Few know of the agent drift problem in multi-agent systems. But it is one of the most common failure modes in multi-agent LLM systems. The more agents interact with each other, the worse they get. Not because individual models are weak. Because it's typical that something
LLMS
-
AI Capabilities Progress from 2021 to 2025 and Beyond
By
–
2021 : ça ne sait même pas autocompléter une ligne
2022 : ça ne sait même pas écrire une fonction complète
2023 : ça ne peut même pas réussir un entretien de code
2024 : ça ne peut même pas construire une application
2025 : ça ne peut même pas gérer des projets complexes
2026 : -
Vibe Coding Eases Attention Demands of Manual Programming
By
–
This is so true. Programming "by hand" is a *VERY* attention demanding and attention-to-the-details kind of endeavor. It is almost impossible to do right if you are stretched with other obligations in life. Vibe coding changes that.
-
Spurious Rewards Testing Implementation on Qwen-2.5 1.5B
By
–
We provide an implementation to test spurious rewards out on Qwen-2.5 1.5B. Check it out here!
-
Qwen excels at RL finetuning with random rewards
By
–
Qwen is hands-down the best model for RL finetuning Turns out you can even tune it with random rewards and it *still* improves. Paper shows it's not magic, pre-trained code reasoning patterns are uncovered with RLVR We provide hassle-free bash scripts to reproduce this below!
-
MCP’s Missing Piece: Stateful Execution Environments for Agent Work
By
–
2/4 MCP standardizes tool calls (what/how), but it doesn’t carry the missing piece for stateful work: where the call runs. In a coding run, subagents don’t just “think”. They edit → run tests → edit again. If two branches share a working directory, you get nondeterminism fast:
-
QwenLong-L1.5: AI Model Reasons Across 4-Million-Token Documents
By
–
Cool, this AI could reason across a 4-million-token document—like reading 100 novels at once. Enter QwenLong-L1.5. The team trained it using a new method: 1) generating complex, multi-step reasoning questions from documents, 2) a stabilized training process to avoid bias, and
-
Alibaba’s Learnable RL Policy for LLM Context Management
By
–
This new paper from Alibaba Group makes context management into a learnable RL policy So the agent is capable of deciding when to store, update, or delete long-term info and when to retrieve, summarize, and filter short-term context bringing a fresh new way to tackle LLMs used
-
AWS Model 100x Smaller Outperforms GPT and Claude on Tool Calling
By
–
Everyone is sleeping on this new paper from AWS. A model 100x smaller than GPT and Claude crushed them on tool calling. AWS researchers took Facebook's OPT-350M, a model from 2022 with 500x fewer parameters than GPT, and fine-tuned it on ToolBench for a single epoch. The
-
GPT-5.2 Codex Represents Major Step Function Improvement
By
–
gpt-5.2 in codex is a step function improvement