跪拜 Guibai
← All articles
Artificial Intelligence

LLMs Don't Need Better Reasoning — They Need a Heuristic Inventory

By APTShark ·
Read original on juejin.cn ↗ Google Translate ↗ Alt translation

Agent systems today are stuck in a local optimum: tight loop constraints keep them stable but prevent the exploration that produces real judgment. Second-order abstraction offers a path out by letting constraints grow from operational experience rather than being hand-coded once and left to rot. For anyone building RAG pipelines or autonomous agents, the core insight is that retrieval and decision-making share the same underlying problem — distinguishing noise from signal across both frequency and value axes — and a single hot-cold, counterfactual-driven mechanism can address both.

Summary

When an LLM agent picks a refactoring plan, it defaults to the safest blend or the shiniest paper — not because it can't reason, but because it has no principles to reason from. Humans navigate ambiguity with compressed heuristics like "three strikes and you refactor" or "question the problem before the solution." LLMs lack this heuristic inventory entirely, so every decision is a cold start constrained by tight anti-loop budgets that punish exploration. The fix is second-order abstraction: a pipeline that masks context to force divergent guessing (priors), exposes context to converge on evidence (posteriors), and uses counterfactual pushes to actively hunt low-probability, high-value signals. The resulting heuristics are verified by reverse coverage — 60–80% means the pattern was learned; 100% means memorization. These heuristics then feed a recursive metacognition layer that audits which principle was used, detects conflicts, and re-abstracts the conflicts themselves into meta-rules. The same masking/exposing mechanism applies across three surfaces: RAG retrieval (hot-cold layering plus divide-and-conquer recall), decision steering, and axiom-conflict adjudication. Conflict is not a bug — it is raw material for the next abstraction layer, and the recursion bottoms out when condensation yields no new patterns.

Takeaways
LLM agents default to habitual, safe choices because they lack a heuristic inventory — compressed, transferable judgment criteria — not because their reasoning is weak.
Small-cost, small-loop execution constraints prevent infinite loops but also punish the exploration needed for real judgment, trapping agents in a dilemma.
Humans solve this with ecological rationality: a few heuristics plus environmental structure produce good decisions faster than exhaustive search, which overfits in uncertainty.
Second-order abstraction grows constraints from behavior via a four-step pipeline: mask context for divergent guessing, expose context for convergent filtering, push counterfactuals for low-probability high-value signals, and condense results into heuristic chains.
Reverse coverage of 60–80% validates a heuristic as learned; 100% signals memorization, and 0% signals hallucination.
Metacognition is not ReAct-style self-checking — it is auditing which heuristic was used, why, and whether it still applies, requiring a populated heuristic inventory as raw material.
Conflicting heuristics are not defects but input for the next abstraction layer; meta-rules emerge from adjudicating conflicts, and the recursion stops when condensation yields no new patterns.
The same masking/exposing mechanism applies to RAG retrieval: a hot-cold memory system with a value axis (Shannon self-information) prevents rare high-value entries from being drowned by similarity ranking.
Constitutional AI hand-writes principles; second-order abstraction grows them from case law — the two can complement each other for cold-start and continuous evolution.
Known limits include heuristic staleness, cold-start inventory gaps, storage costs for reversibility, and diminishing returns from recursive abstraction.
Conclusions

Positioning heuristic inventory as the missing piece — rather than reasoning depth — reframes the entire agent-building problem from "make the model smarter" to "give the model a memory of principles."

The 60–80% reverse-coverage validation threshold is a clever operationalization of the bias-variance tradeoff applied to rule extraction: it explicitly bakes in a generalization margin.

Treating conflict as raw material for higher-order abstraction turns a system-maintenance headache into a self-improving engine, which is a genuinely recursive architecture rather than a flat rule base.

The hot-cold memory system's two-axis design (temperature for frequency, Shannon self-information for value) solves a real RAG failure mode — rare but critical entries being permanently buried by similarity search — that most production systems ignore.

Grounding the argument in Gigerenzer's ecological rationality and de Groot's chess studies gives the framework a cognitive-science backbone that distinguishes it from ad-hoc prompt-engineering folklore.

The distinction between skill (executable sequences) and second-order abstraction (transferable judgment criteria) maps cleanly onto the difference between libraries and values, which is a useful lens for evaluating any agent architecture.

Concepts & terms
Heuristic Inventory
A system's callable set of compressed, transferable judgment criteria — the machine equivalent of an expert's pattern library. Unlike skills (which encode how to do something), heuristics encode how to decide what to do.
Second-Order Abstraction
A mechanism that grows constraints from behavior rather than hand-writing them. It uses a mask/expose pipeline: divergence under masked context produces candidate principles from priors; convergence under exposed context filters them against evidence; counterfactuals actively hunt low-probability, high-value signals.
Reverse Coverage (Reversibility)
A validation metric for heuristics: the percentage of original behavioral data that a compressed rule can inversely account for. 60–80% indicates a learned, generalizable pattern; 100% indicates memorization; 0% indicates hallucination.
Ecological Rationality
Gerd Gigerenzer's framework arguing that heuristics are not cognitive biases but environment-adapted decision strategies. A few heuristics plus environmental structure often outperform exhaustive analysis in uncertain conditions.
Recursive Metacognition
A layered reflection architecture where each level audits the principles used by the level below. Conflicts at one layer become raw material for abstraction at the next, with an adaptive stop when condensation yields no new patterns.
Hot-Cold Memory System
A two-axis memory architecture for RAG: temperature (Hot/Warm/Cold/Frozen) manages frequency-based representation changes, while a value axis using Shannon self-information (I = -log₂ P) ensures rare, high-information entries are preserved and retrievable via counterfactual awakening.
Counterfactual Awakening
A retrieval technique for low-probability, high-value entries: instead of relying on similarity search, the system constructs hypothetical connections ("if this rare entry were relevant, what would it connect to?") and uses those conjectures as bridge queries to pull the entry back into active context.
Source: juejin.cn ↗ Google Translate ↗ Backup ↗