When an LLM picks from fixed options, wouldn't it be handy to adjust after the fact whether it thinks first? 🤔 ThinkJev: One Thought, Every Budget for Calibrated Typed Decisions ❓ What is a typed decision? 💡 A choice among a fixed set of options that code then acts on, such as routing a ticket or checking a rule. Reading option probabilities in one forward pass is fast, but it fails on multi-step decisions like date arithmetic. ❓ How does it solve this? 💡 With no training, one Qwen3 model plays two roles: System 1 answers without thinking, System 2 answers after thinking, both on the same cached prompt. Cutting one thought short and closing it also gives an accuracy curve for every budget. ❓ How well does it work? 💡 On Qwen3.5-27B, direct answers score 62.2% and always thinking scores 82.7%. The anytime policy reaches 81.6% using 72% of the thought tokens. At a 5% target, System 1 answers 24.5% of decisions with a realized error of 3.9%. ❓ Any caveats? 💡 The pre-registered savings criterion failed on every model. System 1 confidence is a weak signal for when to think, and calibration must be refit when the thinking budget changes. #LLMReasoning# #Calibration#

AI policy is becoming politicized. Instead of arguing about the identities and affiliations of people in the space, I think it’s better to focus squarely on the quality of their arguments and ideas: What can the models can do, how do they work, and—as AI systems grow more sophisticated—what steps can we take to keep humans in the loop? I enjoyed speaking with @WNGdotorg’s Grace Snell for this deeply-researched piece on AI’s growing capabilities. The bottom line is that it’s getting harder for researchers to monitor LLM reasoning. This is legitimately dangerous and could put new threats to critical infrastructure as AI is trusted with acting autonomously of human oversight. Labs and the U.S. government need a serious plan for what to do about it.

# Hermes Agent Features and Practical Usage 🚀 More than remembering facts, this memory continuously infers who you are from conversation. The result is an agent that understands you, getting more on-target the longer you use it. 📌 Title and Feature URL Title: Honcho Memory URL: 📝 Overview Honcho is an AI-native memory backend that goes beyond simple key-value storage. After each turn it performs "dialectic reasoning" to automatically derive insights about your preferences, communication style, goals, and behavioral patterns, building a user model that deepens over time. 🔧 How It Works ・Dialectic reasoning is multi-pass analysis: Pass 0 produces an initial assessment, Pass 1 self-audits for gaps, and Pass 2 reconciles contradictions into a final synthesis (depth 1-3). ・New users get cold-start queries probing general preferences and goals; returning users get warm-session queries prioritizing current context. ・Where built-in memory is manually curated static facts, Honcho maintains server-side profiles with automatic reasoning, enabling semantic search over conclusions and multi-agent peer isolation. ・Five tools integrate: honcho_profile (read/update peer identity cards), honcho_search (semantic search over memory and conclusions), honcho_context (retrieve session context with summaries), honcho_reasoning (synthesized reasoning at a chosen depth), and honcho_conclude (create/delete conclusions, useful for PII management). 🛠 Practical Usage ・Run `hermes memory setup honcho` for guided setup; config is written to `~/.honcho/config.json` (global) or `$HERMES_HOME/honcho.json` (per-profile). ・Key config keys: `contextCadence` (base-context refresh interval), `dialecticCadence` (LLM reasoning interval), `dialecticDepth` (multi-pass depth), `recallMode` (hybrid / context / tools), `writeFrequency` (async / turn / session), plus `apiKey` / `peerName` / `aiPeer` / `workspace`. ・In the default hybrid mode, base context and dialectic supplements inject into the system prompt automatically while tools stay available. ・Once Honcho is active, subcommands like `hermes honcho status` become available. 🎯 Use Cases ・As a long-term assistant, track a user's working preferences across months. ・Keep a coding assistant and a personal assistant talking to the same user with independent models, preventing context bleed. ・Reduce repetitive topic recaps and raise suggestion accuracy via session-scoped injection. ⚠️ Caveats ・Cost scales with reasoning depth (depth 2-3 means more LLM calls; balance via cadence settings). ・New peers need a background prewarm; if it doesn't complete in time, a bounded synchronous fallback applies. ・In recallMode, tools mode gives the agent control but needs explicit reasoning calls, while context mode hides tools and reduces flexibility. State is server-side, so migrating from file memory requires a data export. #HermesAgent# #Memory#

🤔 Even for simple factual questions that need no logic, making a model "think" (CoT) somehow raises accuracy. Google Research explains why. Title: Thinking to Recall: How Reasoning Unlocks Parametric Knowledge in LLMs URL: Reasoning is not just task decomposition; it is also a device for pulling knowledge out of the weights. Three highlights. 🧮 Computational buffer The reasoning tokens themselves act as extra room for latent internal processing. In fact, replacing a meaningful trace with meaningless repeated text ("Let me think") of the same length still improved recall over doing nothing, though it never matched natural reasoning and plateaued when stretched too far. 🔗 Factual priming Generating related facts first primes recall of the correct answer, much like spreading activation in human cognition. Conditioning on just the facts extracted from a trace recovered most of reasoning's gains and helped even with reasoning disabled. ⚠️ Fragile to hallucination The mechanism is a double-edged sword: a single hallucinated fact in the trace sharply reduces correct final answers. Conversely, simply selecting hallucination-free trajectories at test time considerably improved accuracy. The practical takeaway: training with process rewards that encourage factually-grounded intermediate steps could boost reliability and cut hallucination vulnerability. #LLM# #Reasoning#