Progressive Localisation in Localist LLMs
📝 Abstract
This paper demonstrates that progressive localization, the gradual increase of attention locality from early distributed layers to late localized layers, represents the optimal architecture for creating interpretable large language models (LLMs) while preserving performance. Through systematic experimentation with GPT-2 fine-tuned on The Psychology of Artificial Superintelligence, we evaluate five locality configurations: two uniform baselines (fully distributed and fully localist) and three progressive polynomial schedules. We investigate whether interpretability constraints can be aligned with natural semantic structure while being applied strategically across network depth. We demonstrate that progressive semantic localization, combining adaptive semantic block partitioning with steep polynomial locality schedules, achieves near-baseline language modeling performance while providing interpretable attention patterns. Multiple independent training runs with different random seeds establish that results are statistically robust and highly reproducible. The approach dramatically outperforms both fixed-window localization and naive uniform locality constraints. Analysis reveals that maintaining flexibility through low-fidelity constraints preserves model capacity while providing interpretability benefits, and that steep schedules concentrating locality in decision-critical final layers while preserving distributed learning in early layers achieve near-baseline attention distribution characteristics. These findings demonstrate that interpretability mechanisms should align with semantic structure to achieve practical performance-interpretability tradeoffs for trustworthy AI systems.
💡 Analysis
This paper demonstrates that progressive localization, the gradual increase of attention locality from early distributed layers to late localized layers, represents the optimal architecture for creating interpretable large language models (LLMs) while preserving performance. Through systematic experimentation with GPT-2 fine-tuned on The Psychology of Artificial Superintelligence, we evaluate five locality configurations: two uniform baselines (fully distributed and fully localist) and three progressive polynomial schedules. We investigate whether interpretability constraints can be aligned with natural semantic structure while being applied strategically across network depth. We demonstrate that progressive semantic localization, combining adaptive semantic block partitioning with steep polynomial locality schedules, achieves near-baseline language modeling performance while providing interpretable attention patterns. Multiple independent training runs with different random seeds establish that results are statistically robust and highly reproducible. The approach dramatically outperforms both fixed-window localization and naive uniform locality constraints. Analysis reveals that maintaining flexibility through low-fidelity constraints preserves model capacity while providing interpretability benefits, and that steep schedules concentrating locality in decision-critical final layers while preserving distributed learning in early layers achieve near-baseline attention distribution characteristics. These findings demonstrate that interpretability mechanisms should align with semantic structure to achieve practical performance-interpretability tradeoffs for trustworthy AI systems.
📄 Content
The interpretability of large language models has become critical for AI safety, regulatory compliance, and trustworthy deployment in high-stakes domains.
However, the dominant transformer architecture operates through distributed representations where information about any concept spreads across millions of parameters (Vaswani et al., 2017). This distribution makes it fundamentally difficult to isolate which model components encode specific facts, execute particular reasoning steps, or contribute to individual decisions. While distributed representations offer computational advantages (Hinton et al., 1986), they create opacity that limits human oversight and verification.
Localist language models address this challenge by constraining attention patterns to focus on semantically coherent blocks of tokens, enabling clearer mapping between attention and reasoning (Diederich, 2025a). Initial implementations demonstrated feasibility but incurred substantial performance costs. Early experiments using fixed positional windows (uniform 5-token blocks) showed degradation. Perplexity increased approximately 6.6-fold relative to distributed baselines, rendering the approach less than optimal for any production application. The early experiments revealed a fundamental insight: interpretability constraints must align with linguistic structure, not just impose arbitrary boundaries.
This paper demonstrates that combining progressive locality schedules with semantic clustering eliminates this performance gap. Through multi-seed experimentation (n=5 per configuration), we establish that progressive quintic localization with adaptive semantic blocks achieves 7.87±0.65 perplexity, only 4.0% above the distributed baseline of 7.57±0.67, with strong statistical significance (p<0.001) and reproducibility (CV=8.2%). This represents a 93% reduction in performance cost compared to fixed-window approaches (from 6.6× to 1.040× gap) and an 82% reduction compared to naive uniform localization (9.27±0.58 PPL,22.4% gap).
The key innovations are threefold. First, semantic block partitioning adapts locality constraints to natural language structure rather than imposing fixed positional windows. Second, progressive locality schedules delay localization to later layers where decisions occur, allowing early layers to learn distributed representations essential for feature extraction. Third, quintic schedule functions (β=5) provide optimal performance by concentrating localization pressure in final layers while maintaining near-zero constraints in early layers.
Crucially, the low fidelity scores observed in the best-performing configuration (0.194 vs 0.509 for uniform localization) reveal that semantic blocks function as guidance rather than rigid constraints. This flexibility allows the model to adapt attention patterns to contextual requirements while maintaining interpretable structure: the blocks shape but do not dictate information flow. This finding suggests that effective interpretability mechanisms need not eliminate model flexibility but rather provide structured pathways that preserve adaptive capacity.
Localist models represent concepts through dedicated, interpretable units rather than distributed patterns across many neurons (Diederich et al., 2010). In the context of decoder-only language models like GPT-2, this means constraining self-attention mechanisms to operate within semantically meaningful token groups, what we term “semantic blocks.” Rather than allowing each token to attend to all previous tokens (as in standard autoregressive transformers), localist LLMs encourage attention to concentrate within and between well-defined semantic units such as entities, phrases, clauses, or arguments.
The Localist LLM framework (Diederich, 2025a) extends transformer attention through a locality penalty that encourages block-structured patterns. For a sequence partitioned into semantic blocks B = {B₁, B₂, …, B_K}, the training objective becomes:
where L_LM is the standard language modeling loss and L_locality measures attention spread across block boundaries. The locality penalty can be formulated as:
where A^(ℓ,h)_ij represents attention weight from token i to token j in layer ℓ and head h, and d(•,•) measures the distance between blocks (0 for within-block attention, increasing for cross-block attention). The hyperparameter λ controls the strength of locality enforcement: λ=0 recovers standard distributed transformers, while λ→∞ enforces strict block-local attention.
Rather than applying uniform locality strength across all layers of the decoderonly architecture, progressive localization varies λ as a function of network depth:
where ℓ indexes the layer (0 to L-1), λ_max is the maximum locality strength, and
Effective locality constraints require meaningful semantic units. Fixed-size windows (e.g., every 5 tokens) ignore linguistic structure and force arbitrary boundaries. This work employs semantic clustering that analyz
This content is AI-processed based on ArXiv data.