Proves that Rademacher complexity of depth-d compositional trees over finite operator vocabulary is controlled by (K b L)^{d} / sqrt(n) under Lipschitz conditions on operators.
Title resolution pending
9 Pith papers cite this work, alongside 6,006 external citations. Polarity classification is still indexing.
9
Pith papers citing it
6,006
external citations · OpenAlex
citation-role summary
background 2
citation-polarity summary
years
2026 9verdicts
UNVERDICTED 9roles
background 2polarities
background 2representative citing papers
Approximate MDL with additive slack C guarantees finite expected squared prediction error for λ ≥ 1 via telescoping and stopping-time arguments, but fails for λ < 1 or multiplicative approximations.
QuBD extends algorithmic complexity estimation to quantized DNN weights, revealing that complexity decreases during learning, increases with overfitting, follows grokking patterns, and correlates with generalization.
Exhaustive symbolic regression identifies low-complexity functional forms for luminosity and mass functions that outperform Schechter and Press-Schechter parametrizations while satisfying physical extrapolation and integration constraints.
Authors create a benchmark across discrete/continuous and static/dynamical systems and introduce the Causal Abstraction Error (CAE) metric that reliably distinguishes valid from invalid causal abstractions when it includes faithfulness testing.
Derives an upper bound on frozen LM expected risk from proxy risk, SAE reconstruction gap, concept-pool mismatch and sparse complexity, with non-vacuous bounds observed on GPT-2, Gemma-2B and Llama-3-8B.
FlexLAM trains variable-length latent actions with nested dropout to match or exceed fixed-capacity LAMs at every token budget under scarce-label supervision without new architectures or losses.
Post-selection with DL or FBF after multi-objective GP search improves test-set performance over AIC/BIC baselines on noisy synthetic and real regression tasks, while using DL directly as fitness often causes premature convergence to overly simple models.
MDL and BIC most reliably select low test-error models and recover ground-truth expressions in symbolic regression benchmarks.
citing papers explorer
-
Sample Complexity of Scientific Discovery: PAC Learnability of Compositional Function Trees
Proves that Rademacher complexity of depth-d compositional trees over finite operator vocabulary is controlled by (K b L)^{d} / sqrt(n) under Lipschitz conditions on operators.
-
Prediction Under Imperfect Compression: A Theory of Approximate MDL
Approximate MDL with additive slack C guarantees finite expected squared prediction error for λ ≥ 1 via telescoping and stopping-time arguments, but fails for λ < 1 or multiplicative approximations.
-
Characterizing Learning in Deep Neural Networks using Tractable Algorithmic Complexity Analysis
QuBD extends algorithmic complexity estimation to quantized DNN weights, revealing that complexity decreases during learning, increases with overfitting, follows grokking patterns, and correlates with generalization.
-
The functional form of galaxy and halo luminosity and mass functions
Exhaustive symbolic regression identifies low-complexity functional forms for luminosity and mass functions that outperform Schechter and Press-Schechter parametrizations while satisfying physical extrapolation and integration constraints.
-
Validating Causal Abstraction Metrics on Simulated Complex Systems
Authors create a benchmark across discrete/continuous and static/dynamical systems and introduce the Causal Abstraction Error (CAE) metric that reliably distinguishes valid from invalid causal abstractions when it includes faithfulness testing.
-
From Sparse Features to Trustworthy Proxies: Certifying SAE-Based Interpretability
Derives an upper bound on frozen LM expected risk from proxy risk, SAE reconstruction gap, concept-pool mismatch and sparse complexity, with non-vacuous bounds observed on GPT-2, Gemma-2B and Llama-3-8B.
-
FlexLAM: Resolving the Bottleneck Trade-off in Latent Action Learning
FlexLAM trains variable-length latent actions with nested dropout to match or exceed fixed-capacity LAMs at every token budget under scarce-label supervision without new architectures or losses.
-
Guiding Multi-Objective Genetic Programming with Description Length Improves Symbolic Regression Solutions
Post-selection with DL or FBF after multi-objective GP search improves test-set performance over AIC/BIC baselines on noisy synthetic and real regression tasks, while using DL directly as fitness often causes premature convergence to overly simple models.
-
A Comparative Study of Model Selection Criteria for Symbolic Regression
MDL and BIC most reliably select low test-error models and recover ground-truth expressions in symbolic regression benchmarks.