A unifying framework decomposes concept alignment into instance-wise and distributional translation and concept consistency, introduces the InterVenchA benchmark, and shows that joint optimization via CoSAE recovers strong alignment even with 0.1% paired data.
Revisiting the platonic representation hypothesis: An aristotelian view.arXiv preprint arXiv:2602.14486
11 Pith papers cite this work. Polarity classification is still indexing.
11
Pith papers citing it
abstract
The Platonic Representation Hypothesis suggests that representations from neural networks are converging to a common statistical model of reality. We show that the existing metrics used to measure representational similarity are confounded by network scale: increasing model depth or width can systematically inflate representational similarity scores. To correct these effects, we introduce a permutation-based null-calibration framework that transforms any representational similarity metric into a calibrated score with statistical guarantees. We revisit the Platonic Representation Hypothesis with our calibration framework, which reveals a nuanced picture: the apparent convergence reported by global spectral measures largely disappears after calibration, while local neighborhood similarity, but not local distances, retains significant agreement across different modalities. Based on these findings, we propose the Aristotelian Representation Hypothesis: representations in neural networks are converging to shared local neighborhood relationships.
citation-role summary
background 1
citation-polarity summary
roles
background 1polarities
background 1representative citing papers
Representational convergence across 16 LLMs on 800 reasoning problems is stronger for failed tasks and pre-decision stages but shows minimal causal influence on predictions, pointing to shared processing constraints over shared reasoning.
Foundation models yield less human-interpretable features than supervised vision transformers, with interpretability tied to activation locality and coarse semantic alignment rather than task performance.
Fusing embeddings from four Earth models (AlphaEarth, Tessera, GeoCLIP, SatCLIP) outperforms the best single model on four of six tasks, with gains depending on task and location.
Vision models converge on universal object dimensions that are semantically interpretable and align more closely with biological vision than model-specific ones.
Neighborhood re-ranking via Hungarian matching and query-conditioned local steering improve CLIP retrieval on attribute-binding and compositional tasks by addressing local geometric inconsistencies.
Representational alignment varies monotonically with SNR and non-monotonically with sample size (minimized near interpolation threshold) across linear and nonlinear networks, and is decoupled from generalization error.
Cross-modal representational alignment between text, image, audio, and video models degrades substantially when evaluated at million-sample scale and under realistic many-to-many pairing assumptions.
Using an asymmetric rank-based measure, the paper maps which layers of large language and vision models carry shared semantic information, finding central-layer peaks for translations and cross-modal pairs, and systematic predictability asymmetries favoring English and larger models.
Order is distinct from control, where control is defined as a local receiver-gated response law demonstrated across biological circuits and LLM response panels with reported prediction accuracies of 72-84%.
A mechanics of the learning process is emerging in deep learning theory, characterized by dynamics, coarse statistics, and falsifiable predictions across idealized settings, limits, laws, hyperparameters, and universal behaviors.
citing papers explorer
-
A Unifying Framework for Concept-Based Representational Similarity
A unifying framework decomposes concept alignment into instance-wise and distributional translation and concept consistency, introduces the InterVenchA benchmark, and shows that joint optimization via CoSAE recovers strong alignment even with 0.1% paired data.
-
Convergence Without Understanding: When Language Models Agree on Representations but Disagree on Reasoning
Representational convergence across 16 LLMs on 800 reasoning problems is stronger for failed tasks and pre-decision stages but shows minimal causal influence on predictions, pointing to shared processing constraints over shared reasoning.
-
Capability $\neq$ Interpretability: Human Interpretability of Vision Foundation Models
Foundation models yield less human-interpretable features than supervised vision transformers, with interpretability tied to activation locality and coarse semantic alignment rather than task performance.
-
Better Together: Evaluating the Complementarity of Earth Embedding Models
Fusing embeddings from four Earth models (AlphaEarth, Tessera, GeoCLIP, SatCLIP) outperforms the best single model on four of six tasks, with gains depending on task and location.
-
Characterizing Universal Object Representations Across Vision Models
Vision models converge on universal object dimensions that are semantically interpretable and align more closely with biological vision than model-specific ones.
-
Geometry-Aware CLIP Retrieval via Local Cross-Modal Alignment and Steering
Neighborhood re-ranking via Hungarian matching and query-conditioned local steering improve CLIP retrieval on attribute-binding and compositional tasks by addressing local geometric inconsistencies.
-
Signal-to-Noise Ratio and Sample Size Govern Representational Alignment in Neural Networks
Representational alignment varies monotonically with SNR and non-monotonically with sample size (minimized near interpolation threshold) across linear and nonlinear networks, and is decoupled from generalization error.
-
Back into Plato's Cave: Examining Cross-modal Representational Convergence at Scale
Cross-modal representational alignment between text, image, audio, and video models degrades substantially when evaluated at million-sample scale and under realistic many-to-many pairing assumptions.
-
A quantitative analysis of semantic information in deep representations of text and images
Using an asymmetric rank-based measure, the paper maps which layers of large language and vision models carry shared semantic information, finding central-layer peaks for translations and cross-modal pairs, and systematic predictability asymmetries favoring English and larger models.
-
Order Is Not Control
Order is distinct from control, where control is defined as a local receiver-gated response law demonstrated across biological circuits and LLM response panels with reported prediction accuracies of 72-84%.
-
There Will Be a Scientific Theory of Deep Learning
A mechanics of the learning process is emerging in deep learning theory, characterized by dynamics, coarse statistics, and falsifiable predictions across idealized settings, limits, laws, hyperparameters, and universal behaviors.