Taming Polysemanticity in LLMs: Provable Feature Recovery via Sparse Autoencoders

• URL: http://arxiv.org/abs/2506.14002v1
• Date: Mon, 16 Jun 2025 20:58:05 GMT
• Title: Taming Polysemanticity in LLMs: Provable Feature Recovery via Sparse Autoencoders
• Authors: Siyu Chen, Heejune Sheen, Xuyuan Xiong, Tianhao Wang, Zhuoran Yang,
• Abstract summary: Existing SAE training algorithms often lack rigorous mathematical guarantees and suffer from practical limitations.<n>We first propose a novel statistical framework for the feature recovery problem, which includes a new notion of feature identifiability.<n>We introduce a new SAE training algorithm based on bias adaptation'', a technique that adaptively adjusts neural network bias parameters to ensure appropriate activation sparsity.
• Score: 50.52694757593443
• License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
• Abstract: We study the challenge of achieving theoretically grounded feature recovery using Sparse Autoencoders (SAEs) for the interpretation of Large Language Models. Existing SAE training algorithms often lack rigorous mathematical guarantees and suffer from practical limitations such as hyperparameter sensitivity and instability. To address these issues, we first propose a novel statistical framework for the feature recovery problem, which includes a new notion of feature identifiability by modeling polysemantic features as sparse mixtures of underlying monosemantic concepts. Building on this framework, we introduce a new SAE training algorithm based on ``bias adaptation'', a technique that adaptively adjusts neural network bias parameters to ensure appropriate activation sparsity. We theoretically \highlight{prove that this algorithm correctly recovers all monosemantic features} when input data is sampled from our proposed statistical model. Furthermore, we develop an improved empirical variant, Group Bias Adaptation (GBA), and \highlight{demonstrate its superior performance against benchmark methods when applied to LLMs with up to 1.5 billion parameters}. This work represents a foundational step in demystifying SAE training by providing the first SAE algorithm with theoretical recovery guarantees, thereby advancing the development of more transparent and trustworthy AI systems through enhanced mechanistic interpretability.

Related papers

• Reparameterized LLM Training via Orthogonal Equivalence Transformation [54.80172809738605]
  We present POET, a novel training algorithm that uses Orthogonal Equivalence Transformation to optimize neurons.<n>POET can stably optimize the objective function with improved generalization.<n>We develop efficient approximations that make POET flexible and scalable for training large-scale neural networks.
  arXiv  Detail & Related papers  (2025-06-09T17:59:34Z)
• Supervised Optimism Correction: Be Confident When LLMs Are Sure [91.7459076316849]
  We establish a novel theoretical connection between supervised fine-tuning and offline reinforcement learning.<n>We show that the widely used beam search method suffers from unacceptable over-optimism.<n>We propose Supervised Optimism Correction, which introduces a simple yet effective auxiliary loss for token-level $Q$-value estimations.
  arXiv  Detail & Related papers  (2025-04-10T07:50:03Z)
• Theoretical Insights in Model Inversion Robustness and Conditional Entropy Maximization for Collaborative Inference Systems [89.35169042718739]
  collaborative inference enables end users to leverage powerful deep learning models without exposure of sensitive raw data to cloud servers.<n>Recent studies have revealed that these intermediate features may not sufficiently preserve privacy, as information can be leaked and raw data can be reconstructed via model inversion attacks (MIAs)<n>This work first theoretically proves that the conditional entropy of inputs given intermediate features provides a guaranteed lower bound on the reconstruction mean square error (MSE) under any MIA.<n>Then, we derive a differentiable and solvable measure for bounding this conditional entropy based on the Gaussian mixture estimation and propose a conditional entropy algorithm to enhance the inversion robustness
  arXiv  Detail & Related papers  (2025-03-01T07:15:21Z)
• Sample Complexity of Linear Quadratic Regulator Without Initial Stability [11.98212766542468]
  Inspired by REINFORCE, we introduce a novel receding-horizon algorithm for the Linear Quadratic Regulator (LQR) problem with unknown parameters.<n>Unlike prior methods, our algorithm avoids reliance on two-point gradient estimates while maintaining the same order of sample complexity.
  arXiv  Detail & Related papers  (2025-02-20T02:44:25Z)
• Assessing Correctness in LLM-Based Code Generation via Uncertainty Estimation [0.0]
  We explore uncertainty estimation as a proxy for correctness in LLM-generated code.<n>We adapt two state-of-the-art techniques from natural language generation to the domain of code generation.<n>Our findings indicate a strong correlation between the uncertainty computed through these techniques and correctness.
  arXiv  Detail & Related papers  (2025-02-17T10:03:01Z)
• Preconditioned Inexact Stochastic ADMM for Deep Model [35.37705488695026]
  This paper develops an algorithm, PISA, which enables scalable parallel computing and supports various second-moment schemes.<n>Grounded in rigorous theoretical guarantees, the algorithm converges under the sole assumption of Lipschitz of the gradient.<n> Comprehensive experimental evaluations for or fine-tuning diverse FMs, including vision models, large language models, reinforcement learning models, generative adversarial networks, and recurrent neural networks, demonstrate its superior numerical performance compared to various state-of-the-art Directions.
  arXiv  Detail & Related papers  (2025-02-15T12:28:51Z)
• Refining Salience-Aware Sparse Fine-Tuning Strategies for Language Models [14.68920095399595]
  sparsity-based PEFT (SPEFT) introduces trainable sparse adaptations to the weight matrices in the model.<n>We conduct the first systematic evaluation of salience metrics for SPEFT, inspired by zero-cost NAS proxies.<n>We compare static and dynamic masking strategies, finding that static masking, which predetermines non-zero entries before training, delivers efficiency without sacrificing performance.
  arXiv  Detail & Related papers  (2024-12-18T04:14:35Z)
• Uncertainty Aware Learning for Language Model Alignment [97.36361196793929]
  We propose uncertainty-aware learning (UAL) to improve the model alignment of different task scenarios. We implement UAL in a simple fashion -- adaptively setting the label smoothing value of training according to the uncertainty of individual samples. Experiments on widely used benchmarks demonstrate that our UAL significantly and consistently outperforms standard supervised fine-tuning.
  arXiv  Detail & Related papers  (2024-06-07T11:37:45Z)
• Towards Stable Machine Learning Model Retraining via Slowly Varying Sequences [6.067007470552307]
  We propose a model-agnostic framework for finding sequences of models that are stable across retraining iterations.<n>We develop a mixed-integer optimization formulation that is guaranteed to recover optimal models.<n>We find that, on average, a 2% reduction in predictive power leads to a 30% improvement in stability.
  arXiv  Detail & Related papers  (2024-03-28T22:45:38Z)
• How Many Pretraining Tasks Are Needed for In-Context Learning of Linear Regression? [92.90857135952231]
  Transformers pretrained on diverse tasks exhibit remarkable in-context learning (ICL) capabilities. We study ICL in one of its simplest setups: pretraining a linearly parameterized single-layer linear attention model for linear regression.
  arXiv  Detail & Related papers  (2023-10-12T15:01:43Z)
• Adaptive Sampling for Minimax Fair Classification [40.936345085421955]
  We propose an adaptive sampling algorithm based on the principle of optimism, and derive theoretical bounds on its performance. By deriving algorithm independent lower-bounds for a specific class of problems, we show that the performance achieved by our adaptive scheme cannot be improved in general.
  arXiv  Detail & Related papers  (2021-03-01T04:58:27Z)
• Reparameterized Variational Divergence Minimization for Stable Imitation [57.06909373038396]
  We study the extent to which variations in the choice of probabilistic divergence may yield more performant ILO algorithms. We contribute a re parameterization trick for adversarial imitation learning to alleviate the challenges of the promising $f$-divergence minimization framework. Empirically, we demonstrate that our design choices allow for ILO algorithms that outperform baseline approaches and more closely match expert performance in low-dimensional continuous-control tasks.
  arXiv  Detail & Related papers  (2020-06-18T19:04:09Z)
• Joint Stochastic Approximation and Its Application to Learning Discrete Latent Variable Models [19.07718284287928]
  We show that the difficulty of obtaining reliable gradients for the inference model and the drawback of indirectly optimizing the target log-likelihood can be gracefully addressed. We propose to directly maximize the target log-likelihood and simultaneously minimize the inclusive divergence between the posterior and the inference model. The resulting learning algorithm is called joint SA (JSA)
  arXiv  Detail & Related papers  (2020-05-28T13:50:08Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.