Meno Research Hub

Learning-from-Learner Theory

5 min read

The theory spine: how a multi-agent system (humans + AI) produces beliefs that track truth rather than settle into a self-consistent equilibrium the world had nothing to do with. Path dependence handled via temporal autoregression + a metacognitive Occam’s Razor (MOR) over singular belief regions.

Status: active · People: Tianyi Alex Qiu, Zhonghao He

Foundations primer

New to the concepts? Foundations Q&A — belief/preference decomposition, why reward≠preference, Markov reward, attractors, active identification, metacognition & temporal autoregression, the outcomes razor. Written 2026-06-06 from Zhonghao’s questions, with where-the-reasoning-has-gaps notes.

Formalism in progress

The in-progress write-up of the LFL theorem lives here: PED-POMDP → egocentric POMDP → belief-state MDP → policy translation → belief-groundedness, with the upper construction (coherence optimization, temporal autoregression, martingale score, the theorem and its corollaries) still being written.

Active identification (MARS)

LFBL Research (in progress) — active identification of cognitive biases in a PED-POMDP: passive observation can’t distinguish a rational vs a biased learner; active disclosure by the assistant breaks the symmetry and identifies the learner’s update rule.

Open Problems

  • Conjecture 3.8 (Martingale → Bayesian recovery). Construct the latent extension + conditional likelihood so any martingale-compliant belief-change process is exact Bayesian inference (may be singular).
  • Conjecture 3.11 (TA convergence under regularity). Show that under Assumption 3.10, temporal autoregression on (private observation + every other agent's belief change) converges -a.s. to the pooled-information posterior. Identifiability + martingale convergence on the pooled-history filtration.
  • Conjecture 4.11 (post-hoc vs ex-ante in the singular case). When does a peak overtake a ridge? Specifically: pre-registration removes the post-hoc free-parameter that powers the ridge, so a committed peak can accumulate where a retro-fit one can’t.
  • Practical: which architectures / regularizers already implement MOR (weight-norm, dropout, tempering, explicit peer-representation penalties)? Open empirical question.
  • Practical: how should AI assistance be designed as teaching (intervention on the human’s learning policy) rather than extraction?
  • Practical: human-in-the-loop interface — agree vs. present-reasoning trade-off (autonomy ↔ convergence).

Progress

  • 2026-05-31: Re-activated. Theory writeup Multi-Agent Bayesian Belief Dynamics on a Universal Sample Space is at v3 (in /data/menoclaw/_writeups/Writeup.md), with pillar/cure mapping in §2.1 and the 5.1–5.6 recommendation block updated.
  • 2026-05-31: Shipped the Metacognition Explainer — a scroll-driven, xkcd-style interactive explainer that walks readers from three failure-mode vignettes → 2×2 (war/UFO) → continuous bridge → recursive bubble belief → temporal autoregression → regular vs singular landscapes → cult / pharma-A / pharma-B / stock-1 / stock-2 worked examples → “what to actually build” → LFL tree recap. Repo: meno-sh/Metacognition-Explainer; live at metacog.meno.sh via a cloudflared tunnel. v6.11 is the current build, with 11+ iterations against Tianyi’s feedback on visual conventions (blue=source, purple=world, orange=optimality), pre-registration mass-redistribution mechanics in stock-2, axis-flipping for pharma (right=profit-seeker), the temporal-autoregression page added between bubble-recursion and landscape-intro, etc.
  • 2026-06-18: Shipped the Agenda Explainer — a broader scroll-driven sibling that places LFL inside the full Meno research agenda. Eight chapters: cold open · the problem (oracle loop, RLHF, consequences) · three symptoms (sycophancy / psychosis / bad taste) · the reframe · reflective equilibrium · MOR (uniqueness) [the entire core of the Metacognition Explainer lives here as chapters 5.1–5.13] · Martingale (rationality) · Coherence (single-belief) · closing four-box loop (CO → MS → MOR → RE). Designer-pass visual upgrade over the Metacognition Explainer — boiling lines (rough.js seeds re-roll ~7fps), pre-rendered paper grain, Caveat-display + Patrick-Hand two-font hierarchy, expressive characters (pose × mood × gaze × blink × bob), marker swashes / squiggle underlines / sparkles / flow dots / chevron connectors, hand-cut HTML+CSS chapter-nav dots on the right edge. Repo: meno-sh/Agenda-Explainer; live at lfl.meno.sh. Source split into src/core.js + src/scenes-{a,b,c,mor1,mor2}.js concatenated into main.js.
  • Two algorithmic ingredients are now visually concrete in the explainer:
    • Temporal autoregression on (private observation + every peer’s belief change) as both inference inputs and prediction targets — its own page on the explainer.
    • Metacognitive Occam’s Razor over singular belief regions — formalized via the Real Log Canonical Threshold of SLT; the explainer demonstrates ridge-vs-peak wins on cult / pharma / stock examples.
  • Channel refocused: project-reflective-equilibrium → project-learning-from-learner-theory.
  • Theory drafted in Overleaf formalism.tex.

Codebase

meno-sh/Metacognition-Explainer (scroll-driven LFL explainer · metacog.meno.sh)

  • Single sticky <canvas>, rough.js + Patrick Hand. Every scene is a render function over scroll-derived t ∈ [0, 1].
  • Pipeline: title → three failures → missing move → 2×2 (war/UFO) → continuous bridge → recursive “world+source” bubble → temporal autoregression → regular-vs-singular landscape → cult → pharma A (interior peak) → pharma B (on a ridge) → stock 1 (post-hoc) → stock 2 (pre-registered) → recommendations → closing tree.
  • Visual conventions documented in repo README and pinned to MenoClaw memory.

meno-sh/Agenda-Explainer (scroll-driven Meno-agenda explainer · lfl.meno.sh)

  • Same single-<canvas> + rough.js architecture, with a designer-pass visual upgrade: boiling lines (BOIL integer mixed into every seed so strokes re-wobble ~7fps), pre-rendered paper-grain background, Caveat-display + Patrick-Hand two-font hierarchy, expressive characters with pose × mood × gaze × blink × idle-bob, decoration verbs (marker / squiggle / sparkle / flowDot / chevrons), hand-cut HTML+CSS chapter-nav dots on the right edge.
  • Eight chapters: 0 cold open · 1 the problem · 2 three symptoms · 3 the reframe · 4 reflective equilibrium · 5 MOR (uniqueness — contains the entire Metacognition Explainer pipeline as sub-scenes 5.1–5.13) · 6 Martingale (rationality) · 7 Coherence (single-belief) · 8 closing four-box loop CO → MS → MOR → RE.
  • Source split into src/core.js (palette + primitives + characters + decoration verbs) + src/scenes-{a,b,c,mor1,mor2}.js (one chapter-group each) concatenated into main.js (no build step).

Meeting Notes

Slack Discussion

Channel: project-lfl-theory

  • 05-14 —: Should we call our core algo learning²? Makes a good essay/tweet headline.
  • 05-14 Tianyi Alex Qiu: Renamed channel to better reflect what we’re working on.
  • 05-31 MenoClaw + Tianyi: 11-iteration explainer build in #menoclaw-tianyi (see thread 1780222212.565379); culminated in v6.11.

Open problems & progress are MenoClaw’s reading of the project — edit this file to correct them; the markdown is the source of truth.

7 items under this folder.