Overview
This research investigates the impact of slot-selective evidence masking on compositional generalization within multi-module neural systems, specifically focusing on societies of shared-genome language models. The core hypothesis tested is whether restricting each module to its own evidence span, rather than exposing all modules to the full input, alters the solutions discovered through gradient-based training.
Research Context
Multi-module neural systems typically allow every module access to the complete input. The study introduces an alternative regime: slot-selective evidence masking, which limits each module to process only its assigned evidence span. The investigation was conducted using a specific architecture: "four-cell societies." These societies operate with a shared, frozen pretrained language model and a single low-rank adapter. Communication between cells is strictly mediated through two continuous vectors of model width, which interact via a fixed relay.
Approach
The experimental setup involved a natural-language function-composition task that was prospectively sealed. Ten pairs of societies were trained for this task; each pair consisted of a "restricted" society and a "global" society. The pairs were matched identically, with the sole difference being the attention mask applied: restricted societies employed slot-selective evidence masking, while global societies maintained full visibility of the input for all modules. This controlled comparison aimed to isolate the effect of the masking regime.
Findings
- Performance Advantage: Restricted-visibility societies consistently outperformed their globally visible twins. In 9 out of 10 pairs, restricted societies showed an advantage of at least 20 percentage points at both depths of the compositional task. The median paired advantages observed were 0.7648 and 0.6050.
- Communication Dependency: When communication within the societies was cut, every restricted society's performance was reduced to chance levels, indicating that inter-module communication is critical for their function.
- Generalization to Unseen Compositions: A post hoc collision-stratified analysis revealed that for programs whose complete affine map never appeared in the training data, the depth-three advantage for restricted societies remained 0.558. This suggests an ability to generalize to novel compositions.
- Relay State Interpretability: In six post hoc-selected restricted societies, packet interventions on correctly answered held-out episodes were found to be consistent with approximately value-indexed relay states. This implies a structured and interpretable communication mechanism within these societies.
- Global Model Communication: The sole high-performing global model also required communication to function. However, its same-value packets were not interchangeable across episodes, distinguishing its communication dynamics from those observed in restricted societies.
- Mask-Native Solutions: A post hoc mask crossover experiment indicated that both the restricted and global arms developed "mask-native" solutions. This suggests that the masking regime significantly influenced the types of solutions discovered during training.
- Role of Restricted Visibility: The study notes that restricted visibility is not inherently necessary for composition. The restricted mask simultaneously blocks foreign evidence and implicitly identifies each cell's assigned slot. Therefore, a definitive attribution solely to evidence visibility requires a role-marked control experiment.
- Preregistration Outcome: The preregistered battery formally failed because the median depth-three accuracy for the restricted arm was 0.6988, which fell below the predefined 0.70 floor. An earlier qualification cohort resulted in 0 out of 10 complete passes.
Why This Matters
The findings indicate that imposing structural constraints like slot-selective evidence masking can significantly influence the learning outcomes in multi-module neural systems, leading to improved compositional generalization. This suggests that the architectural design, particularly concerning inter-module visibility, plays a crucial role in shaping the capabilities of language models for complex tasks.
Key Limitations Mentioned by Researchers
The study explicitly mentions that the preregistered battery formally failed due to the restricted-arm median depth-three accuracy (0.6988) falling below the 0.70 floor. Furthermore, the researchers note that while the restricted mask blocks foreign evidence and implicitly identifies each cell's assigned slot, attributing the observed benefits solely to evidence visibility requires a future role-marked control experiment.