cs.AISep 15, 2026

What You Can't See Is Still What You Learn: A Preregistered Sixty-Society Confirmation That Evidence Masking Drives Compositional Generalization

Authors: Narcis Marincat

Abstract

Restricting what a module can read may improve what a system learns to compute. We test this in a preregistered confirmation with sixty four-cell systems sharing a frozen language-model backbone and communicating through learned continuous packets. Five conditions vary evidence masking, ownership markers, and replacement of foreign evidence with neutral filler, across six initialization clusters, each with two data orders, on one fresh task world. With markers available in both regimes, masking improved accuracy on held-out two- and three-operation compositions by median paired differences of 0.846 and 0.859; all twelve pairs cleared the required margins, and the full preregistered behavioral criterion passed. The unmarked replication also passed. No globally visible system passed the marker-following check, so the effect of usable role information remains unresolved. The filler condition yielded seven full generalizers, but its decomposition criteria were inconclusive. Packet interventions in all eighteen audited masked systems followed the predicted intermediate-value changes on eligible cases; these finite, success-conditioned audits do not establish mediation. The results confirm a large advantage of the tested masking regime, while leaving its finer attribution and generality open. Protocols, results, and checkpoints are public.

Explore similar work

CardsList
Aug 2, 2026cs.LG

An Identifiability Theory of Masked Prediction: Mode Blindness and Mask Schedules

Masked prediction learns by inferring missing variables from visible context. When does optimizing this conditional task recover the true joint data distribution? We study this question using an ε\varepsilon-identifiability modulus, which measures the worst-case joint-distribution error permitted by excess risk at most ε\varepsilon. For distributions with separated global modes, schedules retaining large visible contexts can permit substantial mode-weight errors at exponentially small excess risk. An exact information decomposition explains why: for a fixed mask, the loss penalizes only the mode-weight mismatch that remains unresolved by the visible context. For small mode-weight perturbations, the objective's sensitivity is proportional to residual mode uncertainty averaged over masks. Under joint masked-block log loss, low-visibility masks that retain mode uncertainty restore this sensitivity, while positive full-mask probability bounds joint-distribution error in terms of excess risk. We empirically validate these predictions through exact calculations and controlled stochastic optimization.
Yichao Cai, Javen Qinfeng Shi
IdentifiabilityHidden Markov ModelsLatent Bridge MatchingBlind Spots
AskPDF
May 29, 2026cs.LG

Assign and Add: A Mechanistic Study of Compositional Arithmetic

Large language models are able to compose skills in order to perform complex tasks, many of which might not have been seen during training. The details of how exactly this composition occurs remain elusive. In this paper, we study a mechanism for compositional generalization in transformers by considering a simple controlled setting involving variable assignment and modular addition. By partitioning our training data into disjoint sets, we observe that small transformers are able to generalize to previously unseen combinations of variables and numbers. Our mechanistic analysis shows that the same ``modular addition'' MLP module is used whether the inputs are given directly or indirectly through a separate variable assignment mechanism. We also analyze the training dynamics from an empirical lens, which reveals three phases of learning: first, modular addition is learned, then the structure required for variable assignment, and finally a refinement phase where the model generalizes to some hard sequences not seen in training. Finally, we provide a theoretical framework to explain how compositionality emerges from training dynamics. These results suggest that compositional generalization can be a natural consequence of the compositionality of internal mechanisms in~transformers.
Brady Exoo, Alberto Bietti, John Sous
Multilayer PerceptronsArithmetic
AskPDF
May 27, 2026cs.AI

The Confidence Shortcut: A Reasoning Failure Mode of Masked Diffusion Models

Masked diffusion language models (MDMs) uniquely support any-order generation, with confidence-based decoding currently serving as the de facto standard inference policy. To optimize for this, recent training schemes attempt to align training mask patterns directly with those observed during generation. However, we argue that confidence-based decoding is inherently misaligned with the logical-flow trajectories required for complex reasoning, and that confidence-aligned training actively entrenches this misalignment. We make this concrete using multi-digit addition, where the decoding strategy prematurely predicts locally easy digits before resolving their long-range dependencies, producing high-confidence errors on challenging inputs. While traditional random masking keeps the failure rate low on this challenging tail, confidence-aligned training amplifies the error rate by an order of magnitude. Across five distinct reasoning tasks, this same pattern emerges with task-dependent severity: confidence-based decoding induces failures on highly complex inputs, and confidence-aligned training exacerbates them. In contrast, random masking -- despite its perceived inefficiency -- robustly preserves the reasoning-trajectory conditionals essential for solving the challenging tail.
Dueun Kim, Albert No
Masked Diffusion Language ModelsTraining-Inference MismatchConfidence EstimationReasoning FailuresData GenerationRandom+1
AskPDF