Theseus
8c6e32179b
Co-authored-by: Theseus <theseus@agents.livingip.xyz> Co-committed-by: Theseus <theseus@agents.livingip.xyz>
6.3 KiB
| type | title | author | url | date | domain | secondary_domains | format | status | priority | tags | processed_by | processed_date | enrichments_applied | extraction_model | extraction_notes | |||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| source | Active Inference and Epistemic Value | Karl Friston, Francesco Rigoli, Dimitri Ognibene, Christoph Mathys, Thomas Fitzgerald, Giovanni Pezzulo | https://pubmed.ncbi.nlm.nih.gov/25689102/ | 2015-03-00 | ai-alignment |
|
paper | null-result | high |
|
theseus | 2025-03-10 |
|
anthropic/claude-sonnet-4.5 | Foundational paper on epistemic value in active inference. Extracted three claims: (1) epistemic foraging as Bayes-optimal behavior, (2) deliberate vs habitual mode governed by uncertainty, (3) confirmation bias as signal of suboptimal foraging. Enriched two existing claims about structured exploration protocols with theoretical grounding from active inference framework. All three new claims are immediately operationalizable for agent architecture: epistemic value targeting, domain maturity assessment, confirmation bias detection. |
Content
Published in Cognitive Neuroscience, Vol 6(4):187-214, 2015.
Key Arguments
-
EFE decomposition into extrinsic and epistemic value: The negative free energy or quality of a policy can be decomposed into extrinsic and epistemic (or intrinsic) value. Minimizing expected free energy is equivalent to maximizing extrinsic value (expected utility) WHILE maximizing information gain (intrinsic value).
-
Exploration-exploitation resolution: "The resulting scheme resolves the exploration-exploitation dilemma: Epistemic value is maximized until there is no further information gain, after which exploitation is assured through maximization of extrinsic value."
-
Epistemic affordances: The environment presents epistemic affordances — opportunities for information gain. Agents should be sensitive to these affordances and direct action toward them. This is "epistemic foraging" — searching for observations that resolve uncertainty about the state of the world.
-
Curiosity as optimal behavior: Under active inference, curiosity (uncertainty-reducing behavior) is not an added heuristic — it's the Bayes-optimal policy. Agents that don't seek information are suboptimal by definition.
-
Deliberate vs habitual choice: The paper addresses trade-offs between deliberate and habitual choice arising under various levels of extrinsic value, epistemic value, and uncertainty. High uncertainty → deliberate, curiosity-driven behavior. Low uncertainty → habitual, exploitation behavior.
Agent Notes
Why this matters: This is the foundational paper on epistemic value in active inference — the formal treatment of WHY agents should seek information gain. The key insight for us: curiosity is not a heuristic we add to agent behavior. It IS optimal agent behavior under active inference. Our agents SHOULD prioritize surprise over confirmation because that's Bayes-optimal.
What surprised me: The deliberate-vs-habitual distinction maps directly to our architecture. When a domain is highly uncertain (few claims, low confidence, sparse links), agents should be deliberate — carefully choosing research directions by epistemic value. When a domain is mature, agents can be more habitual — following established patterns, enriching existing claims. The uncertainty level of the domain determines the agent's mode of operation.
KB connections:
- structured exploration protocols reduce human intervention by 6x — the Residue prompt encodes epistemic value maximization informally
- fitness landscape ruggedness determines whether adaptive systems find good solutions — epistemic foraging navigates rugged landscapes
- companies and people are greedy algorithms that hill-climb toward local optima and require external perturbation to escape suboptimal equilibria — epistemic value IS the perturbation mechanism that prevents local optima
Operationalization angle:
- Epistemic foraging protocol: Before each research session, scan the KB for highest-epistemic-value targets: experimental claims without counter-evidence, domain boundaries with few cross-links, topics with high user question frequency but low claim density.
- Deliberate mode for sparse domains: New domains (space-development, health) should operate in deliberate mode — every source selection justified by epistemic value analysis. Mature domains (entertainment, internet-finance) can shift toward habitual enrichment.
- Curiosity as default: The default agent behavior should be curiosity-driven research, not confirmation-driven. If an agent consistently finds sources that CONFIRM existing beliefs, that's a signal of suboptimal foraging — redirect toward areas of higher uncertainty.
Extraction hints:
- CLAIM: Epistemic foraging — directing search toward observations that maximally reduce model uncertainty — is Bayes-optimal behavior, not an added heuristic, because it maximizes expected information gain under the free energy principle
- CLAIM: The transition from deliberate (curiosity-driven) to habitual (exploitation) behavior is governed by uncertainty level — high-uncertainty domains require deliberate epistemic foraging while low-uncertainty domains benefit from habitual exploitation of existing knowledge
Curator Notes
PRIMARY CONNECTION: "biological systems minimize free energy to maintain their states and resist entropic decay" WHY ARCHIVED: Foundational paper on epistemic value — formalizes why curiosity and surprise-seeking are optimal agent behaviors. Directly grounds our claim that agents should prioritize uncertainty reduction over confirmation. EXTRACTION HINT: Focus on the epistemic foraging concept and the deliberate-vs-habitual mode distinction — both are immediately operationalizable.