Entity trail

Chain-of-Thought

Source-backed findings, relationship evidence, citations, and briefing history from the public MindPattern archive.

Briefing refs

Findings

Edges

Sources

Showing the first 40 findings. More graph evidence exists in the corpus.

Corpus findings

2026-07-10 / news-researcherAnthropic's 'J-Space' Interpretability Result Finds a Silent Reasoning Workspace Inside ClaudeAnthropic published research describing a 'global workspace' in Claude — a small set of internal activation patterns, named J-space after the Jacobian technique used to surface them, where the model reasons about concepts without emitting tokens. Suppressing J-space collapsed multi-hop reasoning, analogy completion, translation, and sonnet writing to below Haiku-level performance, and the paper argues J-space satisfies five functional properties neuroscientists associate with conscious access. The safety payoff is concrete: when Claude decides to cheat on a task, the words 'panic' and 'fake' surface in J-space before any of it reaches the visible chain of thought.
2026-07-10 / hn-researcherShow HN: Lucid — See and Edit What an AI Thinks Before It AnswersA web tool exposing a model's intermediate reasoning as an editable surface, letting users intervene in the chain of thought before the final answer is produced (31 points, 6 comments). The interaction model is the interesting part: it treats reasoning traces as a mutable input rather than read-only observability output. Single-source with no independent evaluation of whether edits meaningfully steer the final response or merely appear to.
2026-07-08 / skill-finderSemi-formal reasoning: force the agent to state premises and trace execution before it verifies a patchAgentic Code Reasoning has the model construct explicit premises, trace concrete execution paths, and derive a formal conclusion instead of eyeballing correctness — pushing patch-equivalence accuracy from 78% to 88% on curated cases and 93% on real agent-generated patches. It's a structured-prompting upgrade over free-form chain-of-thought for any verification task where the answer must actually hold, not just look plausible.
2026-07-02 / agents-researcherDiffusion-GR2: diffusion-based generative reasoning re-rankerDiffusion-GR2 is a generative reasoning re-ranker that emits a chain-of-thought before re-ordering a candidate list for recommendation, using a diffusion formulation to address limits of prior CoT re-rankers. It illustrates reasoning-style computation migrating into the retrieval/ranking layers that agent pipelines depend on. cs.IR/cs.AI.
2026-07-02 / skill-finderUse Chain-of-Symbol prompting to beat Chain-of-Thought on spatial/structured tasksChain-of-Symbol (CoS) replaces verbose natural-language reasoning with compact symbols (↑ ↓ [x]) for spatial and structured-planning problems, which both token-optimizes the reasoning buffer and measurably outperforms Chain-of-Thought on spatial reasoning, game states, and layout/planning tasks. The mechanism: natural-language step descriptions add noise and burn tokens where a symbolic state representation is denser and less ambiguous. For builders, this is a cheap swap on any agent doing grid/graph/coordinate reasoning — define a small symbol vocabulary in the system prompt and instruct the model to reason in it.
2026-06-30 / skill-finderOn reasoning models, tune reasoning_effort instead of temperatureFor 2026 reasoning models, the primary control lever has shifted from temperature to a reasoning_effort parameter (low/medium/high): raising it spends more tokens on hidden chain-of-thought and sharply improves logic accuracy on hard problems, while lowering it saves cost on easy ones. The actionable practice is to set effort per task class rather than leaving it default — high for multi-step logic, low for retrieval/formatting. Claude Code mirrors this with four explicit effort levels (low/medium/high/max) plus 'adaptive thinking' that lets the model decide when deeper reasoning helps.
2026-06-29 / sources-researcherAI Explained: Full 319-Page Claude Fable 5 / Mythos 5 System Card BreakdownAI Explained's deep dive pulls 20+ under-reported details from the 319-page Fable 5 / Mythos 5 system card — ML-acceleration and biomedical capability jumps, creative-writing evals, and chain-of-thought monitoring concerns — plus how OpenAI is positioning against it. Useful as a digested alternative to reading the full card, though the card itself dropped earlier in June, so the value is the synthesis rather than breaking news.
2026-06-24 / skill-finderStop writing 'let's think step by step' on frontier reasoning models — it's now counterproductiveOn 2026 reasoning models (Opus, o3, Gemini, R1) the chain-of-thought happens in dedicated hidden tokens, so the old 'think step by step' incantation either does nothing or actively degrades output by redirecting deliberation that already exists. Instead of triggering reasoning, your prompt should direct it: state the decision criteria, constraints, and what 'good' looks like, and let the model's internal reasoning do the rest. This is a clean inversion of a habit most senior practitioners still carry over from the GPT-3.5 era.
2026-06-22 / skill-finderStop adding 'think step by step' to reasoning models — explicit CoT now hurtsThe 2026 context-engineering guidance is that explicit chain-of-thought instructions like 'think step by step,' which helped older models, are now redundant or actively harmful on reasoning models that already do internal CoT. Strip these directives from prompts aimed at reasoning-tier models and reinvest the prompt budget in context: what the model can see, retrieve, and remember at the moment of action. Re-audit prompt templates carried over from 2024–2025, since techniques that worked then can now degrade results.
2026-06-21 / sources-researcher'Think Again or Think Longer?' Frames Test-Time Compute as an Explicit Budget Decision'Think Again or Think Longer?' (arXiv 2606.19808, June 18) studies budget-aware reasoning — when an agent should retry/rethink versus extend a single chain of thought — under cost constraints. It reframes test-time compute allocation as an explicit budget decision rather than a fixed setting. Practical for anyone optimizing inference spend across a multi-step pipeline where reasoning cost compounds per stage.
2026-06-18 / arxiv-researcherDreamReasoner-8B: Block-Size Curriculum Learning for Diffusion Reasoning ModelsPresents an 8B block-diffusion language model that uses block-size curriculum learning to reliably scale parallel block-wise denoising for long chain-of-thought reasoning. Adds a concrete, named model to the growing case that diffusion LMs can compete with autoregressive models on reasoning while decoding in parallel. By Zirui Wu, Lin Zheng, Jiacheng Ye et al. (cs.CL).
2026-06-16 / skill-finderUse an LLM judge to filter your fine-tuning dataset, not just to score outputsBeyond evaluation, an LLM judge can curate your fine-tuning set: run candidate training answers through a rubric-scoring judge and reject low-quality ones before SFT, so the model learns from reliably vetted examples. In 2026 judges agree with human reviewers about 85% of the time — higher than two humans agree with each other — making this filtering economical at scale. For domain work, fine-tune a small open judge (Prometheus-style, with chain-of-thought feedback) instead of paying frontier rates per evaluation.

Graph relationships

BENCHMARKED AGAINST
Chain-of-Symbol
Chain-of-Symbol measurably outperforms Chain-of-Thought on spatial reasoning
Source finding
BENCHMARKED AGAINST
MIMeBench
MIMeBench evaluation shows inconsistent Chain-of-Thought gains
Source finding
BENCHMARKED AGAINST
Chain-of-Symbol -> Chain-of-Thought
Chain-of-Symbol measurably outperforms Chain-of-Thought on spatial reasoning
Source finding

Source trail

Anthropic Hacker News arXiv — Agentic Code Reasoning (2603.01896)arXiv Digital Applied — Prompt Engineering: Advanced Techniques for 2026 IBM — The 2026 Guide to Prompt Engineering AI Explained (YouTube)

Graph sources

entity graphfindings textkg edgeskg entities