While the activations of neurons in deep neural networks usually do not have a simple human-understandable interpretation, sparse autoencoders (SAEs) can be used to transform these activations into a higher-dimensional latent space which may be more easily interpretable. However, these SAEs can have millions of distinct latent features, making it infeasible for humans to manually interpret each one. In this work, we build an open-source automated pipeline to generate and evaluate natural language explanations for SAE features using LLMs. We test our framework on SAEs of varying sizes, activation functions, and losses, trained on two different open-weight LLMs. We introduce five new techniques to score the quality of explanations that are cheaper to run than the previous state of the art. One of these techniques, intervention scoring, evaluates the interpretability of the effects of intervening on a feature, which we find explains features that are not recalled by existing methods. We propose guidelines for generating better explanations that remain valid for a broader set of activating contexts, and discuss pitfalls with existing scoring techniques. We use our explanations to measure the semantic similarity of independently trained SAEs, and find that SAEs trained on nearby layers of the residual stream are highly similar. Our large-scale analysis confirms that SAE latents are indeed much more interpretable than neurons, even when neurons are sparsified using top-$k$ postprocessing. Our code is available at https://github.com/EleutherAI/sae-auto-interp, and our explanations are available at https://huggingface.co/datasets/EleutherAI/auto_interp_explanations.

Methods for knowledge editing and unlearning in large language models seek to edit or remove undesirable knowledge or capabilities without compromising general language modeling performance. This work investigates how mechanistic interpretability -- which, in part, aims to identify model components (circuits) associated to specific interpretable mechanisms that make up a model capability -- can improve the precision and effectiveness of editing and unlearning. We find a stark difference in unlearning and edit robustness when training components localized by different methods. We highlight an important distinction between methods that localize components based primarily on preserving outputs, and those finding high level mechanisms with predictable intermediate states. In particular, localizing edits/unlearning to components associated with the lookup-table mechanism for factual recall 1) leads to more robust edits/unlearning across different input/output formats, and 2) resists attempts to relearn the unwanted information, while also reducing unintended side effects compared to baselines, on both a sports facts dataset and the CounterFact dataset across multiple models. We also find that certain localized edits disrupt the latent knowledge in the model more than any other baselines, making unlearning more robust to various attacks.

Large Language Models (LLMs) are increasingly embedded into software products across diverse industries, enhancing user experiences, but at the same time introducing numerous challenges for developers. Unique characteristics of LLMs force developers, who are accustomed to traditional software development and evaluation, out of their comfort zones as the LLM components shatter standard assumptions about software systems. This study explores the emerging solutions that software developers are adopting to navigate the encountered challenges. Leveraging a mixed-method research, including 26 interviews and a survey with 332 responses, the study identifies 19 emerging solutions regarding quality assurance that practitioners across several product teams at Microsoft are exploring. The findings provide valuable insights that can guide the development and evaluation of LLM-based products more broadly in the face of these challenges.

The development of vision-language and generative models has significantly advanced text-guided image editing, which seeks the \textit{preservation} of core elements in the source image while implementing \textit{modifications} based on the target text. However, existing metrics have a \textbf{context-blindness} problem, indiscriminately applying the same evaluation criteria on completely different pairs of source image and target text, biasing towards either modification or preservation. Directional CLIP similarity, the only metric that considers both source image and target text, is also biased towards modification aspects and attends to irrelevant editing regions of the image. We propose \texttt{AugCLIP}, a \textbf{context-aware} metric that adaptively coordinates preservation and modification aspects, depending on the specific context of a given source image and target text. This is done by deriving the CLIP representation of an ideally edited image, that preserves the source image with necessary modifications to align with target text. More specifically, using a multi-modal large language model, \texttt{AugCLIP} augments the textual descriptions of the source and target, then calculates a modification vector through a hyperplane that separates source and target attributes in CLIP space. Extensive experiments on five benchmark datasets, encompassing a diverse range of editing scenarios, show that \texttt{AugCLIP} aligns remarkably well with human evaluation standards, outperforming existing metrics. The code will be open-sourced for community use.

Long-context language models (LCLMs), characterized by their extensive context window, are becoming increasingly popular. However, despite they are nearly perfect at standard long-context retrieval, we find they are actually not good at all of them. Specifically, we identify 2 basic cases, "multi-matching retrieval," and "logic-based retrieval", which LLMs struggle to solve under normal settings. Moreover, we find these cases can only be well addressed by specific CoT prompting, with enough reasoning steps. This finding reminds the developers and users of LCLMs that relying on LCLMs to directly perform even basic retrieval tasks may be unreliable, rather, a sufficiently long reasoning process is necessary.

The application of large language models (LLMs) in domain-specific contexts, including finance, has expanded rapidly. Domain-specific LLMs are typically evaluated based on their performance in various downstream tasks relevant to the domain. In this work, we present a detailed analysis of fine-tuning LLMs for such tasks. Somewhat counterintuitively, we find that in domain-specific cases, fine-tuning exclusively on the target task is not always the most effective strategy. Instead, multi-task finetuning - where models are trained on a cocktail of related tasks - can significantly enhance performance. We demonstrate how this approach enables a small model, such as Phi-3-Mini, to achieve state-of-the-art results, even surpassing the much larger GPT-4-o model on financial benchmarks. Our study involves a large-scale experiment, conducting over 200 training experiments using several widely adopted LLMs as baselines, and empirically confirms the benefits of multi-task fine-tuning. Additionally, we explore the use of general instruction data as a form of regularization, suggesting that it helps minimize performance degradation. We also investigate the inclusion of mathematical data, finding improvements in numerical reasoning that transfer effectively to financial tasks. Finally, we note that while fine-tuning for downstream tasks leads to targeted improvements in task performance, it does not necessarily result in broader gains in domain knowledge or complex domain reasoning abilities.

We present a novel autonomous driving framework, DualAD, designed to imitate human reasoning during driving. DualAD comprises two layers: a rule-based motion planner at the bottom layer that handles routine driving tasks requiring minimal reasoning, and an upper layer featuring a rule-based text encoder that converts driving scenarios from absolute states into text description. This text is then processed by a large language model (LLM) to make driving decisions. The upper layer intervenes in the bottom layer's decisions when potential danger is detected, mimicking human reasoning in critical situations. Closed-loop experiments demonstrate that DualAD, using a zero-shot pre-trained model, significantly outperforms rule-based motion planners that lack reasoning abilities. Our experiments also highlight the effectiveness of the text encoder, which considerably enhances the model's scenario understanding. Additionally, the integrated DualAD model improves with stronger LLMs, indicating the framework's potential for further enhancement. Code and benchmarks are available at github.com/TUM-AVS/DualAD.

Addressing the disparity between forecasts and actual results can enable individuals to expand their thought processes and stimulate self-reflection, thus promoting accurate planning. In this research, we present **PreAct**, an agent framework that integrates **pre**diction, **rea**soning, and **act**ion. By utilizing the information derived from predictions, the large language model (LLM) agent can provide a wider range and more strategically focused reasoning. This leads to more efficient actions that aid the agent in accomplishing intricate tasks. Our experimental results show that PreAct surpasses the ReAct method in completing complex tasks and that PreAct's performance can be further improved when paired with other memory or selection strategy techniques. We presented the model with varying quantities of historical predictions and discovered that these predictions consistently enhance LLM planning.The variances in single-step reasoning between PreAct and ReAct indicate that PreAct indeed has benefits in terms of diversity and strategic orientation over ReAct.

Recent work in language modeling has raised the possibility of self-improvement, where a language models evaluates and refines its own generations to achieve higher performance without external feedback. It is impossible for this self-improvement to create information that is not already in the model, so why should we expect that this will lead to improved capabilities? We offer a new perspective on the capabilities of self-improvement through a lens we refer to as sharpening. Motivated by the observation that language models are often better at verifying response quality than they are at generating correct responses, we formalize self-improvement as using the model itself as a verifier during post-training in order to ``sharpen'' the model to one placing large mass on high-quality sequences, thereby amortizing the expensive inference-time computation of generating good sequences. We begin by introducing a new statistical framework for sharpening in which the learner aims to sharpen a pre-trained base policy via sample access, and establish fundamental limits. Then we analyze two natural families of self-improvement algorithms based on SFT and RLHF. We find that (i) the SFT-based approach is minimax optimal whenever the initial model has sufficient coverage, but (ii) the RLHF-based approach can improve over SFT-based self-improvement by leveraging online exploration, bypassing the need for coverage. Finally, we empirically validate the sharpening mechanism via inference-time and amortization experiments. We view these findings as a starting point toward a foundational understanding that can guide the design and evaluation of self-improvement algorithms.

Large language models have emerged as powerful instruments that can automate reasoning process in a host of domains ranging from Olympiad level math problem solving [1] to code generation [2] and financial planning [3]. LLMs have seen widespread adoption among people to automate and refine tasks involving decision making, logical thinking and critical reasoning. However, despite such promising results on numerous benchmarks, LLMs still possess surprising knowledge gaps caused due to a host of reasons varying from engineering heuristics such as tokenization [4] to limitations in the training data itself, such as data bias or a lack of up-to-date information. Furthermore, due to being trained on extensive data accumulated and refined by humans over the years, these LLMs present grounds for inherent cognitive bias as characterized in humans [5]. While there have been promising demonstrations of GPT-based models on tasks involving creativity [6] and critical thinking [7], slight variations in input prompt requests can lead to vastly different output responses[8]. Furthermore, due to being next token predictors scaled to a huge text, the world models formed by language models possess knowledge gaps unapparent on the surface level.