arXiv subjects

Nir Shavit

Publications and source records attributed to Nir Shavit.

At least 19 recordsLinked to original sources

arXiv2026

The Sparsity Whisperer

Pruning reduces the inference cost of large language models, but existing criteria primarily preserve large activations or reconstruct layer outputs. We argue that this overlooks a key computation performed by particularly sparsity-sensitive neurons in the MLP up and gate projections: separating similar inputs into dissimilar outputs. This suggests that effective pruning should preserve not only activations, but also the differences between outputs more broadly. We introduce a family of difference-informed pruning methods built upon this principle. Wisp is a first-order, update-free method that scores weights using input-difference norms, and Wisp+ refines this score neuronwise using the input pairs each neuron separates most strongly. Finally, Whisper is a second-order method that uses a lightly regularized difference Hessian as its reconstruction objective. Across Llama 2 and 3.1 models from 7B to 405B parameters, our second-order variant consistently improves over strong reconstruction-based baselines, while our update-free variants improve over activation-aware baselines, especially in constrained settings. The improvements over Wanda and SparseGPT extend to structured sparsity, downstream evaluations, and other model families. Augmenting stronger techniques such as RIA and ALPS with our difference-informed criteria yields further improvements, shifting the overall accuracy-runtime frontier outward at negligible additional cost. These results suggest that preserving output differences is a broadly useful and composable signal for post-training LLM sparsification.

cs.LG↗

arXiv2026

Toy Combinatorial Interpretability Models Reveal Lottery Tickets in Early Feature Space

The lottery ticket hypothesis posits that dense networks contain sparse subnetworks, ``winning tickets,'' that, when rewound to their initial weights and retrained in isolation, match the performance of the full model. We ask a more mechanistic question: what internal object does a winning ticket preserve? We work in a combinatorial, clause-structured toy setting that admits an interpretable feature-space representation with well-defined combinatorial distances between features. We show that winning tickets in weight space correspond to precursor locations in feature space that are already near, at initialization, to the final feature-channel codes. Dense SGD resolves these locations through structured selection: proximal locations either converge to final codes or are rejected, with rejection concentrated at more crowded neurons, implicating competition under superposition. A winning ticket is thus a family of compatible code locations that jointly balance proximity to final codes with low inter-feature interference. Sparse retraining often re-expresses the same clause/template family on a different row, so the preserved object is family-level rather than microscopic row identity. We validate this account with lightweight probes based on feature-space distance and motion; in our setting, these probes frequently outperform established weight-based ticket discovery methods in both accuracy and exact code recovery. Although these findings are grounded in a toy setting, they suggest that the lottery ticket structure is governed by hidden feature-space geometry rather than weight-space subnetwork identity.

cs.LG↗

arXiv2026

An Interpretable Latency Model for Speculative Decoding in LLM Serving

Speculative decoding (SD) accelerates large language model (LLM) inference by using a smaller draft model to propose multiple tokens that are verified by a larger target model in parallel. While prior work demonstrates substantial speedups in isolated or fixed-batch settings, the behavior of SD in production serving systems remains poorly understood: request load varies over time, and effective batch size emerges from the serving system rather than being directly controlled or observed. In this work, we develop a simple and interpretable latency model for SD in LLM serving. We infer effective batch size from request rate using Little's Law and decompose per-request demand into load-independent and load-dependent components for prefill, drafting, and verification. We validate our model using extensive measurements from vLLM across verifier and drafter model sizes, prefill and decode lengths, request rates, draft lengths, and acceptance probabilities. The model accurately describes observed latency, explains why speedups often diminish as server load increases, and characterizes how draft length, acceptance rate, and verifier-drafter size shape latency across serving conditions, with implications for configuring SD in deployed systems. We further show how the framework extends to mixture of experts models, where sparse expert activation changes the effective service costs across load regimes. Together, our results provide a structured framework for understanding SD in real LLM serving systems.

cs.LG↗

arXiv2026

Understanding Empirical Unlearning with Combinatorial Interpretability

While many recent methods aim to unlearn or remove knowledge from pretrained models, seemingly erased knowledge often persists and can be recovered in various ways. Because large foundation models are far from interpretable, understanding whether and how such knowledge persists remains a significant challenge. To address this, we turn to the recently developed framework of combinatorial interpretability. This framework, designed for two-layer neural networks, enables direct inspection of the knowledge encoded in the model weights. We reproduce baseline unlearning methods within the combinatorial interpretability setting and examine their behavior along two dimensions: (i) whether they truly remove knowledge of a target concept (the concept we wish to remove) or merely inhibit its expression while retaining the underlying information, and (ii) how easily the supposedly erased knowledge can be recovered through various fine-tuning operations. Our results shed light within a fully interpretable setting on how knowledge can persist despite unlearning and when it might resurface.

cs.LG↗

arXiv2025

Scalable Energy-Based Models via Adversarial Training: Unifying Discrimination and Generation

Simultaneously achieving robust classification and high-fidelity generative modeling within a single framework presents a significant challenge. Hybrid approaches, such as Joint Energy-Based Models (JEM), interpret classifiers as EBMs but are often limited by the instability and poor sample quality inherent in training based on Stochastic Gradient Langevin Dynamics (SGLD). We address these limitations by proposing a novel training framework that integrates adversarial training (AT) principles for both discriminative robustness and stable generative learning. The proposed method introduces three key innovations: (1) the replacement of SGLD-based JEM learning with a stable, AT-based approach that optimizes the energy function through a Binary Cross-Entropy (BCE) loss that discriminates between real data and contrastive samples generated via Projected Gradient Descent (PGD); (2) adversarial training for the discriminative component that enhances classification robustness while implicitly providing the gradient regularization needed for stable EBM training; and (3) a two-stage training strategy that addresses normalization-related instabilities and enables leveraging pretrained robust classifiers, generalizing effectively across architectures. Experiments on CIFAR-10/100 and ImageNet demonstrate that our approach: (1) is the first EBM-based hybrid to scale to high-resolution datasets with high training stability, simultaneously achieving state-of-the-art discriminative and generative performance on ImageNet 256x256; (2) uniquely combines generative quality with adversarial robustness, enabling faithful counterfactual explanations; and (3) functions as a competitive standalone generative model, matching autoregressive models and surpassing diffusion models while offering additional versatility.

cs.LG↗

arXiv2025

Expand Neurons, Not Parameters

This work demonstrates how increasing the number of neurons in a network without increasing its total number of non-zero parameters improves performance. We show that this gain corresponds with a decrease in interference between multiple features that would otherwise share the same neurons. On symbolic Boolean tasks, splitting each neuron into sparser sub-neurons with knowledge of the clauses systematically reduces polysemanticity metrics and yields higher task accuracy. Notably, even random splits of neuron weights approximate these gains, indicating that reduced collisions, not precise assignment, are a primary driver. Consistent with the superposition hypothesis, the benefits of this framework grow with increasing interference: when polysemantic load is high, accuracy improvements are the largest. Transferring these insights to more realistic models, including classifiers over CLIP embeddings, convolutional neural networks, and deeper multilayer networks, we find that widening networks while maintaining a constant non-zero parameter count consistently increases accuracy. These results identify an interpretability-grounded mechanism to leverage width against superposition, improving performance without increasing the number of non-zero parameters. Such a direction is well matched to modern accelerators, where memory movement of non-zero parameters, rather than raw compute, is often a dominant bottleneck.

cs.LG↗

arXiv2025

Learning to Interpret Weight Differences in Language Models

Finetuning (pretrained) language models is a standard approach for updating their internal parametric knowledge and specializing them to new tasks and domains. However, the corresponding model weight changes ("weight diffs") are not generally interpretable. While inspecting the finetuning dataset can give a sense of how the model might have changed, these datasets are often not publicly available or are too large to work with directly. Towards the goal of comprehensively understanding weight diffs in natural language, we introduce Diff Interpretation Tuning (DIT), a method that trains models to describe their own finetuning-induced modifications. Our approach uses synthetic, labeled weight diffs to train a DIT-adapter, which can be applied to a compatible finetuned model to make it describe how it has changed. We demonstrate in two proof-of-concept settings (reporting hidden behaviors and summarizing finetuned knowledge) that our method enables models to describe their finetuning-induced modifications using accurate natural language descriptions.

cs.LG↗

arXiv2025

Negative Pre-activations Differentiate Syntax

Modern large language models increasingly use smooth activation functions such as GELU or SiLU, allowing negative pre-activations to carry both signal and gradient. Nevertheless, many neuron-level interpretability analyses have historically focused on large positive activations, often implicitly treating the negative region as less informative, a carryover from the ReLU-era. We challenge this assumption and ask whether and how negative pre-activations are leveraged by models. We address this question by studying a sparse subpopulation of Wasserstein neurons whose output distributions deviate strongly from a Gaussian baseline and that functionally differentiate similar inputs. We show that this negative region plays an active role rather than reflecting a mere gradient optimization side effect. A minimal, sign-specific intervention that zeroes only the negative pre-activations of a small set of Wasserstein neurons substantially increases perplexity and sharply degrades grammatical performance on BLiMP and TSE, whereas both random and perplexity-matched ablations of many more non-Wasserstein neurons in their negative pre-activations leave grammatical performance largely intact. Conversely, on a suite of non-grammatical benchmarks, the perplexity-matched control ablation is more damaging than the Wasserstein neuron ablation, yielding a double dissociation between syntax and other capabilities. Part-of-speech analysis localizes the excess surprisal to syntactic scaffolding tokens, layer-specific interventions show that small local degradations accumulate across depth, and training-dynamics analysis reveals that the same sign-specific ablation becomes more harmful as Wasserstein neurons emerge and stabilize. Together, these results identify negative pre-activations in a sparse subpopulation of Wasserstein neurons as an actively used substrate for syntax in smooth-activation language models.

cs.LG↗

arXiv2025

Budgeted Broadcast: An Activity-Dependent Pruning Rule for Neural Network Efficiency

Most pruning methods remove parameters ranked by impact on loss (e.g., magnitude or gradient). We propose Budgeted Broadcast (BB), which gives each unit a local traffic budget (the product of its long-term on-rate $a_i$ and fan-out $k_i$). A constrained-entropy analysis shows that maximizing coding entropy under a global traffic budget yields a selectivity-audience balance, $\log\frac{1-a_i}{a_i}=\beta k_i$. BB enforces this balance with simple local actuators that prune either fan-in (to lower activity) or fan-out (to reduce broadcast). In practice, BB increases coding entropy and decorrelation and improves accuracy at matched sparsity across Transformers for ASR, ResNets for face identification, and 3D U-Nets for synapse prediction, sometimes exceeding dense baselines. On electron microscopy images, it attains state-of-the-art F1 and PR-AUC under our evaluation protocol. BB is easy to integrate and suggests a path toward learning more diverse and efficient representations.

cs.LG↗

arXiv2025

The Birth of Knowledge: Emergent Features across Time, Space, and Scale in Large Language Models

This paper studies the emergence of interpretable categorical features within large language models (LLMs), analyzing their behavior across training checkpoints (time), transformer layers (space), and varying model sizes (scale). Using sparse autoencoders for mechanistic interpretability, we identify when and where specific semantic concepts emerge within neural activations. Results indicate clear temporal and scale-specific thresholds for feature emergence across multiple domains. Notably, spatial analysis reveals unexpected semantic reactivation, with early-layer features re-emerging at later layers, challenging standard assumptions about representational dynamics in transformer models.

cs.CL↗

arXiv2025

Cascade Detector Analysis and Application to Biomedical Microscopy

As both computer vision models and biomedical datasets grow in size, there is an increasing need for efficient inference algorithms. We utilize cascade detectors to efficiently identify sparse objects in multiresolution images. Given an object's prevalence and a set of detectors at different resolutions with known accuracies, we derive the accuracy, and expected number of classifier calls by a cascade detector. These results generalize across number of dimensions and number of cascade levels. Finally, we compare one- and two-level detectors in fluorescent cell detection, organelle segmentation, and tissue segmentation across various microscopy modalities. We show that the multi-level detector achieves comparable performance in 30-75% less time. Our work is compatible with a variety of computer vision models and data domains.

cs.CV↗

arXiv2025

Towards Combinatorial Interpretability of Neural Computation

We introduce combinatorial interpretability, a methodology for understanding neural computation by analyzing the combinatorial structures in the sign-based categorization of a network's weights and biases. We demonstrate its power through feature channel coding, a theory that explains how neural networks compute Boolean expressions and potentially underlies other categories of neural network computation. According to this theory, features are computed via feature channels: unique cross-neuron encodings shared among the inputs the feature operates on. Because different feature channels share neurons, the neurons are polysemantic and the channels interfere with one another, making the computation appear inscrutable. We show how to decipher these computations by analyzing a network's feature channel coding, offering complete mechanistic interpretations of several small neural networks that were trained with gradient descent. Crucially, this is achieved via static combinatorial analysis of the weight matrices, without examining activations or training new autoencoding networks. Feature channel coding reframes the superposition hypothesis, shifting the focus from neuron activation directionality in high-dimensional space to the combinatorial structure of codes. It also allows us for the first time to exactly quantify and explain the relationship between a network's parameter size and its computational capacity (i.e. the set of features it can compute with low error), a relationship that is implicitly at the core of many modern scaling laws. Though our initial studies of feature channel coding are restricted to Boolean functions, we believe they provide a rich, controlled, and informative research space, and that the path we propose for combinatorial interpretation of neural computation can provide a basis for understanding both artificial and biological neural circuits.

cs.LG↗

arXiv2025

NeuroADDA: Active Discriminative Domain Adaptation in Connectomic

Training segmentation models from scratch has been the standard approach for new electron microscopy connectomics datasets. However, leveraging pretrained models from existing datasets could improve efficiency and performance in constrained annotation budget. In this study, we investigate domain adaptation in connectomics by analyzing six major datasets spanning different organisms. We show that, Maximum Mean Discrepancy (MMD) between neuron image distributions serves as a reliable indicator of transferability, and identifies the optimal source domain for transfer learning. Building on this, we introduce NeuroADDA, a method that combines optimal domain selection with source-free active learning to effectively adapt pretrained backbones to a new dataset. NeuroADDA consistently outperforms training from scratch across diverse datasets and fine-tuning sample sizes, with the largest gain observed at $n=4$ samples with a 25-67\% reduction in Variation of Information. Finally, we show that our analysis of distributional differences among neuron images from multiple species in a learned feature space reveals that these domain "distances" correlate with phylogenetic distance among those species.

cs.CV↗

arXiv2024

Jailbreak Defense in a Narrow Domain: Limitations of Existing Methods and a New Transcript-Classifier Approach

Defending large language models against jailbreaks so that they never engage in a broadly-defined set of forbidden behaviors is an open problem. In this paper, we investigate the difficulty of jailbreak-defense when we only want to forbid a narrowly-defined set of behaviors. As a case study, we focus on preventing an LLM from helping a user make a bomb. We find that popular defenses such as safety training, adversarial training, and input/output classifiers are unable to fully solve this problem. In pursuit of a better solution, we develop a transcript-classifier defense which outperforms the baseline defenses we test. However, our classifier defense still fails in some circumstances, which highlights the difficulty of jailbreak-defense even in a narrow domain.

cs.LG↗

arXiv2024

On the Complexity of Neural Computation in Superposition

Superposition, the ability of neural networks to represent more features than neurons, is increasingly seen as key to the efficiency of large models. This paper investigates the theoretical foundations of computing in superposition, establishing complexity bounds for explicit, provably correct algorithms. We present the first lower bounds for a neural network computing in superposition, showing that for a broad class of problems, including permutations and pairwise logical operations, computing $m'$ features in superposition requires at least $\Omega(\sqrt{m' \log m'})$ neurons and $\Omega(m' \log m')$ parameters. This implies an explicit limit on how much one can sparsify or distill a model while preserving its expressibility, and complements empirical scaling laws by implying the first subexponential bound on capacity: a network with $n$ neurons can compute at most $O(n^2 / \log n)$ features. Conversely, we provide a nearly tight constructive upper bound: logical operations like pairwise AND can be computed using $O(\sqrt{m'} \log m')$ neurons and $O(m' \log^2 m')$ parameters. There is thus an exponential gap between the complexity of computing in superposition (the subject of this work) versus merely representing features, which can require as little as $O(\log m')$ neurons based on the Johnson-Lindenstrauss Lemma. Our work analytically establishes that the number of parameters is a good estimator of the number of features a neural network computes.

cs.CC↗

arXiv2024

Panza: Design and Analysis of a Fully-Local Personalized Text Writing Assistant

The availability of powerful open-source large language models (LLMs) opens exciting use-cases, such as using personal data to fine-tune these models to imitate a user's unique writing style. Two key requirements for such assistants are personalization - in the sense that the assistant should recognizably reflect the user's own writing style - and privacy - users may justifiably be wary of uploading extremely personal data, such as their email archive, to a third-party service. In this paper, we present a new design and evaluation for such an automated assistant, for the specific use case of email generation, which we call Panza. Panza's personalization features are based on a combination of fine-tuning using a variant of the Reverse Instructions technique together with Retrieval-Augmented Generation (RAG). We demonstrate that this combination allows us to fine-tune an LLM to reflect a user's writing style using limited data, while executing on extremely limited resources, e.g. on a free Google Colab instance. Our key methodological contribution is the first detailed study of evaluation metrics for this personalized writing task, and of how different choices of system components--the use of RAG and of different fine-tuning approaches-impact the system's performance. Additionally, we demonstrate that very little data - under 100 email samples - are sufficient to create models that convincingly imitate humans. This finding showcases a previously-unknown attack vector in language models - that access to a small number of writing samples can allow a bad actor to cheaply create generative models that imitate a target's writing style. We are releasing the full Panza code as well as three new email datasets licensed for research use at https://github.com/IST-DASLab/PanzaMail.

cs.CL↗

arXiv2024

Wasserstein Distances, Neuronal Entanglement, and Sparsity

Disentangling polysemantic neurons is at the core of many current approaches to interpretability of large language models. Here we attempt to study how disentanglement can be used to understand performance, particularly under weight sparsity, a leading post-training optimization technique. We suggest a novel measure for estimating neuronal entanglement: the Wasserstein distance of a neuron's output distribution to a Gaussian. Moreover, we show the existence of a small number of highly entangled "Wasserstein Neurons" in each linear layer of an LLM, characterized by their highly non-Gaussian output distributions, their role in mapping similar inputs to dissimilar outputs, and their significant impact on model accuracy. To study these phenomena, we propose a new experimental framework for disentangling polysemantic neurons. Our framework separates each layer's inputs to create a mixture of experts where each neuron's output is computed by a mixture of neurons of lower Wasserstein distance, each better at maintaining accuracy when sparsified without retraining. We provide strong evidence that this is because the mixture of sparse experts is effectively disentangling the input-output relationship of individual neurons, in particular the difficult Wasserstein neurons.

cs.LG↗

arXiv2023

Forbidden Facts: An Investigation of Competing Objectives in Llama-2

LLMs often face competing pressures (for example helpfulness vs. harmlessness). To understand how models resolve such conflicts, we study Llama-2-chat models on the forbidden fact task. Specifically, we instruct Llama-2 to truthfully complete a factual recall statement while forbidding it from saying the correct answer. This often makes the model give incorrect answers. We decompose Llama-2 into 1000+ components, and rank each one with respect to how useful it is for forbidding the correct answer. We find that in aggregate, around 35 components are enough to reliably implement the full suppression behavior. However, these components are fairly heterogeneous and many operate using faulty heuristics. We discover that one of these heuristics can be exploited via a manually designed adversarial attack which we call The California Attack. Our results highlight some roadblocks standing in the way of being able to successfully interpret advanced ML systems. Project website available at https://forbiddenfacts.github.io .

cs.LG↗