arXiv subjects

Rajashree Agrawal

Publications and source records attributed to Rajashree Agrawal.

9 recordsLinked to original sources

arXiv2026

Interactions Between Crosscoder Features: A Compact Proofs Perspective

Dictionary learning methods like Sparse Autoencoders (SAEs) and crosscoders attempt to explain a model by decomposing its activations into independent features. Interactions between features hence induce errors in the reconstruction. We formalize this intuition via compact proofs and make five contributions. First, we show how, \textit{in principle}, a compact proof of model performance can be constructed using a crosscoder. Second, we show that an error term arising in this proof can naturally be interpreted as a measure of interaction between crosscoder features and provide an explicit expression for the interaction term in the Multi-Layer Perceptron (MLP) layers. We then provide three applications of this new interaction measure. In our third contribution we show that the interaction term itself can be used as a differentiable loss penalty. Applying this penalty, we can achieve ``computationally sparse'' crosscoders that retain $60\%$ of MLP performance when only keeping a single feature at each datapoint and neuron, compared to $10\%$ in standard crosscoders. We then show that clustering according to our interaction measure provides semantically meaningful feature clusters, and finally that sleeper agents have significant interactions. Code is available at https://github.com/chainik1125/crosscoders-feature-interactions/tree/arxiv.

cs.LG↗

arXiv2025

Automatic Test-Case Reduction in Proof Assistants: A Case Study in Coq

As the adoption of proof assistants increases, there is a need for efficiency in identifying, documenting, and fixing compatibility issues that arise from proof assistant evolution. We present the Coq Bug Minimizer, a tool for reproducing buggy behavior with minimal and standalone files, integrated with coqbot to trigger automatically on Coq reverse CI failures. Our tool eliminates the overhead of having to download, set up, compile, and then explore and understand large developments: enabling Coq developers to easily obtain modular test-case files for fast experimentation. In this paper, we describe insights about how test-case reduction is different in Coq than in traditional compilers. We expect that our insights will generalize to other proof assistants. We evaluate the Coq Bug Minimizer on over 150 CI failures. Our tool succeeds in reducing failures to smaller test cases in roughly 75% of the time. The minimizer produces a fully standalone test case 89% of the time, and it is on average about one-third the size of the original test. The average reduced test case compiles in 1.25 seconds, with 75% taking under half a second.

cs.SE↗

arXiv2025

Accelerating Verified-Compiler Development with a Verified Rewriting Engine

Compilers are a prime target for formal verification, since compiler bugs invalidate higher-level correctness guarantees, but compiler changes may become more labor-intensive to implement, if they must come with proof patches. One appealing approach is to present compilers as sets of algebraic rewrite rules, which a generic engine can apply efficiently. Now each rewrite rule can be proved separately, with no need to revisit past proofs for other parts of the compiler. We present the first realization of this idea, in the form of a framework for the Coq proof assistant. Our new Coq command takes normal proved theorems and combines them automatically into fast compilers with proofs. We applied our framework to improve the Fiat Cryptography toolchain for generating cryptographic arithmetic, producing an extracted command-line compiler that is about 1000$\times$ faster while actually featuring simpler compiler-specific proofs.

cs.PL↗

arXiv2024

Compact Proofs of Model Performance via Mechanistic Interpretability

We propose using mechanistic interpretability -- techniques for reverse engineering model weights into human-interpretable algorithms -- to derive and compactly prove formal guarantees on model performance. We prototype this approach by formally proving accuracy lower bounds for a small transformer trained on Max-of-K, validating proof transferability across 151 random seeds and four values of K. We create 102 different computer-assisted proof strategies and assess their length and tightness of bound on each of our models. Using quantitative metrics, we find that shorter proofs seem to require and provide more mechanistic understanding. Moreover, we find that more faithful mechanistic understanding leads to tighter performance bounds. We confirm these connections by qualitatively examining a subset of our proofs. Finally, we identify compounding structureless errors as a key challenge for using mechanistic interpretability to generate compact proofs on model performance.

cs.LG↗

arXiv2024

Failures to Find Transferable Image Jailbreaks Between Vision-Language Models

The integration of new modalities into frontier AI systems offers exciting capabilities, but also increases the possibility such systems can be adversarially manipulated in undesirable ways. In this work, we focus on a popular class of vision-language models (VLMs) that generate text outputs conditioned on visual and textual inputs. We conducted a large-scale empirical study to assess the transferability of gradient-based universal image ``jailbreaks" using a diverse set of over 40 open-parameter VLMs, including 18 new VLMs that we publicly release. Overall, we find that transferable gradient-based image jailbreaks are extremely difficult to obtain. When an image jailbreak is optimized against a single VLM or against an ensemble of VLMs, the jailbreak successfully jailbreaks the attacked VLM(s), but exhibits little-to-no transfer to any other VLMs; transfer is not affected by whether the attacked and target VLMs possess matching vision backbones or language models, whether the language model underwent instruction-following and/or safety-alignment training, or many other factors. Only two settings display partially successful transfer: between identically-pretrained and identically-initialized VLMs with slightly different VLM training data, and between different training checkpoints of a single VLM. Leveraging these results, we then demonstrate that transfer can be significantly improved against a specific target VLM by attacking larger ensembles of ``highly-similar" VLMs. These results stand in stark contrast to existing evidence of universal and transferable text jailbreaks against language models and transferable adversarial attacks against image classifiers, suggesting that VLMs may be more robust to gradient-based transfer attacks.

cs.CL↗

arXiv2024

Modular addition without black-boxes: Compressing explanations of MLPs that compute numerical integration

The goal of mechanistic interpretability is discovering simpler, low-rank algorithms implemented by models. While we can compress activations into features, compressing nonlinear feature-maps -- like MLP layers -- is an open problem. In this work, we present the first case study in rigorously compressing nonlinear feature-maps, which are the leading asymptotic bottleneck to compressing small transformer models. We work in the classic setting of the modular addition models, and target a non-vacuous bound on the behaviour of the ReLU MLP in time linear in the parameter-count of the circuit. To study the ReLU MLP analytically, we use the infinite-width lens, which turns post-activation matrix multiplications into approximate integrals. We discover a novel interpretation of} the MLP layer in one-layer transformers implementing the ``pizza'' algorithm: the MLP can be understood as evaluating a quadrature scheme, where each neuron computes the area of a rectangle under the curve of a trigonometric integral identity. Our code is available at https://tinyurl.com/mod-add-integration.

cs.LG↗

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

Towards a Scalable Proof Engine: A Performant Prototype Rewriting Primitive for Coq

We address the challenges of scaling verification efforts to match the increasing complexity and size of systems. We propose a research agenda aimed at building a performant proof engine by studying the asymptotic performance of proof engines and redesigning their building blocks. As a case study, we explore equational rewriting and introduce a novel prototype proof engine building block for rewriting in Coq, utilizing proof by reflection for enhanced performance. Our prototype implementation can significantly improve the development of verified compilers, as demonstrated in a case study with the Fiat Cryptography toolchain. The resulting extracted command-line compiler is about 1000$\times$ faster while featuring simpler compiler-specific proofs. This work lays some foundation for scaling verification efforts and contributes to the broader goal of developing a proof engine with good asymptotic performance, ultimately aimed at enabling the verification of larger and more complex systems.

cs.PL↗

arXiv2024

Is Model Collapse Inevitable? Breaking the Curse of Recursion by Accumulating Real and Synthetic Data

The proliferation of generative models, combined with pretraining on web-scale data, raises a timely question: what happens when these models are trained on their own generated outputs? Recent investigations into model-data feedback loops proposed that such loops would lead to a phenomenon termed model collapse, under which performance progressively degrades with each model-data feedback iteration until fitted models become useless. However, those studies largely assumed that new data replace old data over time, where an arguably more realistic assumption is that data accumulate over time. In this paper, we ask: what effect does accumulating data have on model collapse? We empirically study this question by pretraining sequences of language models on text corpora. We confirm that replacing the original real data by each generation's synthetic data does indeed tend towards model collapse, then demonstrate that accumulating the successive generations of synthetic data alongside the original real data avoids model collapse; these results hold across a range of model sizes, architectures, and hyperparameters. We obtain similar results for deep generative models on other types of real data: diffusion models for molecule conformation generation and variational autoencoders for image generation. To understand why accumulating data can avoid model collapse, we use an analytically tractable framework introduced by prior work in which a sequence of linear models are fit to the previous models' outputs. Previous work used this framework to show that if data are replaced, the test error increases with the number of model-fitting iterations; we extend this argument to prove that if data instead accumulate, the test error has a finite upper bound independent of the number of iterations, meaning model collapse no longer occurs.

cs.LG↗