Search

Static single assignment form, or SSA, has been the dominant compiler intermediate representation for decades. In this paper, we give a type theory for a variant of SSA, including its equational theory, which are strong enough to validate a variety of control and data flow transformations. We also give a categorical semantics for SSA, and show that the type theory is sound and complete with respect to the categorical axiomatization. We demonstrate the utility of our model by exhibiting a variety of concrete models satisfying our axioms, including in particular a model of TSO weak memory. The correctness of the syntactic metatheory, as well as the completeness proof has been mechanized in the Lean proof assistant., Comment: 94 pages, 38 figures, mechanization available at https://github.com/imbrem/debruijn-ssa/tree/toplas-artifact

Safely handling objects and avoiding slippage are fundamental challenges in robotic manipulation, yet traditional techniques often oversimplify the issue by treating slippage as a binary occurrence. Our research presents a framework that both identifies slip incidents and measures their severity. We introduce a set of features based on detailed vector field analysis of tactile deformation data captured by the GelSight Mini sensor. Two distinct machine learning models use these features: one focuses on slip detection, and the other evaluates the slip's severity, which is the slipping velocity of the object against the sensor surface. Our slip detection model achieves an average accuracy of 92%, and the slip severity estimation model exhibits a mean absolute error (MAE) of 0.6 cm/s for unseen objects. To demonstrate the synergistic approach of this framework, we employ both the models in a tactile feedback-guided vertical sliding task. Leveraging the high accuracy of slip detection, we utilize it as the foundational and corrective model and integrate the slip severity estimation into the feedback control loop to address slips without overcompensating., Comment: Accepted at IROS 2024

Adiabatic quantum-flux-parametron (AQFP) logic is a proven energy-efficient superconductor technology for various applications. To address the scalability challenges, we investigated AQFP shift registers with the AQFP footprint area reduced by 25% with respect to prior work and with more than 2x denser overall designs obtained by eliminating the previously used free space between the AQFPs. We also investigated AQFP cells with different designs of flux trapping moats in the superconducting ground plane as well as compact AQFP cells that took advantage of the smaller feature sizes available in the new fabrication process, SFQ5ee+, at MIT Lincoln Laboratory. This new process features nine planarized Nb layers with a 0.25 $\mu$m minimum linewidth. The fabricated circuits were tested in a liquid He probe and in a closed-cycle cryocooler using a controlled cooling rate through the superconducting critical temperature. Using multiple thermal cycles, we investigated flux trapping in the dense AQFP shift registers as well as in the registers using the old (sparse) AQFP designs at two levels of the residual magnetic field, about 0.53 $\mu$T and about 1.2 $\mu$T. The sparse designs demonstrated 95% to almost 100% probability of operation after the cooldown and very wide operation margins, although the flux trapping probability was increasing with circuit complexities. The margins were similarly wide in the newer dense designs, but flux trapping probability that rendered the registers nonoperational was significantly, by an order of magnitude, higher in the denser circuits and was also very sensitive to the moats' shape and location. Our findings indicate that AQFP circuits are amendable to increasing the scale of integration and further densification, but a careful moat design and optimization are required to reduce flux trapping effects in the dense AQFP circuits., Comment: 5 pages, 10 figures, 2 tables, 18 references. Presented at Applied Superconductivity Conference, ASC2024, 1-6 September 2024, Salt Lake City, UT, USA

Current volumetric biomedical foundation models struggle to generalize as public 3D datasets are small and do not cover the broad diversity of medical procedures, conditions, anatomical regions, and imaging protocols. We address this by creating a representation learning method that instead anticipates strong domain shifts at training time itself. We first propose a data engine that synthesizes highly variable training samples that enable generalization to new biomedical contexts. To then train a single 3D network for any voxel-level task, we develop a contrastive learning method that pretrains the network to be stable against nuisance imaging variation simulated by the data engine, a key inductive bias for generalization. This network's features can be used as robust representations of input images for downstream tasks and its weights provide a strong, dataset-agnostic initialization for finetuning on new datasets. As a result, we set new standards across both multimodality registration and few-shot segmentation, a first for any 3D biomedical vision model, all without (pre-)training on any existing dataset of real images., Comment: Code and model weights available at https://github.com/neel-dey/anatomix

Multimodal foundation models offer a promising framework for robotic perception and planning by processing sensory inputs to generate actionable plans. However, addressing uncertainty in both perception (sensory interpretation) and decision-making (plan generation) remains a critical challenge for ensuring task reliability. We present a comprehensive framework to disentangle, quantify, and mitigate these two forms of uncertainty. We first introduce a framework for uncertainty disentanglement, isolating perception uncertainty arising from limitations in visual understanding and decision uncertainty relating to the robustness of generated plans. To quantify each type of uncertainty, we propose methods tailored to the unique properties of perception and decision-making: we use conformal prediction to calibrate perception uncertainty and introduce Formal-Methods-Driven Prediction (FMDP) to quantify decision uncertainty, leveraging formal verification techniques for theoretical guarantees. Building on this quantification, we implement two targeted intervention mechanisms: an active sensing process that dynamically re-observes high-uncertainty scenes to enhance visual input quality and an automated refinement procedure that fine-tunes the model on high-certainty data, improving its capability to meet task specifications. Empirical validation in real-world and simulated robotic tasks demonstrates that our uncertainty disentanglement framework reduces variability by up to 40% and enhances task success rates by 5% compared to baselines. These improvements are attributed to the combined effect of both interventions and highlight the importance of uncertainty disentanglement which facilitates targeted interventions that enhance the robustness and reliability of autonomous systems., Comment: Fine-tuned models, code, and datasets are available at https://tinyurl.com/uncertainty-disentanglement

Machine unlearning -- efficiently removing the effect of a small "forget set" of training data on a pre-trained machine learning model -- has recently attracted significant research interest. Despite this interest, however, recent work shows that existing machine unlearning techniques do not hold up to thorough evaluation in non-convex settings. In this work, we introduce a new machine unlearning technique that exhibits strong empirical performance even in such challenging settings. Our starting point is the perspective that the goal of unlearning is to produce a model whose outputs are statistically indistinguishable from those of a model re-trained on all but the forget set. This perspective naturally suggests a reduction from the unlearning problem to that of data attribution, where the goal is to predict the effect of changing the training set on a model's outputs. Thus motivated, we propose the following meta-algorithm, which we call Datamodel Matching (DMM): given a trained model, we (a) use data attribution to predict the output of the model if it were re-trained on all but the forget set points; then (b) fine-tune the pre-trained model to match these predicted outputs. In a simple convex setting, we show how this approach provably outperforms a variety of iterative unlearning algorithms. Empirically, we use a combination of existing evaluations and a new metric based on the KL-divergence to show that even in non-convex settings, DMM achieves strong unlearning performance relative to existing algorithms. An added benefit of DMM is that it is a meta-algorithm, in the sense that future advances in data attribution translate directly into better unlearning algorithms, pointing to a clear direction for future progress in unlearning.

Evaluations are limited by benchmark availability. As models evolve, there is a need to create benchmarks that can measure progress on new generative capabilities. However, creating new benchmarks through human annotations is slow and expensive, restricting comprehensive evaluations for any capability. We introduce BENCHAGENTS, a framework that methodically leverages large language models (LLMs) to automate benchmark creation for complex capabilities while inherently ensuring data and metric quality. BENCHAGENTS decomposes the benchmark creation process into planning, generation, data verification, and evaluation, each of which is executed by an LLM agent. These agents interact with each other and utilize human-in-the-loop feedback from benchmark developers to explicitly improve and flexibly control data diversity and quality. We use BENCHAGENTS to create benchmarks to evaluate capabilities related to planning and constraint satisfaction during text generation. We then use these benchmarks to study seven state-of-the-art models and extract new insights on common failure modes and model differences.

With models getting stronger, evaluations have grown more complex, testing multiple skills in one benchmark and even in the same instance at once. However, skill-wise performance is obscured when inspecting aggregate accuracy, under-utilizing the rich signal modern benchmarks contain. We propose an automatic approach to recover the underlying skills relevant for any evaluation instance, by way of inspecting model-generated rationales. After validating the relevance of rationale-parsed skills and inferring skills for $46$k instances over $12$ benchmarks, we observe many skills to be common across benchmarks, resulting in the curation of hundreds of skill-slices (i.e. sets of instances testing a common skill). Inspecting accuracy over these slices yields novel insights on model trade-offs: e.g., compared to GPT-4o and Claude 3.5 Sonnet, on average, Gemini 1.5 Pro is $18\%$ more accurate in "computing molar mass", but $19\%$ less accurate in "applying constitutional law", despite the overall accuracies of the three models differing by a mere $0.4\%$. Furthermore, we demonstrate the practical utility of our approach by showing that insights derived from skill slice analysis can generalize to held-out instances: when routing each instance to the model strongest on the relevant skills, we see a $3\%$ accuracy improvement over our $12$ dataset corpus. Our skill-slices and framework open a new avenue in model evaluation, leveraging skill-specific analyses to unlock a more granular and actionable understanding of model capabilities., Comment: Code at: github.com/microsoft/skill-slice-insights

Single metallocene molecules act as sensitive spin detectors when decorating the probe of a scanning tunneling microscope (STM). However, the impact of the atomic-scale electrode details on the molecular spin state has remained elusive to date. Here, a nickelocene (Nc) STM junction is manipulated in an atom-wise manner showing clearly the dependence of the spin excitation spectrum on the anchoring of Nc to Cu(111), a Cu monomer and trimer. Moreover, while the spin state of the same Nc tip is a triplet with tunable spin excitation energies upon contacting the surface, it transitions to a Kondo-screened doublet on a Cu atom. Notably, the non-trivial magnetic exchange interaction of the molecular spin with the electron continuum of the substrate determines the spectral line shape of the spin excitations.

This paper proposes a hybrid quantum neural network (HQNN) for indoor user localization using received signal strength indicator (RSSI) values. We use publicly available RSSI datasets for indoor localization using WiFi, Bluetooth, and Zigbee to test the performance of the proposed HQNN. We also compare the performance of the HQNN with the recently proposed quantum fingerprinting-based user localization method. Our results show that the proposed HQNN performs better than the quantum fingerprinting algorithm since the HQNN has trainable parameters in the quantum circuits, whereas the quantum fingerprinting algorithm uses a fixed quantum circuit to calculate the similarity between the test data point and the fingerprint dataset. Unlike prior works, we also test the performance of the HQNN and quantum fingerprint algorithm on a real IBM quantum computer using cloud quantum computing services. Therefore, this paper examines the performance of the HQNN on noisy intermediate scale (NISQ) quantum devices using real-world RSSI localization datasets. The novelty of our approach lies in the use of simple feature maps and ansatz with fewer neurons, alongside testing on actual quantum hardware using real-world data, demonstrating practical applicability in real-world scenarios., Comment: This work has been accepted for presentation at the IEEE TENSYMP 2024 conference

Recent advancements in Large Language Models (LLMs) have showcased their ability to perform complex reasoning tasks, but their effectiveness in planning remains underexplored. In this study, we evaluate the planning capabilities of OpenAI's o1 models across a variety of benchmark tasks, focusing on three key aspects: feasibility, optimality, and generalizability. Through empirical evaluations on constraint-heavy tasks (e.g., $\textit{Barman}$, $\textit{Tyreworld}$) and spatially complex environments (e.g., $\textit{Termes}$, $\textit{Floortile}$), we highlight o1-preview's strengths in self-evaluation and constraint-following, while also identifying bottlenecks in decision-making and memory management, particularly in tasks requiring robust spatial reasoning. Our results reveal that o1-preview outperforms GPT-4 in adhering to task constraints and managing state transitions in structured environments. However, the model often generates suboptimal solutions with redundant actions and struggles to generalize effectively in spatially complex tasks. This pilot study provides foundational insights into the planning limitations of LLMs, offering key directions for future research on improving memory management, decision-making, and generalization in LLM-based planning. Code available at https://github.com/VITA-Group/o1-planning., Comment: Code available at https://github.com/VITA-Group/o1-planning

In this paper, we propose a novel flexible optimization pipeline for determining the optimal adsorption sites, named AUGUR (Aware of Uncertainty Graph Unit Regression). Our model combines graph neural networks and Gaussian processes to create a flexible, efficient, symmetry-aware, translation, and rotation-invariant predictor with inbuilt uncertainty quantification. This predictor is then used as a surrogate for a data-efficient Bayesian Optimization scheme to determine the optimal adsorption positions. This pipeline determines the optimal position of large and complicated clusters with far fewer iterations than current state-of-the-art approaches. Further, it does not rely on hand-crafted features and can be seamlessly employed on any molecule without any alterations. Additionally, the pooling properties of graphs allow for the processing of molecules of different sizes by the same model. This allows the energy prediction of computationally demanding systems by a model trained on comparatively smaller and less expensive ones, Comment: 17 pages 8 figures

Inference-time techniques are emerging as highly effective tools to enhance large language model (LLM) capabilities. However, best practices for developing systems that combine these techniques remain underdeveloped due to our limited understanding of the utility of individual inference-time techniques and the interactions between them. Additionally, efficiently and automatically searching the space of model choices, inference-time techniques, and their compositions is challenging due to the large design space. To address these challenges, we introduce Archon, a modular framework for selecting, combining, and stacking layers of inference-time techniques to construct optimized LLM systems for target benchmarks. Rather than relying on a single LLM called once, we leverage a diverse set of LLMs and inference-time techniques, creating LLM systems greater than the sum of their parts. Archon defines an extensible design space, encompassing techniques such as generation ensembling, repeated sampling, ranking, fusion, critiquing, verification, and unit testing. It transforms the problem of building LLM systems into a hyperparameter optimization objective. Given the available LLMs, inference-time techniques, and compute budget, Archon utilizes hyperparameter search techniques to discover optimized architectures for target benchmark(s). We evaluate Archon architectures across a range of instruction-following, reasoning, and coding benchmarks, including MT-Bench, Arena-Hard-Auto, AlpacaEval 2.0, MixEval, MixEval Hard, MATH, and CodeContests. Archon architectures outperform frontier models, such as GPT-4o and Claude 3.5 Sonnet, on these benchmarks, achieving an average accuracy increase of 15.1 percentage points by using all available LLMs. We make our code and datasets available publicly on Github: https://github.com/ScalingIntelligence/Archon.

Due to the numerous limitations of current quantum devices, quantum error mitigation methods become potential solutions for realizing practical quantum applications in the near term. Zero-Noise Extrapolation (ZNE) and Clifford Data Regression (CDR) are two promising quantum error mitigation methods. Based on the characteristics of these two methods, we propose a new method named extrapolated CDR (eCDR). To benchmark our method, we embed eCDR into a quantum application, specifically multi-layer quantum routing. Quantum routers direct a quantum signal from one input path to a quantum superposition of multiple output paths and are considered important elements of future quantum networks. Multi-layer quantum routers extend the scalability of quantum networks by allowing for further superposition of paths. We benchmark the performance of multi-layer quantum routers implemented on current superconducting quantum devices instantiated with the ZNE, CDR, and eCDR methods. Our experimental results show that the new eCDR method significantly outperforms ZNE and CDR for the 2-layer quantum router. Our work highlights how new mitigation methods built from different components of pre-existing methods, and designed with a core application in mind, can lead to significant performance enhancements.

Although powerful, current cutting-edge LLMs may not fulfil the needs of highly specialised sectors. We introduce KodeXv0.1, a family of large language models that outclass GPT-4 in financial question answering. We utilise the base variants of Llama 3.1 8B and 70B and adapt them to the financial domain through a custom training regime. To this end, we collect and process a large number of publicly available financial documents such as earnings calls and business reports. These are used to generate a high-quality, synthetic dataset consisting of Context-Question-Answer triplets which closely mirror real-world financial tasks. Using the train split of this dataset, we perform RAG-aware 4bit LoRA instruction tuning runs of Llama 3.1 base variants to produce KodeX-8Bv0.1 and KodeX-70Bv0.1. We then complete extensive model evaluations using FinanceBench, FinQABench and the withheld test split of our dataset. Our results show that KodeX-8Bv0.1 is more reliable in financial contexts than cutting-edge instruct models in the same parameter regime, surpassing them by up to 9.24%. In addition, it is even capable of outperforming state-of-the-art proprietary models such as GPT-4 by up to 7.07%. KodeX-70Bv0.1 represents a further improvement upon this, exceeding GPT-4's performance on every tested benchmark., Comment: 11 pages, 8 figures

Rigorous and reproducible evaluation is critical for assessing the state of the art and for guiding scientific advances in Artificial Intelligence. Evaluation is challenging in practice due to several reasons, including benchmark saturation, lack of transparency in methods used for measurement, development challenges in extracting measurements for generative tasks, and, more generally, the extensive number of capabilities required for a well-rounded comparison across models. We make three contributions to alleviate the above challenges. First, we present Eureka, an open-source framework for standardizing evaluations of large foundation models beyond single-score reporting and rankings. Second, we introduce Eureka-Bench as an extensible collection of benchmarks testing capabilities that (i) are still challenging for state-of-the-art models and (ii) represent fundamental but overlooked language and multimodal capabilities. The inherent space for improvement in non-saturated benchmarks enables us to discover meaningful differences between models at a capability level. Third, using Eureka, we conduct an analysis of 12 state-of-the-art models, providing in-depth insights into failure understanding and model comparison, which can be leveraged to plan targeted improvements. In contrast to recent trends in reports and leaderboards showing absolute rankings and claims for one model or another to be the best, our analysis shows that there is no such best model. Different models have different strengths, but there are models that appear more often than others as best performers for some capabilities. Despite the recent improvements, current models still struggle with several fundamental capabilities including detailed image understanding, benefiting from multimodal input when available rather than fully relying on language, factuality and grounding for information retrieval, and over refusals.

This letter reports the first complete observation of magnetized collisionless shock precursors formed through the compression of Biermann-battery magnetic fields in laser produced plasmas. At OMEGA, lasers produce a supersonic CH plasma flow which is magnetized with Biermann-battery magnetic fields. The plasma flow collides with an unmagnetized hydrogen gas jet plasma to create a magnetized shock precursor. The situation where the flowing plasma carries the magnetic field is similar to the Venusian bow shock. Imaging 2$\omega$ Thomson scattering confirms that the interaction is collisionless and shows density and temperature jumps. Proton radiographs have regions of strong deflections and FLASH magnetohydrodynamic (MHD) simulations show the presence of Biermann fields in the Thomson scattering region. Electrons are accelerated to energies of up to 100 keV in a power-law spectrum. OSIRIS particle-in-cell (PIC) simulations, initialized with measured parameters, show the formation of a magnetized shock precursor and corroborate the experimental observables., Comment: 6 pages, 5 figures

The end of Moore's Law is driving cloud providers to offload virtualization and the network data plane to SmartNICs to improve compute efficiency. Even though individual OS control plane tasks consume up to 5% of cycles across the fleet, they remain on the host CPU because they are tightly intertwined with OS mechanisms. Moreover, offloading puts the slow PCIe interconnect in the critical path of OS decisions. We propose Tide, a new split OS architecture that separates OS control plane policies from mechanisms and offloads the control plane policies onto a SmartNIC. Tide has a new host-SmartNIC communication API, state synchronization mechanism, and communication mechanisms that overcome the PCIe bottleneck, even for $\mu$s-scale workloads. Tide frees up host compute for applications and unlocks new optimization opportunities, including machine learning-driven policies, scheduling on the network I/O path, and reducing on-host interference. We demonstrate that Tide enables OS control planes that are competitive with on-host performance for the most difficult $\mu$s-scale workloads. Tide outperforms on-host control planes for memory management (saving 16 host cores), Stubby network RPCs (saving 8 cores), and GCE virtual machine management (11.2% performance improvement)., Comment: About 11 pages

Block Encoding (BE) is a crucial subroutine in many modern quantum algorithms, including those with near-optimal scaling for simulating quantum many-body systems, which often rely on Quantum Signal Processing (QSP). Currently, the primary methods for constructing BEs are the Linear Combination of Unitaries (LCU) and the sparse oracle approach. In this work, we demonstrate that QSP-based techniques, such as Quantum Singular Value Transformation (QSVT) and Quantum Eigenvalue Transformation for Unitary Matrices (QETU), can themselves be efficiently utilized for BE implementation. Specifically, we present several examples of using QSVT and QETU algorithms, along with their combinations, to block encode Hamiltonians for lattice bosons, an essential ingredient in simulations of high-energy physics. We also introduce a straightforward approach to BE based on the exact implementation of Linear Operators Via Exponentiation and LCU (LOVE-LCU). We find that, while using QSVT for BE results in the best asymptotic gate count scaling with the number of qubits per site, LOVE-LCU outperforms all other methods for operators acting on up to $\lesssim11$ qubits, highlighting the importance of concrete circuit constructions over mere comparisons of asymptotic scalings. Using LOVE-LCU to implement the BE, we simulate the time evolution of single-site and two-site systems in the lattice $\varphi^4$ theory using the Generalized QSP algorithm and compare the gate counts to those required for Trotter simulation., Comment: 20 pages, 16 Figures, 2 Tables, 2 Appendices

Current backdoor defense methods are evaluated against a single attack at a time. This is unrealistic, as powerful machine learning systems are trained on large datasets scraped from the internet, which may be attacked multiple times by one or more attackers. We demonstrate that simultaneously executed data poisoning attacks can effectively install multiple backdoors in a single model without substantially degrading clean accuracy. Furthermore, we show that existing backdoor defense methods do not effectively prevent attacks in this setting. Finally, we leverage insights into the nature of backdoor attacks to develop a new defense, BaDLoss, that is effective in the multi-attack setting. With minimal clean accuracy degradation, BaDLoss attains an average attack success rate in the multi-attack setting of 7.98% in CIFAR-10 and 10.29% in GTSRB, compared to the average of other defenses at 64.48% and 84.28% respectively.

Sparse autoencoders (SAEs) are an unsupervised method for learning a sparse decomposition of a neural network's latent representations into seemingly interpretable features. Despite recent excitement about their potential, research applications outside of industry are limited by the high cost of training a comprehensive suite of SAEs. In this work, we introduce Gemma Scope, an open suite of JumpReLU SAEs trained on all layers and sub-layers of Gemma 2 2B and 9B and select layers of Gemma 2 27B base models. We primarily train SAEs on the Gemma 2 pre-trained models, but additionally release SAEs trained on instruction-tuned Gemma 2 9B for comparison. We evaluate the quality of each SAE on standard metrics and release these results. We hope that by releasing these SAE weights, we can help make more ambitious safety and interpretability research easier for the community. Weights and a tutorial can be found at https://huggingface.co/google/gemma-scope and an interactive demo can be found at https://www.neuronpedia.org/gemma-scope, Comment: 12 main text pages, and 14 pages of acknowledgements, references and appendices

Precisely estimating lumen boundaries in intravascular ultrasound (IVUS) is needed for sizing interventional stents to treat deep vein thrombosis (DVT). Unfortunately, current segmentation networks like the UNet lack the precision needed for clinical adoption in IVUS workflows. This arises due to the difficulty of automatically learning accurate lumen contour from limited training data while accounting for the radial geometry of IVUS imaging. We propose the Geo-UNet framework to address these issues via a design informed by the geometry of the lumen contour segmentation task. We first convert the input data and segmentation targets from Cartesian to polar coordinates. Starting from a convUNet feature extractor, we propose a two-task setup, one for conventional pixel-wise labeling and the other for single boundary lumen-contour localization. We directly combine the two predictions by passing the predicted lumen contour through a new activation (named CDFeLU) to filter out spurious pixel-wise predictions. Our unified loss function carefully balances area-based, distance-based, and contour-based penalties to provide near clinical-grade generalization in unseen patient data. We also introduce a lightweight, inference-time technique to enhance segmentation smoothness. The efficacy of our framework on a venous IVUS dataset is shown against state-of-the-art models., Comment: Accepted into the 15th workshop on Machine Learning in Medical Imaging at MICCAI 2024. (* indicates equal contribution)

AI progress is creating a growing range of risks and opportunities, but it is often unclear how they should be navigated. In many cases, the barriers and uncertainties faced are at least partly technical. Technical AI governance, referring to technical analysis and tools for supporting the effective governance of AI, seeks to address such challenges. It can help to (a) identify areas where intervention is needed, (b) identify and assess the efficacy of potential governance actions, and (c) enhance governance options by designing mechanisms for enforcement, incentivization, or compliance. In this paper, we explain what technical AI governance is, why it is important, and present a taxonomy and incomplete catalog of its open problems. This paper is intended as a resource for technical researchers or research funders looking to contribute to AI governance., Comment: Ben Bucknall and Anka Reuel contributed equally and share the first author position

Sparse autoencoders (SAEs) are a promising unsupervised approach for identifying causally relevant and interpretable linear features in a language model's (LM) activations. To be useful for downstream tasks, SAEs need to decompose LM activations faithfully; yet to be interpretable the decomposition must be sparse -- two objectives that are in tension. In this paper, we introduce JumpReLU SAEs, which achieve state-of-the-art reconstruction fidelity at a given sparsity level on Gemma 2 9B activations, compared to other recent advances such as Gated and TopK SAEs. We also show that this improvement does not come at the cost of interpretability through manual and automated interpretability studies. JumpReLU SAEs are a simple modification of vanilla (ReLU) SAEs -- where we replace the ReLU with a discontinuous JumpReLU activation function -- and are similarly efficient to train and run. By utilising straight-through-estimators (STEs) in a principled manner, we show how it is possible to train JumpReLU SAEs effectively despite the discontinuous JumpReLU function introduced in the SAE's forward pass. Similarly, we use STEs to directly train L0 to be sparse, instead of training on proxies such as L1, avoiding problems like shrinkage., Comment: v2: new appendix H comparing kernel functions & bug-fixes to pseudo-code in Appendix J v3: further bug-fix to pseudo-code in Appendix J

The use of quantum cryptography in everyday applications has gained attention in both industrial and academic fields. Due to advancements in quantum electronics, practical quantum devices are already available in the market, and ready for wider use. Quantum Key Distribution (QKD) is a crucial aspect of quantum cryptography, which involves generating and distributing symmetric cryptographic keys between geographically separated users using principles of quantum physics. Many successful QKD networks have been established to test different solutions. The objective of this paper is to delve into the potential of utilizing established routing design techniques in the context of quantum key distribution, a field distinguished by its unique properties rooted in the principles of quantum mechanics. However, the implementation of these techniques poses substantial challenges, including quantum memory decoherence, key rate generation, latency delays, inherent noise in quantum systems, limited communication ranges, and the necessity for highly specialized hardware. This paper conducts an in-depth examination of essential research pertaining to the design methodologies for quantum key distribution. It also explores the fundamental aspects of quantum routing and the associated properties inherent to quantum QKD. This paper elucidates the necessary steps for constructing efficient and resilient QKD networks. In summarizing the techniques relevant to QKD networking and routing, including their underlying principles, protocols, and challenges, this paper sheds light on potential applications and delineates future research directions in this burgeoning field.

The evolution of science education is a dynamic process driven by advances in pedagogy, technology, and especially, our understanding of how students learn. Educators are exploring innovative teaching and learning methodologies such as active learning, incorporated technology, interdisciplinary approach, flipped classrooms, personalized teaching, and many more. The goal of all these evolving methodologies is to empower students with not only a strong foundation in scientific knowledge but also with the skills and mindset required to thrive in the future world. By adopting these innovative approaches, educators can help students become effective problem solvers, critical thinkers, and life learning citizens. Our focus is redesigning individual student learning activities into group learning activities that will benefit various ways of education by building each others support, connection, and communication. Unloading individual homework and loading in-class group work during a synchronous classroom setting will benefit all levels of learnings effectively. Assigning most of the classwork to be completed during class time has potential benefits since students are more likely to be engaged and focused individually and jointly by sharing and communicating subject matter effectively. We are discussing how to facilitate the group work effectively by creating group assignments for personalized classrooms, setting specific rules and class ethics to follow for the best learning practice, imposing a list of roles to each group member, introducing a list of ground rules related to justice, equity, and diversity inclusion (JDEI) within the group. The students performances, progress and effectiveness are analyzed and compared., Comment: 7 pages, 4 figures, 1 table

Test set contamination, wherein test data from a benchmark ends up in a newer model's training set, is a well-documented obstacle for fair LLM evaluation and can quickly render benchmarks obsolete. To mitigate this, many recent benchmarks crowdsource new prompts and evaluations from human or LLM judges; however, these can introduce significant biases, and break down when scoring hard questions. In this work, we introduce a new benchmark for LLMs designed to be immune to both test set contamination and the pitfalls of LLM judging and human crowdsourcing. We release LiveBench, the first benchmark that (1) contains frequently-updated questions from recent information sources, (2) scores answers automatically according to objective ground-truth values, and (3) contains a wide variety of challenging tasks, spanning math, coding, reasoning, language, instruction following, and data analysis. To achieve this, LiveBench contains questions that are based on recently-released math competitions, arXiv papers, news articles, and datasets, and it contains harder, contamination-free versions of tasks from previous benchmarks such as Big-Bench Hard, AMPS, and IFEval. We evaluate many prominent closed-source models, as well as dozens of open-source models ranging from 0.5B to 110B in size. LiveBench is difficult, with top models achieving below 65% accuracy. We release all questions, code, and model answers. Questions will be added and updated on a monthly basis, and we will release new tasks and harder versions of tasks over time so that LiveBench can distinguish between the capabilities of LLMs as they improve in the future. We welcome community engagement and collaboration for expanding the benchmark tasks and models.

Similar to reading the transistor state in classical computers, identifying the quantum bit (qubit) state is a fundamental operation to translate quantum information. However, identifying quantum state has been the slowest and most susceptible to errors operation on superconducting quantum processors. Most existing state discrimination algorithms have only been implemented and optimized "after the fact" - using offline data transferred from control circuits to host computers. Real-time state discrimination is not possible because a superconducting quantum state only survives for a few hundred us, which is much shorter than the communication delay between the readout circuit and the host computer (i.e., tens of ms). Mid-circuit measurement (MCM), where measurements are conducted on qubits at intermediate stages within a quantum circuit rather than solely at the end, represents an advanced technique for qubit reuse. For MCM necessitating single-shot readout, it is imperative to employ an in-situ technique for state discrimination with low latency and high accuracy. This paper introduces QubiCML, a field-programmable gate array (FPGA) based system for real-time state discrimination enabling MCM - the ability to measure the state at the control circuit before/without transferring data to a host computer. A multi-layer neural network has been designed and deployed on an FPGA to ensure accurate in-situ state discrimination. For the first time, ML-powered quantum state discrimination has been implemented on a radio frequency system-on-chip FPGA platform. The deployed lightweight network on the FPGA only takes 54 ns to complete each inference. We evaluated QubiCML's performance on superconducting quantum processors and obtained an average accuracy of 98.5% with only 500 ns readout. QubiCML has the potential to be the standard real-time state discrimination method for the quantum community.

Decomposing model activations into interpretable components is a key open problem in mechanistic interpretability. Sparse autoencoders (SAEs) are a popular method for decomposing the internal activations of trained transformers into sparse, interpretable features, and have been applied to MLP layers and the residual stream. In this work we train SAEs on attention layer outputs and show that also here SAEs find a sparse, interpretable decomposition. We demonstrate this on transformers from several model families and up to 2B parameters. We perform a qualitative study of the features computed by attention layers, and find multiple families: long-range context, short-range context and induction features. We qualitatively study the role of every head in GPT-2 Small, and estimate that at least 90% of the heads are polysemantic, i.e. have multiple unrelated roles. Further, we show that Sparse Autoencoders are a useful tool that enable researchers to explain model behavior in greater detail than prior work. For example, we explore the mystery of why models have so many seemingly redundant induction heads, use SAEs to motivate the hypothesis that some are long-prefix whereas others are short-prefix, and confirm this with more rigorous analysis. We use our SAEs to analyze the computation performed by the Indirect Object Identification circuit (Wang et al.), validating that the SAEs find causally meaningful intermediate variables, and deepening our understanding of the semantics of the circuit. We open-source the trained SAEs and a tool for exploring arbitrary prompts through the lens of Attention Output SAEs.

We revisit the efficacy of several practical methods for approximate machine unlearning developed for large-scale deep learning. In addition to complying with data deletion requests, one often-cited potential application for unlearning methods is to remove the effects of training on poisoned data. We experimentally demonstrate that, while existing unlearning methods have been demonstrated to be effective in a number of evaluation settings (e.g., alleviating membership inference attacks), they fail to remove the effects of data poisoning, across a variety of types of poisoning attacks (indiscriminate, targeted, and a newly-introduced Gaussian poisoning attack) and models (image classifiers and LLMs); even when granted a relatively large compute budget. In order to precisely characterize unlearning efficacy, we introduce new evaluation metrics for unlearning based on data poisoning. Our results suggest that a broader perspective, including a wider variety of evaluations, is required to avoid a false sense of confidence in machine unlearning procedures for deep learning without provable guarantees. Moreover, while unlearning methods show some signs of being useful to efficiently remove poisoned datapoints without having to retrain, our work suggests that these methods are not yet "ready for prime time", and currently provide limited benefit over retraining.

Despite their widespread use, the mechanisms by which large language models (LLMs) represent and regulate uncertainty in next-token predictions remain largely unexplored. This study investigates two critical components believed to influence this uncertainty: the recently discovered entropy neurons and a new set of components that we term token frequency neurons. Entropy neurons are characterized by an unusually high weight norm and influence the final layer normalization (LayerNorm) scale to effectively scale down the logits. Our work shows that entropy neurons operate by writing onto an unembedding null space, allowing them to impact the residual stream norm with minimal direct effect on the logits themselves. We observe the presence of entropy neurons across a range of models, up to 7 billion parameters. On the other hand, token frequency neurons, which we discover and describe here for the first time, boost or suppress each token's logit proportionally to its log frequency, thereby shifting the output distribution towards or away from the unigram distribution. Finally, we present a detailed case study where entropy neurons actively manage confidence in the setting of induction, i.e. detecting and continuing repeated subsequences., Comment: NeurIPS 2024

Large language models (LLMs) and vision-language models (VLMs) have demonstrated remarkable performance across a wide range of tasks and domains. Despite this promise, spatial understanding and reasoning -- a fundamental component of human cognition -- remains under-explored. We propose SpatialEval, a novel benchmark that covers diverse aspects of spatial reasoning such as relationship understanding, navigation, and counting. We conduct a comprehensive evaluation of competitive language and vision-language models. Our findings reveal several counter-intuitive insights that have been overlooked in the literature: (1) Spatial reasoning poses significant challenges where competitive models can fall behind random guessing; (2) Despite additional visual input, VLMs often under-perform compared to their LLM counterparts; (3) When both textual and visual information is available, multi-modal language models become less reliant on visual information if sufficient textual clues are provided. Additionally, we demonstrate that leveraging redundancy between vision and text can significantly enhance model performance. We hope our study will inform the development of multimodal models to improve spatial intelligence and further close the gap with human intelligence., Comment: Accepted to NeurIPS 2024

It is becoming cheaper to launch disinformation operations at scale using AI-generated content, in particular 'deepfake' technology. We have observed instances of deepfakes in political campaigns, where generated content is employed to both bolster the credibility of certain narratives (reinforcing outcomes) and manipulate public perception to the detriment of targeted candidates or causes (adversarial outcomes). We discuss the threats from deepfakes in politics, highlight model specifications underlying different types of deepfake generation methods, and contribute an accessible evaluation of the efficacy of existing detection methods. We provide this as a summary for lawmakers and civil society actors to understand how the technology may be applied in light of existing policies regulating its use. We highlight the limitations of existing detection mechanisms and discuss the areas where policies and regulations are required to address the challenges of deepfakes., Comment: Accepted at the AAAI 2024 conference, AI for Credible Elections Workshop-AI4CE 2024

Tuberculosis persists as a global health crisis, especially in resource-limited populations and remote regions, with more than 10 million individuals newly infected annually. It stands as a stark symbol of inequity in public health. Tuberculosis impacts roughly a quarter of the global populace, with the majority of cases concentrated in eight countries, accounting for two-thirds of all tuberculosis infections. Although a severe ailment, tuberculosis is both curable and manageable. However, early detection and screening of at-risk populations are imperative. Chest x-ray stands as the predominant imaging technique utilized in tuberculosis screening efforts. However, x-ray screening necessitates skilled radiologists, a resource often scarce, particularly in remote regions with limited resources. Consequently, there is a pressing need for artificial intelligence (AI)-powered systems to support clinicians and healthcare providers in swift screening. However, training a reliable AI model necessitates large-scale high-quality data, which can be difficult and costly to acquire. Inspired by these challenges, in this work, we introduce an explainable self-supervised self-train learning network tailored for tuberculosis case screening. The network achieves an outstanding overall accuracy of 98.14% and demonstrates high recall and precision rates of 95.72% and 99.44%, respectively, in identifying tuberculosis cases, effectively capturing clinically significant features., Comment: 9 pages, 3 figures

Scanning tunneling microscopy using a CO-functionalized tip is combined with simulations to explore the impact of the CO tilt angle on topographies of a single Cu atom and CO molecule adsorbed on Cu(111). Images of the Cu atom acquired with varying tip tilt angles and sample voltages are reproduced by the calculations. The agreement between measured and simulated data allows to unveil the tip-orbital composition of the tunneling current, which highlights the role of the different p-orbitals of the CO tip. Microscope data of adsorbed CO and their dependence on voltage and probe-surface distance are captured for the nontilted junction geometry and in the limit of weak tip-surface interaction assuming sufficiently large tip-sample separations., Comment: 13 pages, 12 figures, accepted for publication in Physical Review B

A key goal in mechanistic interpretability is circuit analysis: finding sparse subgraphs of models corresponding to specific behaviors or capabilities. However, MLP sublayers make fine-grained circuit analysis on transformer-based language models difficult. In particular, interpretable features -- such as those found by sparse autoencoders (SAEs) -- are typically linear combinations of extremely many neurons, each with its own nonlinearity to account for. Circuit analysis in this setting thus either yields intractably large circuits or fails to disentangle local and global behavior. To address this we explore transcoders, which seek to faithfully approximate a densely activating MLP layer with a wider, sparsely-activating MLP layer. We introduce a novel method for using transcoders to perform weights-based circuit analysis through MLP sublayers. The resulting circuits neatly factorize into input-dependent and input-invariant terms. We then successfully train transcoders on language models with 120M, 410M, and 1.4B parameters, and find them to perform at least on par with SAEs in terms of sparsity, faithfulness, and human-interpretability. Finally, we apply transcoders to reverse-engineer unknown circuits in the model, and we obtain novel insights regarding the "greater-than circuit" in GPT2-small. Our results suggest that transcoders can prove effective in decomposing model computations involving MLPs into interpretable circuits. Code is available at https://github.com/jacobdunefsky/transcoder_circuits/., Comment: 29 pages, 6 figures, 4 tables, 2 algorithms. NeurIPS 2024

Conversational large language models are fine-tuned for both instruction-following and safety, resulting in models that obey benign requests but refuse harmful ones. While this refusal behavior is widespread across chat models, its underlying mechanisms remain poorly understood. In this work, we show that refusal is mediated by a one-dimensional subspace, across 13 popular open-source chat models up to 72B parameters in size. Specifically, for each model, we find a single direction such that erasing this direction from the model's residual stream activations prevents it from refusing harmful instructions, while adding this direction elicits refusal on even harmless instructions. Leveraging this insight, we propose a novel white-box jailbreak method that surgically disables refusal with minimal effect on other capabilities. Finally, we mechanistically analyze how adversarial suffixes suppress propagation of the refusal-mediating direction. Our findings underscore the brittleness of current safety fine-tuning methods. More broadly, our work showcases how an understanding of model internals can be leveraged to develop practical methods for controlling model behavior.

Most public instruction finetuning datasets are relatively small compared to the closed source datasets used to train industry models. To study questions about finetuning at scale, such as curricula and learning rate cooldown schedules, there is a need for industrial-scale datasets. However, this scale necessitates a data generation process that is almost entirely automated. In this work, we study methods for generating large instruction datasets from a single prompt. With little human oversight, we get LLMs to write diverse sets of instruction examples ranging from simple completion tasks to complex multi-turn dialogs across a variety of subject areas. When finetuning a Llama-3 8B base model, our dataset meets or exceeds both WizardLM and Ultrachat on both knowledge-intensive leaderboard tasks as well as conversational evaluations. We release our dataset, the "generator" prompts that created it, and our finetuned model checkpoints., Comment: 9.5 pages, 6 Figures, and 3 tables in the main body. Dataset available at https://huggingface.co/datasets/tomg-group-umd/GenQA

Large language models can memorize and repeat their training data, causing privacy and copyright risks. To mitigate memorization, we introduce a subtle modification to the next-token training objective that we call the goldfish loss. During training, randomly sampled subsets of tokens are excluded from the loss computation. These dropped tokens are not memorized by the model, which prevents verbatim reproduction of a complete chain of tokens from the training set. We run extensive experiments training billion-scale Llama-2 models, both pre-trained and trained from scratch, and demonstrate significant reductions in extractable memorization with little to no impact on downstream benchmarks., Comment: 10 pages, 8 figures, and 1 table in the main body. Code available at https://github.com/ahans30/goldfish-loss and checkpoints at https://huggingface.co/collections/tomg-group-umd/goldfish-loss-mitigating-memorization-in-llms-66c175becb6aab07744f7272

We present the Keras domain packages KerasCV and KerasNLP, extensions of the Keras API for Computer Vision and Natural Language Processing workflows, capable of running on either JAX, TensorFlow, or PyTorch. These domain packages are designed to enable fast experimentation, with a focus on ease-of-use and performance. We adopt a modular, layered design: at the library's lowest level of abstraction, we provide building blocks for creating models and data preprocessing pipelines, and at the library's highest level of abstraction, we provide pretrained ``task" models for popular architectures such as Stable Diffusion, YOLOv8, GPT2, BERT, Mistral, CLIP, Gemma, T5, etc. Task models have built-in preprocessing, pretrained weights, and can be fine-tuned on raw inputs. To enable efficient training, we support XLA compilation for all models, and run all preprocessing via a compiled graph of TensorFlow operations using the tf.data API. The libraries are fully open-source (Apache 2.0 license) and available on GitHub., Comment: Submitted to Journal of Machine Learning Open Source Software

The poor performance of transformers on arithmetic tasks seems to stem in large part from their inability to keep track of the exact position of each digit inside of a large span of digits. We mend this problem by adding an embedding to each digit that encodes its position relative to the start of the number. In addition to the boost these embeddings provide on their own, we show that this fix enables architectural modifications such as input injection and recurrent layers to improve performance even further. With positions resolved, we can study the logical extrapolation ability of transformers. Can they solve arithmetic problems that are larger and more complex than those in their training data? We find that training on only 20 digit numbers with a single GPU for one day, we can reach state-of-the-art performance, achieving up to 99% accuracy on 100 digit addition problems. Finally, we show that these gains in numeracy also unlock improvements on other multi-step reasoning tasks including sorting and multiplication.

Current developments in autonomous off-road driving are steadily increasing performance through higher speeds and more challenging, unstructured environments. However, this operating regime subjects the vehicle to larger inertial effects, where consideration of higher-order states is necessary to avoid failures such as rollovers or excessive impact forces. Aggressive driving through Model Predictive Control (MPC) in these conditions requires dynamics models that accurately predict safety-critical information. This work aims to empirically quantify this aggressive operating regime and its effects on the performance of current models. We evaluate three dynamics models of varying complexity on two distinct off-road driving datasets: one simulated and the other real-world. By conditioning trajectory data on higher-order states, we show that model accuracy degrades with aggressiveness and simpler models degrade faster. These models are also validated across datasets, where accuracies over safety-critical states are reported and provide benchmarks for future work., Comment: Accepted to ICRA 2024 Workshop on Resilient Off-road Autonomy

Running deep neural networks for large medical images is a resource-hungry and time-consuming task with centralized computing. Outsourcing such medical image processing tasks to hybrid clouds has benefits, such as a significant reduction of execution time and monetary cost. However, due to privacy concerns, it is still challenging to process sensitive medical images over clouds, which would hinder their deployment in many real-world applications. To overcome this, we first formulate the overall optimization objectives of the privacy-preserving distributed system model, i.e., minimizing the amount of information about the private data learned by the adversaries throughout the process, reducing the maximum execution time and cost under the user budget constraint. We propose a novel privacy-preserving and cost-effective method called PriCE to solve this multi-objective optimization problem. We performed extensive simulation experiments for artifact detection tasks on medical images using an ensemble of five deep convolutional neural network inferences as the workflow task. Experimental results show that PriCE successfully splits a wide range of input gigapixel medical images with graph-coloring-based strategies, yielding desired output utility and lowering the privacy risk, makespan, and monetary cost under user's budget., Comment: Acccepted at Europar 2024

The performance of physical workers is significantly influenced by the quantity of their motions. However, monitoring and assessing these motions is challenging due to the complexities of motion sensing, tracking, and quantification. Recent advancements have utilized in-situ video analysis for real-time observation of worker behaviors, enabling data-driven quantification of motion amounts. Nevertheless, there are limitations to monitoring worker movements using video data. This paper introduces a novel framework based on computer vision to track and quantify the motion of workers' upper and lower limbs, issuing alerts when the motion reaches critical thresholds. Using joint position data from posture estimation, the framework employs Hotelling's T$^2$ statistic to quantify and monitor motion amounts, integrating computer vision tools to address challenges in automated worker training and enhance exploratory research in this field. We collected data of participants performing lifting and moving tasks with large boxes and small wooden cubes, to simulate macro and micro assembly tasks respectively. It was found that the correlation between workers' joint motion amount and the Hotelling's T$^2$ statistic was approximately 35% greater for micro tasks compared to macro tasks, highlighting the framework's ability to identify fine-grained motion differences. This study demonstrates the effectiveness of the proposed system in real-time applications across various industry settings. It provides a tool for enhancing worker safety and productivity through precision motion analysis and proactive ergonomic adjustments.

Quantum key distribution (QKD) has often been hailed as a reliable technology for secure communication in cyber-physical microgrids. Even though unauthorized key measurements are not possible in QKD, attempts to read them can disturb quantum states leading to mutations in the transmitted value. Further, inaccurate quantum keys can lead to erroneous decryption producing garbage values, destabilizing microgrid operation. QKD can also be vulnerable to node-level manipulations incorporating attack values into measurements before they are encrypted at the communication layer. To address these issues, this paper proposes a secure QKD protocol that can identify errors in keys and/or nodal measurements by observing violations in control dynamics. Additionally, the protocol uses a dynamic adjacency matrix-based formulation strategy enabling the affected nodes to reconstruct a trustworthy signal and replace it with the attacked signal in a multi-hop manner. This enables microgrids to perform nominal operations in the presence of adversaries who try to eavesdrop on the system causing an increase in the quantum bit error rate (QBER). We provide several case studies to showcase the robustness of the proposed strategy against eavesdroppers and node manipulations. The results demonstrate that it can resist unwanted observation and attack vectors that manipulate signals before encryption.

Disentangling model activations into meaningful features is a central problem in interpretability. However, the absence of ground-truth for these features in realistic scenarios makes validating recent approaches, such as sparse dictionary learning, elusive. To address this challenge, we propose a framework for evaluating feature dictionaries in the context of specific tasks, by comparing them against \emph{supervised} feature dictionaries. First, we demonstrate that supervised dictionaries achieve excellent approximation, control, and interpretability of model computations on the task. Second, we use the supervised dictionaries to develop and contextualize evaluations of unsupervised dictionaries along the same three axes. We apply this framework to the indirect object identification (IOI) task using GPT-2 Small, with sparse autoencoders (SAEs) trained on either the IOI or OpenWebText datasets. We find that these SAEs capture interpretable features for the IOI task, but they are less successful than supervised features in controlling the model. Finally, we observe two qualitative phenomena in SAE training: feature occlusion (where a causally relevant concept is robustly overshadowed by even slightly higher-magnitude ones in the learned features), and feature over-splitting (where binary features split into many smaller, less interpretable features). We hope that our framework will provide a useful step towards more objective and grounded evaluations of sparse dictionary learning methods.

Optical helicity quantifies the handedness of light, and plays a central role in the description of interactions between light and chiral matter. In free space, it is related to the duality symmetry of the electromagnetic field, a continuous symmetry encapsulating the invariance of Maxwell's equations under the interchange of electric and magnetic fields. However, in materials the situation is not so straightforward, as the free space transformation must be extended to encompass mixing of both the $\mathbf{E}$/$\mathbf{H}$ and $\mathbf{D}$/$\mathbf{B}$ field pairs. The simultaneous direct interchange of $\mathbf{E}$/$\mathbf{H}$ and of $\mathbf{D}$/$\mathbf{B}$ is incompatible with the presence of linear constitutive relations. In this work, we extend the duality transform in a way that resolves this incompatibility, and use this to define the optical helicity in a general medium, which may be dispersive, lossy, chiral or nonreciprocal. We find that the helicity density must contain an explicit contribution associated with the polarisation and magnetisation of the matter, and we show that the form of this matter contribution is independent of the details of the medium. We also show that the in-medium helicity can be naturally expressed in terms of the elementary quantised excitations of the system., Comment: 4 pages + supplemental. 2 figures. Comments welcome

Recent work has found that sparse autoencoders (SAEs) are an effective technique for unsupervised discovery of interpretable features in language models' (LMs) activations, by finding sparse, linear reconstructions of LM activations. We introduce the Gated Sparse Autoencoder (Gated SAE), which achieves a Pareto improvement over training with prevailing methods. In SAEs, the L1 penalty used to encourage sparsity introduces many undesirable biases, such as shrinkage -- systematic underestimation of feature activations. The key insight of Gated SAEs is to separate the functionality of (a) determining which directions to use and (b) estimating the magnitudes of those directions: this enables us to apply the L1 penalty only to the former, limiting the scope of undesirable side effects. Through training SAEs on LMs of up to 7B parameters we find that, in typical hyper-parameter ranges, Gated SAEs solve shrinkage, are similarly interpretable, and require half as many firing features to achieve comparable reconstruction fidelity., Comment: 15 main text pages, 22 appendix pages

Activation patching is a popular mechanistic interpretability technique, but has many subtleties regarding how it is applied and how one may interpret the results. We provide a summary of advice and best practices, based on our experience using this technique in practice. We include an overview of the different ways to apply activation patching and a discussion on how to interpret the results. We focus on what evidence patching experiments provide about circuits, and on the choice of metric and associated pitfalls., Comment: A tutorial on activation patching. 13 pages, 2 figures

Large language models (LLMs) has experienced exponential growth, they demonstrate remarkable performance across various tasks. Notwithstanding, contemporary research primarily centers on enhancing the size and quality of pretraining data, still utilizing the next token prediction task on autoregressive transformer model structure. The efficacy of this task in truly facilitating the model's comprehension of code logic remains questionable, we speculate that it still interprets code as mere text, while human emphasizes the underlying logical knowledge. In order to prove it, we introduce a new task, "Logically Equivalent Code Selection," which necessitates the selection of logically equivalent code from a candidate set, given a query code. Our experimental findings indicate that current LLMs underperform in this task, since they understand code by unordered bag of keywords. To ameliorate their performance, we propose an advanced pretraining task, "Next Token Prediction+". This task aims to modify the sentence embedding distribution of the LLM without sacrificing its generative capabilities. Our experimental results reveal that following this pretraining, both Code Llama and StarCoder, the prevalent code domain pretraining models, display significant improvements on our logically equivalent code selection task and the code completion task.