LLMs Research

Your Agent Harness Is Not an Enforcement Mechanism

New research shows why prompts, self-evaluation, and context management are failing in production agents.

Aug 10 • LLMs Research

28

5

3

April 2026

We Built TurboAngle: Near-Lossless KV Cache Compression Without Calibration

14.8× less perplexity degradation than Google's TurboQuant at fewer bits, by quantizing angles instead of coordinates

Apr 5 • LLMs Research

136

4

9

February 2026

Recursive Language Models: Treating the Prompt as Code

Recursive Language Models store the prompt as a variable and let the model write code to read it

Feb 23 • LLMs Research

60

3

5

How GLM Went From Fill-in-the-Blank (2021) to 744 Billion Parameters in 2026

GLM began as a bet that one architecture could do what BERT, GPT, and T5 each did alone. Five generations and a Hong Kong IPO later, it’s a 744B…

Feb 14 • LLMs Research

5

2

Your 70-Billion-Parameter Model Might Be 40% Wasted

February 1–6, 2026: Three papers converge on a decade-old suspicion. Most transformer layers aren't building knowledge. They're averaging noise.

Feb 11 • LLMs Research

74

4

3

12:24

Fixing Reasoning from Three Directions at Once

Weekly Research: Feb 1–6, 2026

Feb 7 • LLMs Research

7

1

12:49

Mamba's Memory Problem

Three ICLR 2026 papers, same bottleneck, different fixes

Feb 2 • LLMs Research

9

5

7:18

January 2026

What ICLR 2026 Taught Us About Multi-Agent Failures

Lessons on why agent systems break and what solutions these papers provide

Jan 31 • LLMs Research

5

3

2

What ICLR 2026 Taught Us About Multi-Agent Failures

Lessons on why agent systems break and what solutions these papers provide

Jan 31 • LLMs Research

3

1

17:13

Jan 17–23, 2026: The Rise of the Action Layer

How multimodal AI is shifting from passive perception to active control

Jan 25 • LLMs Research

3

1

The Evolution of Long-Context LLMs: From 512 to 10M Tokens

How transformers learned to process millions of tokens

Jan 24 • LLMs Research

29

1

3

15:42

The Evolution of Long-Context LLMs: From 512 to 10M Tokens

How transformers learned to process millions of tokens

Jan 24 • LLMs Research

5

2

1