Browse by topic
- LLM (8)
- AI Agents (8)
- Math (3)
- Analysis (1)
- Algebra & Foundations (0)
- Geometry & Topology (0)
- Probability & Finance (2)
- Physics (0)
- Reading (4)
- Books (0)
- Papers (4)
- Robotics & Embodied AI (1)
- Reinforcement Learning (1)
- Conferences (1)
- ICML 2026 (1)
-
Multi-Agent Orchestration Patterns: A Map
The field guide named orchestration as a sub-problem and gave you the reliability math; this post gives you the taxonomy the math was missing — Orchestrator-Workers, Multi-Agent Debate, Mixture-of-Agents, and Reflexion, compared on who talks to whom and which failure mode each actually targets.
-
The Latent / Continuous Reasoning Landscape: A Map
Five real papers, one question: what do you do instead of verbalizing a reasoning step? A survey-style map (comparison table, tradeoffs, one-line thesis) to sit alongside the deep-dive posts — Pause Tokens, CoCoMix, Soft Thinking, Coconut, and CoGHP, compared on what each actually generates, whether it needs training, and whether it's inspectable.
-
Coconut: What It Actually Looks Like to Reason Without Words
CoGHP showed latent chain-of-thought works in a control domain; the faithfulness post argued verbalized CoT is a real liability. Coconut is the LLM-native mechanism that closes the loop — reasoning as fed-back hidden states instead of decoded tokens — plus the honest cost: what you give up to get there.
-
Does the Chain of Thought Say What the Model Actually Did? The Faithfulness Problem
Explicit chain-of-thought reads as a reasoning trace, but two papers show it can be a post-hoc story instead — and, more troublingly, that more capable models tend to produce less faithful explanations. What 'faithful' means, the two core results, and why this is the real argument for latent reasoning.
-
Ambient Attention for Long-Running Agent Sessions: kmux-observer
Dashboards fail for watching multiple long-running AI agent sessions — context-switch cost, staleness, visual noise, disconnection from the actual session. kmux-observer's answer: treat sessions like browser tabs, not dashboard rows. What it is, what's actually built (v0.1) vs. planned, and how it connects to the agent-loop escalation question.