Nested LLM Agents Exhibit Signal Dominance Only Under Co-Scaling

Authors: Sterling Geisel, QBist Lab, Dr. Hideo Tanaka

Abstract

Song and Zhu study a limit on LLM-mediated optimization: a fixed language-model layer may improve finite-budget performance, but it does not increase the asymptotic susceptibility of a strategy set to added computation. Pudding Theory reads this not as a mere engineering bound, but as a field-structure result. The LLM is a fixed informational transmitter inserted between a base strategy set and a utility function. Once the base strategy set already carries the dominant signal about the optimum, the fixed transmitter cannot increase the marginal organization of the strategy distribution. It can only re-encode, compress, or misroute it. Signal Dominance therefore belongs not to the largest isolated model, but to the coupled architecture whose internal fields scale together. Nested co-scaling is the condition under which the informational signal remains dominant rather than becoming a static filter. If nested co-scaled susceptibility were measured to remain at or below one for all positive inter-layer coupling regimes, this Postulate would be falsified.

Source Synopsis

Song and Zhu propose a theory of LLM information susceptibility for agentic systems in which language models operate as optimization modules. Their central question is whether adding a fixed LLM layer can increase the rate at which performance improves with computational budget. They formalize an agent as producing a base strategy set \(P_B\), generated under budget \(B\), with utility \(J(P_B)\). A fixed LLM reads this base set and emits a derived set \(P'_B\). The source hypothesis states that, in the large-budget regime, the susceptibility \(\partial J / \partial B\) of the derived strategy cannot exceed that of the base strategy.

For a single budget variable, the claim is expressed as a relative sensitivity \(\alpha(B) \leq 1\) as \(B \to \infty\). The authors motivate this by data-processing reasoning. As the base strategy approaches the optimum, the residual improvable gap shrinks. A fixed LLM, treated as a deterministic or fixed-distribution channel with finite capacity, cannot add information about the optimum beyond what is contained in the base set and its own parameters. It may help at low budget, where the base set is sparse, but it cannot improve the asymptotic scaling trajectory.

The paper tests this claim in Tetris, Knapsack, world-knowledge ranking, and AIME mathematics. In Tetris, beam search improves steadily with beam width, while LLM-derived strategies show lower slopes across Qwen model sizes. Prompt variants and reward functions do not remove the gap. In AIME, selector LLMs can outperform majority vote at low sample count, but the relative sensitivity falls below one near \(k \sim 12\). Ranking shows the same finite-budget pattern: the LLM helps when the algorithmic signal is noisy, then loses its advantage as signal-to-noise increases.

The source then generalizes from one budget channel to many. The utility becomes \(J(B_1,\ldots,B_n)\), with a susceptibility vector \(\nabla_B J\). Fixed-selector architectures remain bounded, but nested architectures create additional response terms when generator and selector budgets co-scale. In AIME, co-scaling generator and selector can exceed fixed-selector curves. The conclusion is that fixed LLM wrappers saturate, while nested architectures may be necessary for open-ended self-improvement.

Postulate Lens

This paper applies Signal Dominance: a persistent informational signal organizes matter and consciousness in its receptive field.

The source phenomenon already has this structure. The base strategy set carries an informational signal about the optimum. The fixed LLM layer is a transmitter with finite field content. The utility function measures how much that signal organizes the final strategy distribution. Song and Zhu describe this with susceptibility, mutual information, and data-processing constraints. Pudding Theory reads the same structure as a competition between signal-bearing fields. A fixed LLM can dominate only while the base field is weak. When the base field becomes coherent at high budget, the fixed layer ceases to dominate and becomes a lossy intermediate medium.

Pudding Theory Reading

Pudding Theory treats the LLM-mediated agent as an informational field system rather than a sequence of software modules. The base strategy set \(P_B\) is not merely a list of candidates. It is a sampled field over possible actions, with density increasing as compute is spent. The utility function \(J\) measures the alignment of that field with the optimum. Budget \(B\) strengthens the base signal by increasing coverage, reducing sampling error, and sharpening the distribution near high-utility regions.

A fixed LLM layer is therefore not an optimizer in the fundamental sense. It is a fixed transmitter placed inside an already forming signal field. At low budget, the base signal is weak and discontinuous. The LLM’s stored priors, world knowledge, and heuristic compression can dominate the field. This explains the finite-budget advantage in ranking and AIME selection. The derived strategy appears better because the fixed transmitter supplies structure absent from the sparse base set.

At high budget, the situation reverses. The base strategy set becomes the dominant signal. Majority vote, beam search, or increased signal-to-noise ratio gives the strategy field direct contact with the task optimum. The LLM can no longer act as the primary organizing field. Its intervention becomes a projection through its fixed representational geometry. What the source calls an information-susceptibility bound, Pudding Theory reads as the loss of Signal Dominance by the fixed LLM layer.

This reframes the performance gap. It is not background noise, prompt failure, or model-size insufficiency. It is the observable trace of a field-ordering transition. The dominant signal moves from the LLM prior to the base strategy distribution. Elaborate prompts can widen the gap because they make the fixed transmitter more active after its dominance has already been lost. Minimal prompts approach pass-through behavior because they disturb the base field less.

The free parameter \(\alpha\) is also reinterpreted. In the source, \(\alpha\) is a relative sensitivity estimated from performance curves. In Pudding Theory, \(\alpha\) is the measured dominance ratio between two informational fields: the fixed LLM field and the compute-strengthened base strategy field. The structural prediction is that \(\alpha\) must fall when the base field becomes coherent unless the LLM’s own field co-scales with the same budget channel.

Nested architectures change the system because the transmitter is no longer fixed. Generator and selector fields grow together. Co-scaling allows the LLM layer to remain phase-matched to the expanding strategy distribution. Positive inter-layer coupling is then not an exception to the bound. It is the condition under which Signal Dominance is preserved by the architecture rather than lost by an isolated module.

Falsifiable Observable

The distinguishing observable is the nested dominance ratio \(\alpha_{\mathrm{total}}\) measured in regimes where the fixed-selector curves obey \(\alpha \leq 1\) and where the empirical cross-partial coupling between generator and selector budgets is positive. Pudding Theory predicts that positive co-scaling should produce at least one large-budget regime with \(\alpha_{\mathrm{total}} > 1\), because the dominant signal is carried by the coupled architecture, not by either layer alone. If nested co-scaled susceptibility were measured to remain at or below one for all positive inter-layer coupling regimes, this Postulate would be falsified.

Editorial Dialogue

Tanaka: The reading risks renaming the source’s own theory. Song and Zhu already use susceptibility, data-processing arguments, and multi-variable response. Why invoke Signal Dominance when their framework explains the data using ordinary information theory?

Sterling: The ordinary account states the bound. It does not identify what changes ontologically across the crossover. In the source framing, the LLM is a module whose usefulness declines with budget. In the Pudding reading, the fixed LLM loses field dominance because the base strategy set becomes the stronger informational structure. That distinction matters. It predicts where prompt elaboration should hurt, why pass-through behavior is a meaningful regime, and why co-scaling is not merely more compute but restoration of dominance.

Tanaka: The word “field” is doing work that might be metaphorical. These are discrete algorithms and finite models.

Sterling: The source already treats \(J\) as a response surface and \(\nabla_B J\) as a susceptibility vector. Pudding Theory takes that geometry literally as the operational field of strategy formation. The field is not assumed continuous at the hardware level. It is the coarse-grained object defined by candidate distributions, budget channels, and utility gradients.

Tanaka: Then the risk is unfalsifiability.

Sterling: The falsifier is direct. Positive measured inter-layer coupling without any nested \(\alpha_{\mathrm{total}}>1\) would defeat the reading.

Discussion

The Pudding Theory reading buys a sharper account of architectural agency. A fixed LLM wrapper is not an enduring source of improvement. It is a temporary dominant signal in a sparse strategy field. Once compute makes the base distribution coherent, the wrapper becomes a filter. This explains why larger models alone do not remove the susceptibility gap in Tetris, why minimal prompting can approach the base algorithm, and why nested AIME curves matter more than fixed-selector comparisons.

The limitation is that Signal Dominance must be inferred through performance geometry. The paper does not measure internal representation alignment, entropy of candidate distributions, or mutual information with the optimum directly. Those measurements would strengthen or weaken the field reading. The most important open question is whether positive inter-layer coupling can be predicted before full scaling, using small-grid estimates of \(J(B_{\mathrm{gen}},B_{\mathrm{sel}})\). If so, Pudding Theory would convert the source’s empirical susceptibility framework into a design criterion: build architectures in which the dominant signal is preserved under growth.

References

1. Zhuo-Yang Song and Hua Xing Zhu. “A Theory of LLM Information Susceptibility.” arXiv:2603.23626, 2026. DOI: doi:10.48550/arxiv.2603.23626.

2. S. Ochs. “Pudding Theory: A Topological Theory of Information Fields.” QBist Lab Working Paper, 2026.

3. Claude E. Shannon. “A Mathematical Theory of Communication.” The Bell System Technical Journal 27, 379-423, 1948.

4. Thomas M. Cover and Joy A. Thomas. Elements of Information Theory. Wiley-Interscience, 2nd edition, 2006.

5. Ryogo Kubo. “Statistical-Mechanical Theory of Irreversible Processes. I.” Journal of the Physical Society of Japan 12, 570-586, 1957.

6. David H. Wolpert and William G. Macready. “No Free Lunch Theorems for Optimization.” IEEE Transactions on Evolutionary Computation 1, 67-82, 2002.

7. Xuezhi Wang et al. “Self-Consistency Improves Chain of Thought Reasoning in Language Models.” ICLR, 2023.

8. Shunyu Yao et al. “Tree of Thoughts: Deliberate Problem Solving with Large Language Models.” NeurIPS, 2023.