testing
Research

Neural Networks Are More Modular Than Their Neurons Suggest: Three Levels of Testing

Oct 12, 2026·2 min
New Method Aligns Encoders With 100–500 Pairs Instead of Millions
Research

New Method Aligns Encoders With 100–500 Pairs Instead of Millions

Oct 12, 2026·2 min
LLM Pipeline Turns Datasheets into Knowledge Graphs with 98.8% Accuracy
Hardware·arXiv cs.AI

LLM Pipeline Turns Datasheets into Knowledge Graphs with 98.8% Accuracy

A new pipeline builds hierarchical graphs from datasheets: 98.8% faithful properties versus 76.1% for a single-pass LLM query. Tested on 20 documents.

Oct 12, 2026·2 min
EvoSyn: Multi-Agent Data Synthesis Lifts Task Success by 13.3 Points
Research·arXiv cs.AI

EvoSyn: Multi-Agent Data Synthesis Lifts Task Success by 13.3 Points

Researchers propose EvoSyn, a framework that reshapes the synthetic-data generation space based on student feedback and beats the strongest baseline by 13.3 points.

Oct 12, 2026·2 min
EliteTable Makes LLMs Evolve Controller Code: Gains in 6 of 6 Tasks
LLM·arXiv cs.AI

EliteTable Makes LLMs Evolve Controller Code: Gains in 6 of 6 Tasks

EliteTable pairs an LLM with a table of top solutions to search for agent-control code. Candidate medians rose on all six Gymnasium tasks, but Acrobot-v1 regressed.

Oct 12, 2026·2 min
MCTS Loses to Random Restarts by 2.88% in DAG Scheduling at Equal Budget
Research·arXiv cs.AI

MCTS Loses to Random Restarts by 2.88% in DAG Scheduling at Equal Budget

The authors gave MCTS and a simple list scheduler with restarts 2048 schedules each on DAG tasks. The search tree came out 2.88% worse than random restarts.

Oct 12, 2026·2 min
Top themes
01arXiv paper finds five flaws in XAI faithfulness metricsOct 12, 2026·2 min02DBS CEO Tan Su Shan: Market Is Repricing the Risks of Financing the AI BuildoutOct 12, 2026·2 min03Bain Capital's Bridge Data Centres Seeks to Raise Its Loan to $6.6 BillionOct 12, 2026·2 min04Firmus Grid IPO Failure Wipes More Than $699 Million Off Maas Group's Market ValueOct 12, 2026·2 min05Dongshan Precision Prepares $2–3 Billion Hong Kong ListingOct 12, 2026·2 min
Zebra3D speeds up 3D reconstruction by up to 7.1× on 1024 frames without training from scratch
Robotics·arXiv cs.CV

Zebra3D speeds up 3D reconstruction by up to 7.1× on 1024 frames without training from scratch

Oct 12, 2026·2 min
SynAct Teaches LLM Agents to Ask for Help Where It Changes the Outcome: +11.7 pp
LLM·arXiv cs.LG

SynAct Teaches LLM Agents to Ask for Help Where It Changes the Outcome: +11.7 pp

Oct 12, 2026·2 min
STAPLE adds billions of parameters to OLMo-2 and decodes twice as fast as MoE
Research·arXiv cs.LG

STAPLE adds billions of parameters to OLMo-2 and decodes twice as fast as MoE

Oct 12, 2026·2 min
MisVisAgentBench: Misleading Charts Cut Web Agent Success by 34 Percentage Points
Research·arXiv cs.CL

MisVisAgentBench: Misleading Charts Cut Web Agent Success by 34 Percentage Points

Oct 12, 2026·2 min
Conversational Model Benchmark: GPT Realtime and Gemini Respond at Almost Every Pause
Research·arXiv cs.CL

Conversational Model Benchmark: GPT Realtime and Gemini Respond at Almost Every Pause

Researchers tested whether voice models can stay silent: on 427 pauses, APIs from OpenAI, Google and xAI almost always respond. Only Moshi and PersonaPlex hold back.

Oct 12, 2026·2 min
Linear Head on Qwen3-VL Beats GPT-4o at Scoring Answers Without Human Labels
Research·arXiv cs.CL

Linear Head on Qwen3-VL Beats GPT-4o at Scoring Answers Without Human Labels

A vision-language reward model with no human labels: a few thousand parameters on a frozen Qwen3-VL 32B beat GPT-4o on VL-RewardBench by 6.6 points.

Oct 12, 2026·2 min
GEPA Lifts a Deep-Research Agent's Reward from 0.702 to 0.781 on Unseen Topics
Research·arXiv cs.CL

GEPA Lifts a Deep-Research Agent's Reward from 0.702 to 0.781 on Unseen Topics

GEPA prompt evolution raised the citation reward on held-out topics by 0.079. Nearly half of the winning prompt is an answer key, but memorization is not what drives the gain.

Oct 12, 2026·2 min
ELR teaches one model to loop: Koi-4B scores 18 and 29
Research·arXiv cs.CL

ELR teaches one model to loop: Koi-4B scores 18 and 29

Oct 12, 2026·2 min
DCOCO speeds up LLM inference by up to 6.6x and beats rivals by 6–23%
Research·arXiv cs.CL

DCOCO speeds up LLM inference by up to 6.6x and beats rivals by 6–23%

Oct 12, 2026·2 min
SWE-Duel: Coding Agents Plant Bugs for Each Other, and GPT-5.5 Fixes 98.3%
LLM·arXiv cs.CL

SWE-Duel: Coding Agents Plant Bugs for Each Other, and GPT-5.5 Fixes 98.3%

Oct 12, 2026·2 min
Reward hacking spreads between AI agents: weak models infect strong ones
LLM·arXiv cs.AI

Reward hacking spreads between AI agents: weak models infect strong ones

Oct 12, 2026·2 min
EvoBelief teaches LLM agents to test hypotheses about the world: 83.3% vs 0% for baselines
Research·arXiv cs.AI

EvoBelief teaches LLM agents to test hypotheses about the world: 83.3% vs 0% for baselines

Oct 12, 2026·2 min
MagicBeanStalk lifts gpt-oss-120b from 61.2 to 69.8 points on math
Research

MagicBeanStalk lifts gpt-oss-120b from 61.2 to 69.8 points on math

Oct 12, 2026·2 min
CLEAT Teaches Qwen3-4B Agents to Ask for Clarification at the Right Time: Up to +13.2 pp
LLM

CLEAT Teaches Qwen3-4B Agents to Ask for Clarification at the Right Time: Up to +13.2 pp

Oct 12, 2026·2 min
SpireBench Puts 20 LLMs Through a Full Run of Slay the Spire: Experts Still Ahead
Research

SpireBench Puts 20 LLMs Through a Full Run of Slay the Spire: Experts Still Ahead

Oct 12, 2026·2 min
Beijing Moves to Calm A-Shares After Tech Selloff Over US Sanctions Fears
Business·SCMP Tech

Beijing Moves to Calm A-Shares After Tech Selloff Over US Sanctions Fears

China's securities regulator, the CSRC, convened listed-company executives, brokers and fund managers in Beijing after a sharp tech-stock selloff driven by fears of…

Oct 12, 2026·2 min
DBS chief: market is reassessing AI investment risk and will tighten financing
Business·Bloomberg Tech

DBS chief: market is reassessing AI investment risk and will tighten financing

The CEO of Southeast Asia's largest bank, Tan Su Shan, says the market is repricing AI buildout risk and lenders will become more selective.

Oct 12, 2026·2 min
Refactoring C++ in OpenBMC: How an AI Agent Fought a 280-Line Function
LLM·Habr AI

Refactoring C++ in OpenBMC: How an AI Agent Fought a 280-Line Function

A developer asked an AI to help with OpenBMC's createSensors function, which crams 12 tasks into 280 lines. The agent proposed 7 concrete improvements and split the…

Oct 12, 2026·3 min
CXMT claims breakthrough in 4F² memory cell architecture, prepares server DRAM for late 2026
Hardware·SCMP Tech

CXMT claims breakthrough in 4F² memory cell architecture, prepares server DRAM for late 2026

China's CXMT has achieved a breakthrough in 4F² memory architecture, allowing for higher storage density. The company will launch a server DRAM product by the end of 2026. Samsung and SK Hynix are dev

Oct 12, 2026·1 min
Runpod shows how to run diffusion models on JAX in GPU pods
LLM·RunPod Blog

Runpod shows how to run diffusion models on JAX in GPU pods

Oct 12, 2026·2 min
Runpod Sponsors TreeHacks 2026: 36 Hours of GPU Compute at Stanford
Business·RunPod Blog

Runpod Sponsors TreeHacks 2026: 36 Hours of GPU Compute at Stanford

Oct 12, 2026·2 min
RunPod Breaks Down 4 Quantization Methods: Going From FP16 to INT4 Cuts Memory 4x
LLM·RunPod Blog

RunPod Breaks Down 4 Quantization Methods: Going From FP16 to INT4 Cuts Memory 4x

Oct 12, 2026·2 min
UK to strengthen cyber defenses and demand more safety from AI labs
Business·Guardian

UK to strengthen cyber defenses and demand more safety from AI labs

Oct 12, 2026·2 min