6arXiv cs.LG (Machine Learning)·
Data-driven theory explains RoPE frequency usage and long-context generalization in transformers
A new arXiv preprint proposes a data-centered explanation for why trained transformers use RoPE positional frequencies non-uniformly: frequencies are selected to match the relative-distance dependency structure of training data, with optimal frequency scaling as 1/W for dependency width W. The paper formalizes a field-resolution tradeoff and connects this frequency-matching principle to position-interpolation-based length generalization, showing that test-time frequency scaling succeeds when longer-context dependencies are approximate dilations of training-time dependencies. Empirical results demonstrate that natural language exhibits approximate self-similarity across positional scales, providing a mechanistic account of why context-length extrapolation methods work when they do.
Related guides (2)
Related events (8)
You Could Have Designed State of the Art Positional Encoding
A Hugging Face blog post walks through the design space of positional encoding for transformer models, building intuition for why modern schemes like RoPE emerged. The post takes a pedagogical approach, showing how one could derive state-of-the-art positional encoding from first principles. It covers the evolution from absolute to relative positional encodings and the properties that make certain schemes preferable for long-context generalization.
Roofline-inspired scaling model predicts Transformer fine-tuning energy consumption across GPU configurations
A new arXiv preprint presents a framework for modeling energy consumption during Transformer training on multiple GPUs, using BERT architectural sweeps to relate measured energy to proxies for compute, memory traffic, and hardware efficiency. The approach adapts roofline modeling with a speedup-based hardware-efficiency factor that accounts for tensor parallelism and fully sharded data parallelism. The resulting scaling law accurately predicts training energy across heterogeneous configurations, targeting sustainable and cost-aware system design.
Positional vs. Symbolic Attention Heads: Learning Dynamics, RoPE Geometry, and Length Generalization
Researchers train a decoder-only Transformer (GPT-J) on two structurally equivalent multi-hop reasoning tasks to study how attention heads specialize into positional or symbolic roles during learning. They find that successful task learning correlates with the emergence of 'pure' heads—exclusively positional or symbolic—and provide theoretical constructions showing how single-layer RoPE-based attention realizes these functions geometrically. A novel 'discrepancy' metric formalizes the robustness difference between the two head types, with symbolic mechanisms shown to extrapolate more reliably to longer sequences than positional ones. The findings have implications for understanding length generalization failures in RoPE-based models.
Hyperfitting Explained: Terminal Geometric Expansion in Final Transformer Layers Drives Diversity Gains
This paper investigates the 'hyperfitting' phenomenon—where fine-tuning LLMs to near-zero loss on small datasets improves open-ended generation and reduces repetition—and demonstrates it is mechanistically distinct from temperature scaling. Entropy-matched control experiments falsify both the temperature-equivalence and static vocabulary reweighting hypotheses, instead localizing the effect to a 'Terminal Expansion' in the final transformer block where feature-space dimensionality expands by ~80.8 dimensions, enabling promotion of deep-tail tokens via context-dependent rank reordering. The authors introduce Late-Stage LoRA, a targeted fine-tuning strategy updating only the final 5 layers, achieving robust generation with minimal parameter updates.
Conditional Scale Entropy: A Wavelet-Derived Tool for Mechanistic Interpretability of Metaphor Processing in Transformers
This paper introduces Conditional Scale Entropy (CSE), a wavelet-derived measure of how transformer computation engages across frequency scales at each layer, and applies it to study metaphor processing in decoder-only language models. The authors prove CSE is invariant to update magnitude, isolating structural computation patterns from intensity. Across architectures ranging from GPT-2 (124M) to LLaMA-2 7B and GPT-oss 20B, metaphorical tokens consistently produce higher spectral breadth than literal tokens in early-to-mid layers, with the effect surviving permutation correction and specificity controls. The work establishes multi-scale coordination as a consistent mechanistic signature of metaphorical language processing and positions CSE as a general interpretability tool for cross-depth structure in transformers.
Energy-based transformers as unified predictors of reading difficulty in computational psycholinguistics
A new arXiv preprint introduces energy-based transformer measures as predictors of human reading difficulty, evaluated across three reading-time corpora (Natural Stories, UCL eye-tracking, UCL self-paced reading). The energy measure outperforms surprisal alone and appears to subsume both surprisal and attention entropy effects, suggesting it could serve as a single unified predictor. The work connects transformer language models to Hopfield networks and dense associative memory literature, marking the first application of energy-based transformer measures in computational psycholinguistics.
SemRF: Semantic Reference Frames for analyzing residual-stream dynamics in language models
A new arXiv preprint introduces Semantic Reference Frames (SemRF), a formal framework for analyzing how language model computation evolves across transformer layers via the residual stream. The method uses anchor-based coordinates with pseudo-inverse tying to separate semantic measurement from residual dynamics, enabling stable cross-layer comparison without measurement drift. The framework defines layerwise semantic trajectories, Voronoi-based coarse cell assignments, and a minimum-action canonical trace, with theoretical links to parameter efficiency and knowledge density. This is a mechanistic interpretability contribution aimed at understanding internal model computation.
Rosetta Neurons follow sublinear power-law scaling with model size, becoming more monosemantic at scale
A new arXiv paper investigates how neuron populations evolve with scale in both language models (up to 30B parameters) and vision models (up to 5B parameters), focusing on 'Rosetta Neurons' — neurons with similar activation patterns across independently trained models. The authors find Rosetta Neurons grow in absolute count but shrink as a fraction of total neurons, and exhibit a 'Neuron Polarization Effect' where they become increasingly monosemantic while non-Rosetta neurons remain less selective. An analytical model explains the sublinear power-law scaling, and the paper demonstrates practical utility via a targeted data-filtering case study for continued pretraining. The results extend scaling laws to neuron-level interpretability structure, linking model size to systematic changes in universality and specialization.