To control the behavior of language models, steering methods attempt to ensure that outputs of the model satisfy specific pre-defined properties. Adding steering vectors to the model is a promising method of model control that is easier than finetuning, and may be more robust than prompting. However, it can be difficult to anticipate the effects of steering vectors produced by methods such as CAA [Panickssery et al., 2024] or the direct use of SAE latents [Templeton et al., 2024]. In our work, we address this issue by using SAEs to measure the effects of steering vectors, giving us a method that can be used to understand the causal effect of any steering vector intervention. We use this method for measuring causal effects to develop an improved steering method, SAE-Targeted Steering (SAE-TS), which finds steering vectors to target specific SAE features while minimizing unintended side effects. We show that overall, SAE-TS balances steering effects with coherence better than CAA and SAE feature steering, when evaluated on a range of tasks.

Although Large Language Models (LLMs) exhibit superior performance across diverse applications, their empirical deployment remains challenging due to their associated considerable model size and high inference costs. To mitigate these emerging challenges, model compression research such as post-training compression (Ashkboos et al., 2024; Ma et al., 2023) and compression-aware training (Alvarez & Salzmann, 2017; Lym et al., 2019; Liu et al., 2024, 2023c) has been extensively explored to reduce the computational resource demands of serving LLMs (Zhu et al., 2023). However, most existing methods either incur significant accuracy degradation compared to uncompressed models or have high training time. Additionally, their flexibility is often limited by a discrete set of compression formats (e.g., 2:4 sparsity, 3/4-bit quantization), making it challenging to meet the diverse capacity and efficiency requirements of different users. To overcome the above flexibility limitation, we re-formulate the model compression problem into the customized compensation problem: Given a compressed model, we aim to introduce residual low-rank paths to compensate for compression errors under customized requirements from users, such as tasks, compression ratios, etc. Rather than focusing solely on producing compressed models with minimal performance degradation, by incorporating these residual paths, the compensated model gains greater flexibility in adjusting overall capacity, without being constrained by specific compression formats.

Large Language Models (LLMs) excel at understanding the semantic relationships between queries and documents, even with lengthy and complex long-tail queries. These queries are challenging for feedback-based rankings due to sparse user engagement and limited feedback, making LLMs' ranking ability highly valuable. However, the large size and slow inference of LLMs necessitate the development of smaller, more efficient models (sLLMs). Recently, integrating ranking label generation into distillation techniques has become crucial, but existing methods underutilize LLMs' capabilities and are cumbersome. Our research, RRADistill: Re-Ranking Ability Distillation, propose an efficient label generation pipeline and novel sLLM training methods for both encoder and decoder models. We introduce an encoder-based method using a Term Control Layer to capture term matching signals and a decoder-based model with a ranking layer for enhanced understanding. A/B testing on a Korean-based search platform, validates the effectiveness of our approach in improving re-ranking for long-tail queries.

As our understanding of autism and ableism continues to increase, so does our understanding of ableist language towards autistic people. Such language poses a significant challenge in NLP research due to its subtle and context-dependent nature. Yet, detecting anti-autistic ableist language remains underexplored, with existing NLP tools often failing to capture its nuanced expressions. We present AUTALIC, the first benchmark dataset dedicated to the detection of anti-autistic ableist language in context, addressing a significant gap in the field. The dataset comprises 2,400 autism-related sentences collected from Reddit, accompanied by surrounding context, and is annotated by trained experts with backgrounds in neurodiversity. Our comprehensive evaluation reveals that current language models, including state-of-the-art LLMs, struggle to reliably identify anti-autistic ableism and align with human judgments, underscoring their limitations in this domain. We publicly release AUTALIC along with the individual annotations which serve as a valuable resource to researchers working on ableism, neurodiversity, and also studying disagreements in annotation tasks. This dataset serves as a crucial step towards developing more inclusive and context-aware NLP systems that better reflect diverse perspectives.

Contrastive language-image pre-training (CLIP) has shown remarkable generalization ability in image classification. However, CLIP sometimes encounters performance drops on downstream datasets during zero-shot inference. Test-time adaptation methods attempt to mitigate this by adjusting normalization layers or tuning context prompts with large batch sizes and extensive augmentations; yet, these methods are computationally intensive. This raises an important question: Is there a training-free approach that can efficiently address CLIP's performance drop in such cases? To explore this, we benchmark token condensation techniques, originally designed to enhance the efficiency of vision transformers, on CLIP zero-shot inference tasks. We observe that although token condensation may compromise in-domain accuracy, it surprisingly enhances CLIP's performance on certain cross-dataset benchmarks. This motivates two key inquiries: (1) Can token condensation serve as a "free-lunch" solution for CLIP zero-shot inference? (2) What criteria should guide condensation -- how can essential tokens be identified and redundant ones eliminated? To address these questions, we propose Token Condensation as Adaptation (TCA), a training-free adaptation method for CLIP by pruning class-irrelevant visual tokens while merging class-ambiguous tokens. As the first approach for CLIP's token efficiency, TCA demonstrates superior performance across cross-dataset tasks, achieving up to a 21.4\% improvement over the strongest baseline while reducing GFLOPs by 12.2\% to 48.9\%, with minimized hyperparameter dependency.

This paper presents an advanced mathematical problem-solving framework, LLaMA-Berry, for enhancing the mathematical reasoning ability of Large Language Models (LLMs). The framework combines Monte Carlo Tree Search (MCTS) with iterative Self-Refine to optimize the reasoning path and utilizes a pairwise reward model to evaluate different paths globally. By leveraging the self-critic and rewriting capabilities of LLMs, Self-Refine applied to MCTS (SR-MCTS) overcomes the inefficiencies and limitations of conventional step-wise and greedy search algorithms by fostering a more efficient exploration of solution spaces. Pairwise Preference Reward Model~(PPRM), inspired by Reinforcement Learning from Human Feedback (RLHF), is then used to model pairwise preferences between solutions, utilizing an Enhanced Borda Count (EBC) method to synthesize these preferences into a global ranking score to find better answers. This approach addresses the challenges of scoring variability and non-independent distributions in mathematical reasoning tasks. The framework has been tested on general and advanced benchmarks, showing superior performance in terms of search efficiency and problem-solving capability compared to existing methods like ToT and rStar, particularly in complex Olympiad-level benchmarks, including GPQA, AIME24 and AMC23.

Recent advances in deep learning have brought attention to the possibility of creating advanced, general AI systems that outperform humans across many tasks. However, if these systems pursue unintended goals, there could be catastrophic consequences. A key prerequisite for AI systems pursuing unintended goals is whether they will behave in a coherent and goal-directed manner in the first place, optimizing for some unknown goal; there exists significant research trying to evaluate systems for said behaviors. However, the most rigorous definitions of goal-directedness we currently have are difficult to compute in real-world settings. Drawing upon this previous literature, we explore policy goal-directedness within reinforcement learning (RL) environments. In our findings, we propose a different family of definitions of the goal-directedness of a policy that analyze whether it is well-modeled as near-optimal for many (sparse) reward functions. We operationalize this preliminary definition of goal-directedness and test it in toy Markov decision process (MDP) environments. Furthermore, we explore how goal-directedness could be measured in frontier large-language models (LLMs). Our contribution is a definition of goal-directedness that is simpler and more easily computable in order to approach the question of whether AI systems could pursue dangerous goals. We recommend further exploration of measuring coherence and goal-directedness, based on our findings.

Robots in 2024 are far more complex than their single-armed factory-working predecessors. Modern robots can run, jump, do the splits, and even hold down a basic conversation. At the same, despite decades of technical advancements and billions of dollars of investment, even the most advanced robot systems still struggle to do many everyday tasks humans take for granted like folding laundry or stacking blocks. Ironically, robots are quite bad at doing things we find easy. New advances in robot training take some inspiration from massively popular large language models like ChatGPT may change that… eventually.

Can AI replicate individual humans? An experiment simulating more than 1000 real people using the artificial intelligence model behind ChatGPT has successfully replicated their unique thoughts and personalities with high accuracy, sparking concerns about the ethics of mimicking individuals in this way. Joon Sung Park at Stanford University in California and his colleagues wanted to use generative AI tools to model individuals as a way of forecasting the impact of policy changes. Historically, this has been attempted using more simplistic rule-based statistical models, with limited success. How does ChatGPT work and do AI-powered chatbots "think" like…

What if your projects could talk? Not in a creepy, horror-movie way, but in such a way that they could point out that obvious thing you're missing, answer your questions, or help you brainstorm new ideas. That's what this new AI assistant is doing. With Merlin AI, you just upload your project files, and the chatbot transforms into an intelligent knowledge base. Picture it like ChatGPT and Google combined, saving you hours researching and digging through documents.