KV cache pruning has emerged as a promising technique for reducing memory and computation costs in long-context auto-regressive generation. Existing methods for vision-language models (VLMs) typically rely on self-attention scores from large language models (LLMs) to identify and prune irrelevant tokens. However, these approaches overlook the inherent distributional discrepancies between modalities, often leading to inaccurate token importance estimation and the over-pruning of critical visual tokens. To address this, we propose decomposing attention scores into intra-modality attention (within the same modality) and inter-modality attention (across modalities), enabling more precise KV cache pruning by independently managing these distinct attention types. Additionally, we introduce an n-softmax function to counteract distribution shifts caused by pruning, preserving the original smoothness of attention scores and ensuring stable performance. Our final training-free method, \textbf{C}ross-\textbf{S}elf \textbf{P}runing (CSP), achieves competitive performance compared to models with full KV caches while significantly outperforming previous pruning methods. Extensive evaluations on MileBench, a benchmark encompassing 29 multimodal datasets, demonstrate CSP's effectiveness, achieving up to a 41\% performance improvement on challenging tasks like conversational embodied dialogue while reducing the KV cache budget by 13.6\%. The code is available at https://github.com/TerryPei/CSP

The recent advancement of large foundation models (FMs) has increased the demand for fine-tuning these models on large-scale and cross-domain datasets. To address this, federated fine-tuning has emerged as a solution, allowing models to be fine-tuned on distributed datasets across multiple devices while ensuring data privacy. However, the substantial parameter size of FMs and the multi-round communication required by traditional federated fine-tuning algorithms result in prohibitively high communication costs, challenging the practicality of federated fine-tuning. In this paper, we are the first to reveal, both theoretically and empirically, that the traditional multi-round aggregation algorithms may not be necessary for federated fine-tuning large FMs. Our experiments reveal that a single round of communication (i.e., one-shot federated fine-tuning) yields a global model performance comparable to that achieved through multiple rounds of communication. Through rigorous mathematical and empirical analyses, we demonstrate that large FMs, due to their extensive parameter sizes and pre-training on general tasks, achieve significantly lower training loss in one-shot federated fine-tuning compared to smaller models. Our extensive experiments show that one-shot federated fine-tuning not only reduces communication costs but also enables asynchronous aggregation, enhances privacy, and maintains performance consistency with multi-round federated fine-tuning for models larger than 1 billion parameters, on text generation and text-to-image generation tasks. Our findings have the potential to revolutionize federated fine-tuning in practice, enhancing efficiency, reducing costs, and expanding accessibility for large-scale models. This breakthrough paves the way for broader adoption and application of federated fine-tuning across various domains.

Radiology report generation (RRG) has shown great potential in assisting radiologists by automating the labor-intensive task of report writing. While recent advancements have improved the quality and coherence of generated reports, ensuring their factual correctness remains a critical challenge. Although generative medical Vision Large Language Models (VLLMs) have been proposed to address this issue, these models are prone to hallucinations and can produce inaccurate diagnostic information. To address these concerns, we introduce a novel Semantic Consistency-Based Uncertainty Quantification framework that provides both report-level and sentence-level uncertainties. Unlike existing approaches, our method does not require modifications to the underlying model or access to its inner state, such as output token logits, thus serving as a plug-and-play module that can be seamlessly integrated with state-of-the-art models. Extensive experiments demonstrate the efficacy of our method in detecting hallucinations and enhancing the factual accuracy of automatically generated radiology reports. By abstaining from high-uncertainty reports, our approach improves factuality scores by $10$%, achieved by rejecting $20$% of reports using the Radialog model on the MIMIC-CXR dataset. Furthermore, sentence-level uncertainty flags the lowest-precision sentence in each report with an $82.9$% success rate.

Tools for analyzing character portrayal in fiction are valuable for writers and literary scholars in developing and interpreting compelling stories. Existing tools, such as visualization tools for analyzing fictional characters, primarily rely on explicit textual indicators of character attributes. However, portrayal is often implicit, revealed through actions and behaviors rather than explicit statements. We address this gap by leveraging large language models (LLMs) to uncover implicit character portrayals. We start by generating a dataset for this task with greater cross-topic similarity, lexical diversity, and narrative lengths than existing narrative text corpora such as TinyStories and WritingPrompts. We then introduce LIIPA (LLMs for Inferring Implicit Portrayal for Character Analysis), a framework for prompting LLMs to uncover character portrayals. LIIPA can be configured to use various types of intermediate computation (character attribute word lists, chain-of-thought) to infer how fictional characters are portrayed in the source text. We find that LIIPA outperforms existing approaches, and is more robust to increasing character counts (number of unique persons depicted) due to its ability to utilize full narrative context. Lastly, we investigate the sensitivity of portrayal estimates to character demographics, identifying a fairness-accuracy tradeoff among methods in our LIIPA framework -- a phenomenon familiar within the algorithmic fairness literature. Despite this tradeoff, all LIIPA variants consistently outperform non-LLM baselines in both fairness and accuracy. Our work demonstrates the potential benefits of using LLMs to analyze complex characters and to better understand how implicit portrayal biases may manifest in narrative texts.

Large Language Models (LLMs) have demonstrated excellent capabilities in composing various modules together to create programs that can perform complex reasoning tasks on images. In this paper, we propose TANGO, an approach that extends the program composition via LLMs already observed for images, aiming to integrate those capabilities into embodied agents capable of observing and acting in the world. Specifically, by employing a simple PointGoal Navigation model combined with a memory-based exploration policy as a foundational primitive for guiding an agent through the world, we show how a single model can address diverse tasks without additional training. We task an LLM with composing the provided primitives to solve a specific task, using only a few in-context examples in the prompt. We evaluate our approach on three key Embodied AI tasks: Open-Set ObjectGoal Navigation, Multi-Modal Lifelong Navigation, and Open Embodied Question Answering, achieving state-of-the-art results without any specific fine-tuning in challenging zero-shot scenarios.

We present Quasar-1, a novel architecture that introduces temperature-guided reasoning to large language models through the Token Temperature Mechanism (TTM) and Guided Sequence of Thought (GSoT). Our approach leverages the concept of hot and cold tokens, where hot tokens are prioritized for their contextual relevance, while cold tokens provide supplementary information. This dynamic modulation of token importance enables the model to achieve superior logical reasoning capabilities compared to traditional chain-of-thought approaches. Through rigorous mathematical analysis, we prove that our temperature-guided attention mechanism converges to optimal reasoning paths with exponential guarantees. Empirical results show significant improvements in reasoning accuracy and computational efficiency across a wide range of tasks, making advanced AI reasoning accessible to a broader range of applications.

Language model post-training is applied to refine behaviors and unlock new skills across a wide range of recent language models, but open recipes for applying these techniques lag behind proprietary ones. The underlying training data and recipes for post-training are simultaneously the most important pieces of the puzzle and the portion with the least transparency. To bridge this gap, we introduce Tulu 3, a family of fully-open state-of-the-art post-trained models, alongside its data, code, and training recipes, serving as a comprehensive guide for modern post-training techniques. Tulu 3, which builds on Llama 3.1 base models, achieves results surpassing the instruct versions of Llama 3.1, Qwen 2.5, Mistral, and even closed models such as GPT-4o-mini and Claude 3.5-Haiku. The training algorithms for our models include supervised finetuning (SFT), Direct Preference Optimization (DPO), and a novel method we call Reinforcement Learning with Verifiable Rewards (RLVR). With Tulu 3, we introduce a multi-task evaluation scheme for post-training recipes with development and unseen evaluations, standard benchmark implementations, and substantial decontamination of existing open datasets on said benchmarks. We conclude with analysis and discussion of training methods that did not reliably improve performance. In addition to the Tulu 3 model weights and demo, we release the complete recipe -- including datasets for diverse core skills, a robust toolkit for data curation and evaluation, the training code and infrastructure, and, most importantly, a detailed report for reproducing and further adapting the Tulu 3 approach to more domains.

Large language models (LLMs) are increasingly used in applications where LLM inputs may span many different tasks. Recent work has found that the choice of LLM is consequential, and different LLMs may be good for different input samples. Prior approaches have thus explored how engineers might select an LLM to use for each sample (i.e. routing). While existing routing methods mostly require training auxiliary models on human-annotated data, our work explores whether it is possible to perform unsupervised routing. We propose Smoothie, a weak supervision-inspired routing approach that requires no labeled data. Given a set of outputs from different LLMs, Smoothie constructs a latent variable graphical model over embedding representations of observable LLM outputs and unknown "true" outputs. Using this graphical model, we estimate sample-dependent quality scores for each LLM, and route each sample to the LLM with the highest corresponding score. We find that Smoothie's LLM quality-scores correlate with ground-truth model quality (correctly identifying the optimal model on 9/14 tasks), and that Smoothie outperforms baselines for routing by up to 10 points accuracy.

With the recent remarkable advancement of large language models (LLMs), there has been a growing interest in utilizing them in the domains with highly sensitive data that lies outside their training data. For this purpose, retrieval augmented generation (RAG) is particularly effective -- it assists LLMs by directly providing relevant information from the external knowledge sources. However, without extra privacy safeguards, RAG outputs risk leaking sensitive information from the external data source. In this work, we explore RAG under differential privacy (DP), a formal guarantee of data privacy. The main challenge with differentially private RAG is how to generate long accurate answers within a moderate privacy budget. We address this by proposing an algorithm that smartly spends privacy budget only for the tokens that require the sensitive information and uses the non-private LLM for other tokens. Our extensive empirical evaluations reveal that our algorithm outperforms the non-RAG baseline under a reasonable privacy budget of $\epsilon\approx 10$ across different models and datasets.

Large language models (LLMs) have been shown to be effective on tabular prediction tasks in the low-data regime, leveraging their internal knowledge and ability to learn from instructions and examples. However, LLMs can fail to generate predictions that satisfy group fairness, that is, produce equitable outcomes across groups. Critically, conventional debiasing approaches for natural language tasks do not directly translate to mitigating group unfairness in tabular settings. In this work, we systematically investigate four empirical approaches to improve group fairness of LLM predictions on tabular datasets, including fair prompt optimization, soft prompt tuning, strategic selection of few-shot examples, and self-refining predictions via chain-of-thought reasoning. Through experiments on four tabular datasets using both open-source and proprietary LLMs, we show the effectiveness of these methods in enhancing demographic parity while maintaining high overall performance. Our analysis provides actionable insights for practitioners in selecting the most suitable approach based on their specific requirements and constraints.