Forecast First, Ask Later: How DCATS Makes Time Series Smarter with LLMs
DCATS shows how LLM agents can improve time series forecasting by curating auxiliary data, not by inventing a cleverer forecasting model.
Cognaptus Insights
Industry thoughts, trends, and commentary.
From GUI Novice to Digital Native: How SEAgent Teaches Itself Software Autonomously
SEAgent shows how computer-use agents can adapt to unfamiliar software through self-generated curricula, trajectory-level judging, and specialist-to-generalist training.
Scalpels Not Sledgehammers: A New Era of Precision Editing for LLMs
Latent Knowledge Scalpel reframes LLM editing as representation-level surgery: targeted entity updates without rewriting the whole model.
Shattering the Spectrum: How PRISM Revives Signal Processing in Time-Series AI
PRISM shows how symmetric multi-resolution filters can make time-series classifiers smaller, cheaper, and still competitive where temporal signals matter more than architectural fashion.
The Forest Within: How Galaxy Reinvents LLM Agents with Self-Evolving Cognition
Galaxy shows how proactive personal agents may need cognition, system design, privacy handling, and self-repair to evolve as one mechanism rather than four separate features.
From Wallets to Warlords: How AI Agents Are Colonizing Web3
A practical reading of how Web3-AI agents are moving from chat interfaces into infrastructure, DeFi execution, governance, security, and trust mechanisms.
Longer Yet Dumber: Why LLMs Fail at Catching Their Own Coding Mistakes
FPBench shows that many LLM code assistants can detect faulty requirements when prompted, but often fail to question bad premises on their own.
Open-Source, Open Risk? Testing the Limits of Malicious Fine-Tuning
A mechanism-first reading of OpenAI’s malicious fine-tuning study and what it implies for evaluating open-weight model releases.
Reasoning with Both Eyes Open: Why Multimodal Chain-of-Thought Still Trips Up LLMs
Multimodal chain-of-thought looks impressive until visual evidence must be used repeatedly, not merely mentioned.
Thinking in Circles: How Self-Questioning LLMs Learn Without Labels
A mechanism-first reading of SQLM, a self-play post-training method where language models generate their own questions, solve them, and learn from proxy rewards without curated training data.