cs.LGNov 11, 2026

Probabilistic State Space Enumeration in Neural Network Computation

Authors: Victor O. K. Li, Yang Han, Jacqueline C. K. Lam

Organizations: ElohaX Ltd. · The Hong Kong University of Science and Technology (Guangzhou), China · The Chinese University of Hong Kong, Hong Kong SAR, China · CASBS, Stanford University, United States

Abstract

Many methods for pruning neural networks rank weights or groups of weights by importance and retain a single subset. We study how channel groups contribute jointly to the output of multilayer perceptrons (MLPs) in Transformer language models. Inspired by network reliability analysis, probabilistic state space enumeration (PSSE) represents the dense MLP output as the expected importance-corrected masked output over 2G2^G joint states of GG channel groups. Summing the most probable states' probability-weighted outputs gives an approximation whose absolute error is bounded by the omitted probability mass times the maximum masked-output norm. NORDER-II adapts ORDER-II to enumerate states in decreasing probability without sorting the full space. Its first BB states maximize the probability mass captured at that budget. A weight-entropy scorer fixes each layer's ordering from pretrained weights. We evaluate all 168 MLP layers across six models from the Llama, Gemma, and Qwen families on two datasets with 8, 16, and 32 channel groups. In all 1,008 comparisons, the 95%-mass prefix yields lower mean relative error than uniform sampling at the same state count, with or without bias correction. At all three group counts, it also yields lower median and 99th-percentile relative error than a single top-kk mask. For downstream inference, we use the number of distinct channels in each prefix to set pruned MLP widths. With one validation-selected group count per model, PSSE achieves lower perplexity than both top-kk pruning baselines in all 36 matched-FLOP settings. Our pruned decoder MLPs achieve mean speedups of 1.11×1.11\times, 1.23×1.23\times, and 1.36×1.36\times over dense MLPs at 10%, 20%, and 30% target MLP projection FLOP savings, respectively. PSSE offers a principled path toward both probability-mass-controlled approximation and more efficient inference in pretrained language models.

Figures & tables

Appendix figures & tables12 assets

Supplementary material from the paper’s appendix.

Appendix

Explore similar work

Sep 12, 2026

How Should Reasoning Be Organized in a Transformer's Latent Space?

Continuous reasoning has emerged as a promising way to improve reasoning in large language models (LLMs). Yet we still lack a clear principle for deciding what a latent state should preserve. Reasoning by superposition shows that a single latent state can encode several search alternatives and expand them in parallel. We ask how those states should be weighted as reasoning proceeds. A natural choice is to preserve only the states active at the frontier step, since keeping every reached state appears to spread a limited hidden width too thin. We show that the opposite can hold. When later computation draws on several reached states, a cumulative state can guide attention correctly at a smaller hidden width than a frontier state that stores fewer states. At the same width, the cumulative state therefore keeps more intermediate states available for later reasoning. More generally, equal cumulative weights are optimal when future queries are unknown and remain close to the best task-specific weights when those queries are known. Experiments with two-layer and GPT-2 Transformers reproduce the predicted width advantage and show that unequal weights fail first on the states that receive the least weight. This suggests a important principle: keep reached states equally weighted, and restore equal weights as computation proceeds.
May 30, 2026

Refit the Probe: Single-Direction Ablation Is Not a Necessity Test

Probes are routinely paired with an intervention: ablate the direction the probe found, run the model, and read the change in task accuracy, taking a large drop as evidence that the computation depends on what the probe read and a near-zero drop as evidence that it does not. Either inference requires that the ablation have removed the target from the layer. We find that the ablation does not remove what it targets. A probe refitted on the ablated activations recovers its original accuracy in every cell we test, and keeps recovering when the probe's entire row space is deleted rather than a single axis, because the quantity survives in the orthogonal complement. Because a refitted probe recovers, neither a large task drop nor a near-zero one establishes whether the model needed the target, and one probe fit detects this. Replacing the ablation with iterative nullspace projection, scored against random subspaces of matched dimension, reverses the conclusion: representations that looked causally inert carry most of the task. The correction also separates where a variable is most readable from where deleting it does most damage, and those are not the same layer in any pretrained model we study. The erasure is defined by a linear probe family, so removing a nonlinearly encoded quantity remains open.
May 26, 2026

Tracing Computation Density in LLMs

Transformer-based large language models (LLMs) are comprised of billions of parameters arranged in deep and wide computational graphs, but it is not clear that they exploit their full capacity for all inputs. We introduce the s-Trace method to efficiently estimate a subgraph of size s that approximates a full model output. With this method, we find the computation in a variety of LLMs to be organized in two distinct phases. A small subgraph mostly composed of early-layer nodes can reconstruct the head of the full model output distribution. Adding further nodes, mostly located in later layers and increasingly consisting of attention heads, leads to incremental refinements in approximating the full output distribution. We find moreover that the amount of necessary computation per input correlates with model uncertainty, and that sparser subgraphs encode shallow statistics, such as unigram frequency. Overall, our results suggest a consistent modular organization in effective LLM computation, with a sparse early-layer core providing a rough prediction that is further refined through denser computations in later layers.