Welcome back to, TIME's new twice-weekly newsletter about AI. If you're reading this in your browser, why not subscribe to have the next one delivered straight to your inbox? The three most valuable private companies in the U.S. have big reputations: OpenAI, SpaceX, and Anthropic. But the fourth, Databricks, flies a little more under the radar. This company, which is currently raising funds at a valuation of $134 billion according to reports this week, is a quiet workhorse of the AI revolution.

Benchmarks play a significant role in how researchers and the public understand generative AI systems. However, the widespread use of benchmark scores to communicate about model capabilities has led to criticisms of validity, especially whether benchmarks test what they claim to test (i.e. construct validity) and whether benchmark evaluations are representative of how models are used in the wild (i.e. ecological validity). In this work we explore how to create an LLM benchmark that addresses these issues by taking a human-centered approach. We focus on designing a domain-oriented benchmark for journalism practitioners, drawing on insights from a workshop of 23 journalism professionals. Our workshop findings surface specific challenges that inform benchmark design opportunities, which we instantiate in a case study that addresses underlying criticisms and specific domain concerns. Through our findings and design case study, this work provides design guidance for developing benchmarks that are better tuned to specific domains.

The LLM unlearning aims to eliminate the influence of undesirable data without affecting causally unrelated information. This process typically involves using a forget set to remove target information, alongside a retain set to maintain non-target capabilities. While recent localization-based methods demonstrate promise in identifying important neurons to be unlearned, they fail to disentangle neurons responsible for forgetting undesirable knowledge or retaining essential skills, often treating them as a single entangled group. As a result, these methods apply uniform interventions, risking catastrophic over-forgetting or incomplete erasure of the target knowledge. To address this, we turn to circuit discovery, a mechanistic interpretability technique, and propose the Conflict-guided Localization for LLM Unlearning framEwork (CLUE). This framework identifies the forget and retain circuit composed of important neurons, and then the circuits are transformed into conjunctive normal forms (CNF). The assignment of each neuron in the CNF satisfiability solution reveals whether it should be forgotten or retained. We then provide targeted fine-tuning strategies for different categories of neurons. Extensive experiments demonstrate that, compared to existing localization methods, CLUE achieves superior forget efficacy and retain utility through precise neural localization.

The episode length is 150 for RoboBin and RoboKitchen and 1000 for RoboY oga. We show all goals in Figure A.1. For both W alker and Quadruped, the success criterion is based on the largest violation across all joints. Global position is not taken into account for the success computation. For reaching task, the success is based on placing the arm in the correct position within 10 cm.

Mixture-of-experts (MoE) architectures enable scaling large language models (LLMs) to vast parameter counts without a proportional rise in computational costs. However, the significant memory demands of large MoE models hinder their deployment across various computational environments, from cloud servers to consumer devices. This study first demonstrates pronounced task-specific specialization in expert activation patterns within MoE layers. Building on this, we introduce PreMoe, a novel framework that enables efficient deployment of massive MoE models in memory-constrained environments. PreMoe features two main components: probabilistic expert pruning (PEP) and task-adaptive expert retrieval (TAER). PEP employs a new metric, the task-conditioned expected selection score (TCESS), derived from router logits to quantify expert importance for specific tasks, thereby identifying a minimal set of critical experts. TAER leverages these task-specific expert importance profiles for efficient inference. It pre-computes and stores compact expert patterns for diverse tasks. When a user query is received, TAER rapidly identifies the most relevant stored task pattern and reconstructs the model by loading only the small subset of experts crucial for that task. This approach dramatically reduces the memory footprint across all deployment scenarios. DeepSeek-R1 671B maintains 97.2\% accuracy on MATH500 when pruned to 8/128 configuration (50\% expert reduction), and still achieves 72.0\% with aggressive 8/32 pruning (87.5\% expert reduction). Pangu-Ultra-MoE 718B achieves 97.15\% on MATH500 and 81.3\% on AIME24 with 8/128 pruning, while even more aggressive pruning to 4/64 (390GB memory) preserves 96.95\% accuracy on MATH500. We make our code publicly available at https://github.com/JarvisPei/PreMoe.

Large language models demonstrate impressive proficiency in language understanding and generation. Nonetheless, training these models from scratch, even the least complex billion-parameter variant demands significant computational resources rendering it economically impractical for many organizations. With large language models functioning as general-purpose task solvers, this paper investigates their task-specific fine-tuning. We employ task-specific datasets and prompts to fine-tune two pruned LLaMA models having 5 billion and 4 billion parameters. This process utilizes the pre-trained weights and focuses on a subset of weights using the LoRA method. One challenge in fine-tuning the LLaMA model is crafting a precise prompt tailored to the specific task. To address this, we propose a novel approach to fine-tune the LLaMA model under two primary constraints: task specificity and prompt effectiveness. Our approach, Tailored LLaMA initially employs structural pruning to reduce the model sizes from 7B to 5B and 4B parameters. Subsequently, it applies a carefully designed prompt specific to the task and utilizes the LoRA method to accelerate the fine-tuning process. Moreover, fine-tuning a model pruned by 50\% for less than one hour restores the mean accuracy of classification tasks to 95.68\% at a 20\% compression ratio and to 86.54\% at a 50\% compression ratio through few-shot learning with 50 shots. Our validation of Tailored LLaMA on these two pruned variants demonstrates that even when compressed to 50\%, the models maintain over 65\% of the baseline model accuracy in few-shot classification and generation tasks. These findings highlight the efficacy of our tailored approach in maintaining high performance with significantly reduced model sizes.

This study proposes a large language model optimization method based on the improved LoRA fine-tuning algorithm, aiming to improve the accuracy and computational efficiency of the model in natural language processing tasks. We fine-tune the large language model through a low-rank adaptation strategy, which significantly reduces the consumption of computing resources while maintaining the powerful capabilities of the pre-trained model. The experiment uses the QQP task as the evaluation scenario. The results show that the improved LoRA algorithm shows significant improvements in accuracy, F1 score, and MCC compared with traditional models such as BERT, Roberta, T5, and GPT-4. In particular, in terms of F1 score and MCC, our model shows stronger robustness and discrimination ability, which proves the potential of the improved LoRA algorithm in fine-tuning large-scale pre-trained models. In addition, this paper also discusses the application prospects of the improved LoRA algorithm in other natural language processing tasks, emphasizing its advantages in multi-task learning and scenarios with limited computing resources. Future research can further optimize the LoRA fine-tuning strategy and expand its application in larger-scale pre-trained models to improve the generalization ability and task adaptability of the model.

Large Language Models (LLMs) have shown impressive performance across a wide range of tasks. However, the size of LLMs is steadily increasing, hindering their application on computationally constrained environments. On the other hand, despite their general capabilities, there are many situations where only one specific task is performed, rendering all other capabilities unnecessary and wasteful. This leads us to the following question: Is it possible to extract the minimal subset from an LLM that is able to perform a specific task in a faster, standalone manner? Recent works on Mechanistic Interpretability (MI) have shown that specific tasks are performed by a localized subset of components, or circuit. However, current techniques used to identify the circuit cannot be used to extract it for its standalone usage. In this work, we propose a novel approach to automatically extract the subset of the LLM that properly performs a targeted task requiring no additional training and a small amount of data samples. We evaluate our approach on different tasks and show that the resulting models are (i) considerably smaller, reducing the number of parameters up to 82.77% and (ii) more interpretable, as they focus on the circuit that is used to carry out the specific task, and can therefore be understood using MI techniques.

In this study, we introduce FEET, a standardized protocol designed to guide the development and benchmarking of foundation models. While numerous benchmark datasets exist for evaluating these models, we propose a structured evaluation protocol across three distinct scenarios to gain a comprehensive understanding of their practical performance. We define three primary use cases: frozen embeddings, few-shot embeddings, and fully fine-tuned embeddings. Each scenario is detailed and illustrated through two case studies: one in sentiment analysis and another in the medical domain, demonstrating how these evaluations provide a thorough assessment of foundation models' effectiveness in research applications. We recommend this protocol as a standard for future research aimed at advancing representation learning models.