arXiv subjects

Tung-Yu Wu

Publications and source records attributed to Tung-Yu Wu.

10 recordsLinked to original sources

arXiv2026

CLT-Forge: A Scalable Library for Cross-Layer Transcoders and Attribution Graphs

Mechanistic interpretability seeks to understand how Large Language Models (LLMs) represent and process information. Recent approaches based on dictionary learning and transcoders enable representing model computation in terms of sparse, interpretable features and their interactions, giving rise to feature attribution graphs. However, these graphs are often large and redundant, limiting their interpretability in practice. Cross-Layer Transcoders (CLTs) address this issue by sharing features across layers while preserving layer-specific decoding, yielding more compact representations, but remain difficult to train and analyze at scale. We introduce an open-source library for end-to-end training and interpretability of CLTs. Our framework integrates scalable distributed training with model sharding and compressed activation caching, a unified automated interpretability pipeline for feature analysis and explanation, attribution graph computation using Circuit-Tracer, and a flexible visualization interface. This provides a practical and unified solution for scaling CLT-based mechanistic interpretability. Our code is available at: https://github.com/LLM-Interp/CLT-Forge.

cs.LG↗

arXiv2026

Position: Token Taxes Can Mitigate AI's Economic Risks

AI-driven automation threatens to erode government tax bases, lower living standards, and disempower citizens--risks that mirror the 40-year stagnation of wages during the first industrial revolution. While AI safety research has focused primarily on capability risks, comparatively little work has studied how to mitigate the economic risks of AI. This position paper argues that technical governance researchers should prioritize the study of token taxes: usage-based surcharges on model inference applied at the point of sale. We situate token taxes within previous proposals for robot taxes and identify two key advantages: they are enforceable through existing compute governance infrastructure, and they capture value where AI is used rather than where models are hosted. We then present a research roadmap. For enforcement, we outline a staged audit pipeline--black-box token verification, norm-based tax rates, and white-box audits--and identify open technical problems at each stage. For impact, we highlight the need for economic modeling of cost pass-through and deadweight loss. Finally, we discuss why FLOP taxes may be preferable, token taxes could stifle innovation, and how to prevent AI superpowers from vetoing such measures.

cs.CY↗

arXiv2026

Query Circuits: Explaining How Language Models Answer User Prompts

Explaining why a language model produces a particular output requires local, input-level explanations. Existing methods uncover global capability circuits (e.g., indirect object identification), but not why the model answers a specific input query in a particular way. We introduce query circuits, which directly trace the information flow inside a model that maps a specific input to the output. Unlike surrogate-based approaches (e.g., sparse autoencoders), query circuits are identified within the model itself, resulting in more faithful and computationally accessible explanations. To make query circuits practical, we address two challenges. First, we introduce Normalized Deviation Faithfulness (NDF), a robust metric to evaluate how well a discovered circuit recovers the model's decision for a specific input, and is broadly applicable to circuit discovery beyond our setting. Second, we develop sampling-based methods to efficiently identify circuits that are sparse yet faithfully describe the model's behavior. Across benchmarks (IOI, arithmetic, MMLU, and ARC), we find that there exist extremely sparse query circuits within the model that can recover much of its performance on single queries. For example, a circuit covering only 1.3% of model connections can recover about 60% of performance on an MMLU questions. Overall, query circuits provide a step towards faithful, scalable explanations of how language models process individual inputs. The project page is at https://tony10101105.github.io/query-circuit/.

cs.AI↗

arXiv2025

Data-Efficient 3D Visual Grounding via Order-Aware Referring

3D visual grounding aims to identify the target object within a 3D point cloud scene referred to by a natural language description. Previous works usually require significant data relating to point color and their descriptions to exploit the corresponding complicated verbo-visual relations. In our work, we introduce Vigor, a novel Data-Efficient 3D Visual Grounding framework via Order-aware Referring. Vigor leverages LLM to produce a desirable referential order from the input description for 3D visual grounding. With the proposed stacked object-referring blocks, the predicted anchor objects in the above order allow one to locate the target object progressively without supervision on the identities of anchor objects or exact relations between anchor/target objects. In addition, we present an order-aware warm-up training strategy, which augments referential orders for pre-training the visual grounding framework. This allows us to better capture the complex verbo-visual relations and benefit the desirable data-efficient learning scheme. Experimental results on the NR3D and ScanRefer datasets demonstrate our superiority in low-resource scenarios. In particular, Vigor surpasses current state-of-the-art frameworks by 9.3% and 7.6% grounding accuracy under 1% data and 10% data settings on the NR3D dataset, respectively. Our code is publicly available at https://github.com/tony10101105/Vigor.

cs.CV↗

arXiv2025

U-shaped and Inverted-U Scaling behind Emergent Abilities of Large Language Models

Large language models (LLMs) have been shown to exhibit emergent abilities in some downstream tasks, where model performance stagnates at first and then improves sharply and unpredictably with scale beyond a threshold. In this work, we investigate the phenomenon by grouping questions based on difficulty level and provide a possible explanation for emergent abilities. Specifically, we observe U-shaped scaling for hard questions and inverted-U scaling followed by steady improvement for easy questions. The two scaling patterns initially offset each other, causing stagnant overall performance. The performance starts to soar when the scaling pattern of easy questions reverts from inverse to standard scaling, leading to emergent abilities. Based on this finding, we propose a simple yet effective pipeline, called Slice-and-Sandwich, to predict the emergence threshold and model performance beyond the threshold. Our code is publicly available at https://github.com/tony10101105/ExpEmergence.

cs.AI↗

arXiv2024

AND: Audio Network Dissection for Interpreting Deep Acoustic Models

Neuron-level interpretations aim to explain network behaviors and properties by investigating neurons responsive to specific perceptual or structural input patterns. Although there is emerging work in the vision and language domains, none is explored for acoustic models. To bridge the gap, we introduce $\textit{AND}$, the first $\textbf{A}$udio $\textbf{N}$etwork $\textbf{D}$issection framework that automatically establishes natural language explanations of acoustic neurons based on highly-responsive audio. $\textit{AND}$ features the use of LLMs to summarize mutual acoustic features and identities among audio. Extensive experiments are conducted to verify $\textit{AND}$'s precise and informative descriptions. In addition, we demonstrate a potential use of $\textit{AND}$ for audio machine unlearning by conducting concept-specific pruning based on the generated descriptions. Finally, we highlight two acoustic model behaviors with analysis by $\textit{AND}$: (i) models discriminate audio with a combination of basic acoustic features rather than high-level abstract concepts; (ii) training strategies affect model behaviors and neuron interpretability -- supervised training guides neurons to gradually narrow their attention, while self-supervised learning encourages neurons to be polysemantic for exploring high-level features.

cs.SD↗

arXiv2023

Model Extraction Attack against Self-supervised Speech Models

Self-supervised learning (SSL) speech models generate meaningful representations of given clips and achieve incredible performance across various downstream tasks. Model extraction attack (MEA) often refers to an adversary stealing the functionality of the victim model with only query access. In this work, we study the MEA problem against SSL speech model with a small number of queries. We propose a two-stage framework to extract the model. In the first stage, SSL is conducted on the large-scale unlabeled corpus to pre-train a small speech model. Secondly, we actively sample a small portion of clips from the unlabeled corpus and query the target model with these clips to acquire their representations as labels for the small model's second-stage training. Experiment results show that our sampling methods can effectively extract the target model without knowing any information about its model architecture.

cs.SD↗

arXiv2023

The Efficacy of Self-Supervised Speech Models for Audio Representations

Self-supervised learning (SSL) speech models, which can serve as powerful upstream models to extract meaningful speech representations, have achieved unprecedented success in speech representation learning. However, their effectiveness on non-speech datasets is relatively less explored. In this work, we propose an ensemble framework, with a combination of ensemble techniques, to fuse SSL speech models' embeddings. Extensive experiments on speech and non-speech audio datasets are conducted to investigate the representation abilities of our ensemble method and its single constituent model. Ablation studies are carried out to evaluate the performances of different ensemble techniques, such as feature averaging and concatenation. All experiments are conducted during NeurIPS 2021 HEAR Challenge as a standard evaluation pipeline provided by competition officials. Results demonstrate SSL speech models' strong abilities on various non-speech tasks, while we also note that they fail to deal with fine-grained music tasks, such as pitch classification and note onset detection. In addition, feature ensemble is shown to have great potential on producing more holistic representations, as our proposed framework generally surpasses state-of-the-art SSL speech/audio models and has superior performance on various datasets compared with other teams in HEAR Challenge. Our code is available at https://github.com/tony10101105/HEAR-2021-NeurIPS-Challenge -- NTU-GURA.

cs.SD↗

arXiv2019

Mini-batch Metropolis-Hastings MCMC with Reversible SGLD Proposal

Traditional MCMC algorithms are computationally intensive and do not scale well to large data. In particular, the Metropolis-Hastings (MH) algorithm requires passing over the entire dataset to evaluate the likelihood ratio in each iteration. We propose a general framework for performing MH-MCMC using mini-batches of the whole dataset and show that this gives rise to approximately a tempered stationary distribution. We prove that the algorithm preserves the modes of the original target distribution and derive an error bound on the approximation with mild assumptions on the likelihood. To further extend the utility of the algorithm to high dimensional settings, we construct a proposal with forward and reverse moves using stochastic gradient and show that the construction leads to reasonable acceptance probabilities. We demonstrate the performance of our algorithm in both low dimensional models and high dimensional neural network applications. Particularly in the latter case, compared to popular optimization methods, our method is more robust to the choice of learning rate and improves testing accuracy.

stat.ML↗

arXiv2018

Convergence of Contrastive Divergence Algorithm in Exponential Family

The Contrastive Divergence (CD) algorithm has achieved notable success in training energy-based models including Restricted Boltzmann Machines and played a key role in the emergence of deep learning. The idea of this algorithm is to approximate the intractable term in the exact gradient of the log-likelihood function by using short Markov chain Monte Carlo (MCMC) runs. The approximate gradient is computationally-cheap but biased. Whether and why the CD algorithm provides an asymptotically consistent estimate are still open questions. This paper studies the asymptotic properties of the CD algorithm in canonical exponential families, which are special cases of the energy-based model. Suppose the CD algorithm runs $m$ MCMC transition steps at each iteration $t$ and iteratively generates a sequence of parameter estimates $\{θ_t\}_{t \ge 0}$ given an i.i.d. data sample $\{X_i\}_{i=1}^n \sim p_{θ_\star}$. Under conditions which are commonly obeyed by the CD algorithm in practice, we prove the existence of some bounded $m$ such that any limit point of the time average $\left. \sum_{s=0}^{t-1} θ_s \right/ t$ as $t \to \infty$ is a consistent estimate for the true parameter $θ_\star$. Our proof is based on the fact that $\{θ_t\}_{t \ge 0}$ is a homogenous Markov chain conditional on the data sample $\{X_i\}_{i=1}^n$. This chain meets the Foster-Lyapunov drift criterion and converges to a random walk around the Maximum Likelihood Estimate. The range of the random walk shrinks to zero at rate $\mathcal{O}(1/\sqrt[3]{n})$ as the sample size $n \to \infty$.

stat.ML↗