Cognition all the way down 2.0

Research summary

One paper, summarised for reference. Cognition all the way down 2.0: neuroscience beyond neurons in the diverse intelligence era — Robert Chis-Ciure & Michael Levin, Synthese (2025) 206:257. doi:10.1007/s11229-025-05319-6.

The problem it addresses

The basal cognition debate (a strand of the Diverse Intelligence program) is stuck on a definitional deadlock: proponents ascribe “learning”, “memory”, and “decision-making” to cells, tissues, and molecular networks; skeptics charge that these are metaphorical dilutions of concepts whose semantics were fixed for brains. Chis-Ciure & Levin try to dissolve the deadlock by replacing the umbrella concept “cognition” with a single, substrate-agnostic, measurable quantity: how efficiently a system searches its problem space relative to blind chance.

A lexicon for problem spaces

They extend Newell & Simon’s (1972) symbolic problem-solving into a scale-agnostic quintuple:

$$P = \langle S, O, C, E, H\rangle$$

• $S$ — physically realisable states (incl. initial $S_{\text{init}}$ and goal $S_{\text{goal}}$).
• $O$ — operators, elementary transitions $o : S \to S$, each costed by a weight $w : O \to \mathbb{R}_{\ge 0}$ (in physical units — Joules, ATP).
• $C$ — constraints, forbidden state–operator pairs (the nomologically possible paths).
• $E$ — an evaluation functional $S \to \mathbb{R}$ (a fitness proxy; in a variational reading, free energy).
• $H$ — the horizon, the forward look-ahead depth (set by intrinsic timescales / delay lines).

Constraints, evaluation, and horizon are promoted to first-class elements because biological agents modulate them directly — relaxing a constraint or re-tiling the landscape is itself a fingerprint of intelligence, not just navigation within a fixed space.

Intelligence as search efficiency

Intelligence (in William James’s sense — a fixed goal reached by variable means) is operationalised as a scalar. Let $\tau_{\text{blind}}$ be the expected cost of a maximal-entropy random walk on the admissible graph to reach a goal, and $\tau_{\text{agent}}$ the cost under the agent’s policy. Then:

$$K = \log_{10}\!\left(\frac{\tau_{\text{blind}}}{\tau_{\text{agent}}}\right)$$

$K$ counts orders of magnitude of dissipative work saved relative to chance. $K=0$ is chance; $K=n$ is $10^{n}$-fold more efficient. Because both numerator and denominator scale with the combinatorial size of $P$, $K$ is scale-invariant and normalised for task difficulty; it is in physical work units, and it is additive across conditionally independent sub-problems ($K_{\text{total}} = \sum_j K_j$), hence compositional across nested scales. Crucially, $K$ is only as good as its null model — an unfairly handicapped $\tau_{\text{blind}}$ inflates it — so the random walk must run in the exact same $P$.

Two worked estimates

System	Problem space	$K$	Reading
Dictyostelium chemotaxis	cortical-patch occupancy under a cAMP gradient	$\approx 2.2$	~150–200× faster than a random walk (≈7 bits of path-information)
Planarian head regeneration (BaCl₂)	~2,700-gene expression manifold	$\approx 21$	finds a ~10-gene resistance solution in 37 days vs an astronomical blind search ($10^{21}$×)

Both are computed against deliberately conservative (intelligence-underestimating) null models, so they are lower bounds. Even “simple” organisms prune combinatorial explosions by hundreds- to sextillion-fold.

The claim

The “mark of the cognitive” is best sought not in neurons but in the measurable efficiency with which living systems traverse energy and information gradients to tame combinatorial explosions — “one problem space at a time.” The additive decomposition of $K$ across nested scales pinpoints where and by how much intelligence condenses, rendering the “obvious” superiority over chance measurable and therefore refutable. The account connects downstream to the Free Energy Principle (with $E$ as variational free energy and optimal trajectories as least-action flows) and to algorithmic-complexity / minimum-description-length notions of efficient computation.

Why it’s collected here

Two ideas resonate with this lab’s design at the conceptual level (not the specifics of any one task): the multi-scale competency picture — nested agents each searching their own space — is the same “reservoir within an ensemble” nesting the harness is built around; and the insistence that an efficiency claim is only as good as its explicit null model is exactly the discipline the circular-shift and surrogate null tests enforce on the empirical side.

---

References

Chis-Ciure & Levin (2025), Synthese 206:257 · Newell & Simon (1972), Human Problem Solving · Fields & Levin (2022), Entropy 24, 819 · Lyon (2020), Adaptive Behavior 28, 407 · Friston (2010), Nat. Rev. Neurosci. 11, 127 · Emmons-Bell et al. (2019), iScience 22, 147 · Levin (2022), Front. Syst. Neurosci. 16, 768201.