Rui Hou

Hugo Touvron, Louis Martin, Kevin Stone et al. · meta-ai

In this work, we develop and release Llama 2, a collection of pretrained and fine-tuned large language models (LLMs) ranging in scale from 7 billion to 70 billion parameters. Our fine-tuned LLMs, called Llama 2-Chat, are optimized for dialogue use cases. Our models outperform open-source chat models on most benchmarks we tested, and based on our human evaluations for helpfulness and safety, may be a suitable substitute for closed-source models. We provide a detailed description of our approach to fine-tuning and safety improvements of Llama 2-Chat in order to enable the community to build on our work and contribute to the responsible development of LLMs.

Aaron Grattafiori, Abhimanyu Dubey, Abhinav Jauhri et al. · allen-ai, berkeley

Modern artificial intelligence (AI) systems are powered by foundation models. This paper presents a new set of foundation models, called Llama 3. It is a herd of language models that natively support multilinguality, coding, reasoning, and tool usage. Our largest model is a dense Transformer with 405B parameters and a context window of up to 128K tokens. This paper presents an extensive empirical evaluation of Llama 3. We find that Llama 3 delivers comparable quality to leading language models such as GPT-4 on a plethora of tasks. We publicly release Llama 3, including pre-trained and post-trained versions of the 405B parameter language model and our Llama Guard 3 model for input and output safety. The paper also presents the results of experiments in which we integrate image, video, and speech capabilities into Llama 3 via a compositional approach. We observe this approach performs competitively with the state-of-the-art on image, video, and speech recognition tasks. The resulting models are not yet being broadly released as they are still under development.

Davis Liang, Hila Gonen, Yuning Mao et al. · uw

Large multilingual language models typically rely on a single vocabulary shared across 100+ languages. As these models have increased in parameter count and depth, vocabulary size has remained largely unchanged. This \textit{vocabulary bottleneck} limits the representational capabilities of multilingual models like XLM-R. In this paper, we introduce a new approach for scaling to very large multilingual vocabularies by de-emphasizing token sharing between languages with little lexical overlap and assigning vocabulary capacity to achieve sufficient coverage for each individual language. Tokenizations using our vocabulary are typically more semantically meaningful and shorter compared to XLM-R. Leveraging this improved vocabulary, we train XLM-V, a multilingual language model with a one million token vocabulary. XLM-V outperforms XLM-R on every task we tested on ranging from natural language inference (XNLI), question answering (MLQA, XQuAD, TyDiQA), to named entity recognition (WikiAnn). XLM-V is particularly effective on low-resource language tasks and outperforms XLM-R by 11.2% and 5.8% absolute on MasakhaNER and Americas NLI, respectively.

Wenhan Xiong, Jingyu Liu, Igor Molybog et al. · meta-ai

We present a series of long-context LLMs that support effective context windows of up to 32,768 tokens. Our model series are built through continual pretraining from Llama 2 with longer training sequences and on a dataset where long texts are upsampled. We perform extensive evaluation on language modeling, synthetic context probing tasks, and a wide range of research benchmarks. On research benchmarks, our models achieve consistent improvements on most regular tasks and significant improvements on long-context tasks over Llama 2. Notably, with a cost-effective instruction tuning procedure that does not require human-annotated long instruction data, the 70B variant can already surpass gpt-3.5-turbo-16k's overall performance on a suite of long-context tasks. Alongside these results, we provide an in-depth analysis on the individual components of our method. We delve into Llama's position encodings and discuss its limitation in modeling long dependencies. We also examine the impact of various design choices in the pretraining process, including the data mix and the training curriculum of sequence lengths -- our ablation experiments suggest that having abundant long texts in the pretrain dataset is not the key to achieving strong performance, and we empirically verify that long context continual pretraining is more efficient and similarly effective compared to pretraining from scratch with long sequences.

Zhuofeng Wu, Sinong Wang, Jiatao Gu et al. · meta-ai

Prompt tuning is a new, efficient NLP transfer learning paradigm that adds a task-specific prompt in each input instance during the model training stage. It freezes the pre-trained language model and only optimizes a few task-specific prompts. In this paper, we propose a conditional prompt generation method to generate prompts for each input instance, referred to as the Instance-Dependent Prompt Generation (IDPG). Unlike traditional prompt tuning methods that use a fixed prompt, IDPG introduces a lightweight and trainable component to generate prompts based on each input sentence. Extensive experiments on ten natural language understanding (NLU) tasks show that the proposed strategy consistently outperforms various prompt tuning baselines and is on par with other efficient transfer learning methods such as Compacter while tuning far fewer model parameters.

Ming Zhong, Aston Zhang, Xuewei Wang et al. · meta-ai

The development and evaluation of Large Language Models (LLMs) have largely focused on individual capabilities. However, this overlooks the intersection of multiple abilities across different types of expertise that are often required for real-world tasks, which we term cross capabilities. To systematically explore this concept, we first define seven core individual capabilities and then pair them to form seven common cross capabilities, each supported by a manually constructed taxonomy. Building on these definitions, we introduce CrossEval, a benchmark comprising 1,400 human-annotated prompts, with 100 prompts for each individual and cross capability. To ensure reliable evaluation, we involve expert annotators to assess 4,200 model responses, gathering 8,400 human ratings with detailed explanations to serve as reference examples. Our findings reveal that, in both static evaluations and attempts to enhance specific abilities, current LLMs consistently exhibit the "Law of the Weakest Link," where cross-capability performance is significantly constrained by the weakest component. Specifically, across 58 cross-capability scores from 17 models, 38 scores are lower than all individual capabilities, while 20 fall between strong and weak, but closer to the weaker ability. These results highlight the under-performance of LLMs in cross-capability tasks, making the identification and improvement of the weakest capabilities a critical priority for future research to optimize performance in complex, multi-dimensional scenarios.

Anastasia Razdaibiedina, Yuning Mao, Rui Hou et al.

We introduce Progressive Prompts - a simple and efficient approach for continual learning in language models. Our method allows forward transfer and resists catastrophic forgetting, without relying on data replay or a large number of task-specific parameters. Progressive Prompts learns a new soft prompt for each task and sequentially concatenates it with the previously learned prompts, while keeping the base model frozen. Experiments on standard continual learning benchmarks show that our approach outperforms state-of-the-art methods, with an improvement >20% in average test accuracy over the previous best-preforming method on T5 model. We also explore a more challenging continual learning setup with longer sequences of tasks and show that Progressive Prompts significantly outperforms prior methods.

Yuchen Liu, Natasha Ong, Kaiyan Peng et al. · meta-ai

We present Multiscale Multiview Vision Transformers (MMViT), which introduces multiscale feature maps and multiview encodings to transformer models. Our model encodes different views of the input signal and builds several channel-resolution feature stages to process the multiple views of the input at different resolutions in parallel. At each scale stage, we use a cross-attention block to fuse information across different views. This enables the MMViT model to acquire complex high-dimensional representations of the input at different resolutions. The proposed model can serve as a backbone model in multiple domains. We demonstrate the effectiveness of MMViT on audio and image classification tasks, achieving state-of-the-art results.

Chenbin Pan, Rui Hou, Hanchao Yu et al. · meta-ai

Whether by processing videos with fixed resolution from start to end or incorporating pooling and down-scaling strategies, existing video transformers process the whole video content throughout the network without specially handling the large portions of redundant information. In this paper, we present a Supertoken Video Transformer (SVT) that incorporates a Semantic Pooling Module (SPM) to aggregate latent representations along the depth of visual transformer based on their semantics, and thus, reduces redundancy inherent in video inputs.~Qualitative results show that our method can effectively reduce redundancy by merging latent representations with similar semantics and thus increase the proportion of salient information for downstream tasks.~Quantitatively, our method improves the performance of both ViT and MViT while requiring significantly less computations on the Kinectics and Something-Something-V2 benchmarks.~More specifically, with our SPM, we improve the accuracy of MAE-pretrained ViT-B and ViT-L by 1.5% with 33% less GFLOPs and by 0.2% with 55% less FLOPs, respectively, on the Kinectics-400 benchmark, and improve the accuracy of MViTv2-B by 0.2% and 0.3% with 22% less GFLOPs on Kinectics-400 and Something-Something-V2, respectively.

Xueyi Liu, Rui Hou, Yanglei Gan et al.

Aspect-based sentiment classification is a crucial problem in fine-grained sentiment analysis, which aims to predict the sentiment polarity of the given aspect according to its context. Previous works have made remarkable progress in leveraging attention mechanism to extract opinion words for different aspects. However, a persistent challenge is the effective management of semantic mismatches, which stem from attention mechanisms that fall short in adequately aligning opinions words with their corresponding aspect in multi-aspect sentences. To address this issue, we propose a novel Aspect-oriented Opinion Alignment Network (AOAN) to capture the contextual association between opinion words and the corresponding aspect. Specifically, we first introduce a neighboring span enhanced module which highlights various compositions of neighboring words and given aspects. In addition, we design a multi-perspective attention mechanism that align relevant opinion information with respect to the given aspect. Extensive experiments on three benchmark datasets demonstrate that our model achieves state-of-the-art results. The source code is available at https://github.com/AONE-NLP/ABSA-AOAN.

Suyu Ge, Chunting Zhou, Rui Hou et al.

Red-teaming is a common practice for mitigating unsafe behaviors in Large Language Models (LLMs), which involves thoroughly assessing LLMs to identify potential flaws and addressing them with responsible and accurate responses. While effective, manual red-teaming is costly, and existing automatic red-teaming typically discovers safety risks without addressing them. In this paper, we propose a Multi-round Automatic Red-Teaming (MART) method, which incorporates both automatic adversarial prompt writing and safe response generation, significantly increasing red-teaming scalability and the safety of the target LLM. Specifically, an adversarial LLM and a target LLM interplay with each other in an iterative manner, where the adversarial LLM aims to generate challenging prompts that elicit unsafe responses from the target LLM, while the target LLM is fine-tuned with safety aligned data on these adversarial prompts. In each round, the adversarial LLM crafts better attacks on the updated target LLM, while the target LLM also improves itself through safety fine-tuning. On adversarial prompt benchmarks, the violation rate of an LLM with limited safety alignment reduces up to 84.7% after 4 rounds of MART, achieving comparable performance to LLMs with extensive adversarial prompt writing. Notably, model helpfulness on non-adversarial prompts remains stable throughout iterations, indicating the target LLM maintains strong performance on instruction following.

Ashkan Kazemi, Artem Abzaliev, Naihao Deng et al.

We propose a novel system to help fact-checkers formulate search queries for known misinformation claims and effectively search across multiple social media platforms. We introduce an adaptable rewriting strategy, where editing actions for queries containing claims (e.g., swap a word with its synonym; change verb tense into present simple) are automatically learned through offline reinforcement learning. Our model uses a decision transformer to learn a sequence of editing actions that maximizes query retrieval metrics such as mean average precision. We conduct a series of experiments showing that our query rewriting system achieves a relative increase in the effectiveness of the queries of up to 42%, while producing editing action sequences that are human interpretable.

Siddharth Verma, Yuchen Lu, Rui Hou et al.

Masked Language Modeling (MLM) has proven to be an essential component of Vision-Language (VL) pretraining. To implement MLM, the researcher must make two design choices: the masking strategy, which determines which tokens to mask, and the masking rate, which determines how many tokens to mask. Previous work has focused primarily on the masking strategy while setting the masking rate at a default of 15\%. In this paper, we show that increasing this masking rate improves downstream performance while simultaneously reducing performance gap among different masking strategies, rendering the uniform masking strategy competitive to other more complex ones. Surprisingly, we also discover that increasing the masking rate leads to gains in Image-Text Matching (ITM) tasks, suggesting that the role of MLM goes beyond language modeling in VL pretraining.

Banghao Chi, Yining Xie, Mingyuan Wu et al.

Spreadsheet systems (e.g., Microsoft Excel, Google Sheets) play a central role in modern data-centric workflows. As AI agents grow increasingly capable of automating complex tasks, such as controlling computers and generating presentations, building an AI-driven spreadsheet agent has emerged as a promising research direction. Most existing spreadsheet agents rely on specialized prompting over general-purpose LLMs; while this design has potentials on simple spreadsheet operations, it struggles to manage the complex, multi-step workflows typical of real-world applications. We introduce Spreadsheet-RL, a reinforcement learning (RL) fine-tuning framework designed to train specialized spreadsheet agents within a realistic Microsoft Excel environment. Spreadsheet-RL features an automated pipeline for scalable collection of paired start-goal spreadsheets from online forums, as well as domain-specific evaluation tasks in areas such as finance and supply chain management, which we compile into the new Domain-Spreadsheet benchmark dataset. It also includes a Spreadsheet Gym environment designed for multi-turn RL: Spreadsheet Gym exposes extensive Excel functionality through a Python sandbox, along with a refined harness that incorporates a comprehensive tool set and carefully designed tool-routing rules for spreadsheet tasks. Through comprehensive experiments, we show that Spreadsheet-RL substantially enhances AI agent's performance on both general and domain-specific spreadsheet tasks: it improves Qwen3-4B-Thinking-2507's Pass@1 on SpreadsheetBench from 12.0% to 23.4%, and raises Pass@1 from 8.4% to 17.2% on our curated Domain-Spreadsheet dataset. These results highlight Spreadsheet-RL's strong potential for generalization and real-world adoption in spreadsheet automation, and broadly, its promise for advancing LLM-based interactions with data interfaces in everyday work.

Da Luo, Yanglei Gan, Rui Hou et al.

Few-shot Relation Extraction (FSRE) aims to extract relational facts from a sparse set of labeled corpora. Recent studies have shown promising results in FSRE by employing Pre-trained Language Models (PLMs) within the framework of supervised contrastive learning, which considers both instances and label facts. However, how to effectively harness massive instance-label pairs to encompass the learned representation with semantic richness in this learning paradigm is not fully explored. To address this gap, we introduce a novel synergistic anchored contrastive pre-training framework. This framework is motivated by the insight that the diverse viewpoints conveyed through instance-label pairs capture incomplete yet complementary intrinsic textual semantics. Specifically, our framework involves a symmetrical contrastive objective that encompasses both sentence-anchored and label-anchored contrastive losses. By combining these two losses, the model establishes a robust and uniform representation space. This space effectively captures the reciprocal alignment of feature distributions among instances and relational facts, simultaneously enhancing the maximization of mutual information across diverse perspectives within the same relation. Experimental results demonstrate that our framework achieves significant performance enhancements compared to baseline models in downstream FSRE tasks. Furthermore, our approach exhibits superior adaptability to handle the challenges of domain shift and zero-shot relation extraction. Our code is available online at https://github.com/AONE-NLP/FSRE-SaCon.

Ciyan Ouyang, Rui Hou

We introduce MemLineage, a defense for LLM agent memory that attaches both cryptographic provenance and LLM-mediated derivation lineage to every entry. Recent and concurrent work shows that untrusted content can be written into persistent agent state and re-enter later sessions as an instruction; the remaining systems question is how to preserve useful memory recall while preventing such state from justifying sensitive actions. MemLineage treats this as a chain-of-custody problem rather than a filtering problem. It is a six-module design around an RFC-6962 Merkle log over per-principal Ed25519-signed entries: a weighted derivation DAG records which retrieved entries influenced each new memory, and a max-of-strong-edges propagation rule makes Untrusted-Path Persistence hold for any chain whose attribution edges remain above threshold. The sensitive-action gate then refuses dispatches whose active justification descends from an external ancestor, while still allowing benign recall. We evaluate three defense cells against three memory-poisoning workloads on a deterministic mechanism-isolation harness; MemLineage is the only configuration in that harness that drives all three columns to zero ASR, while sub-millisecond per-operation overhead keeps it well below the noise floor of any LLM call. A Codex-backed AgentDojo bridge further separates strong-model behavior from defense-layer behavior: under an intentionally vulnerable tool-output profile, no-defense and signature-only baselines fail on all six banking pairs, while all MemLineage rows reduce strict AgentDojo ASR to zero. The core deterministic artifacts are byte-equal CI-verified; hosted-model AgentDojo and live-model sweeps are recorded as auditable logs rather than byte-pinned artifacts.

Peiyao Wang, Haotian Xu, Noranart Vesdapunt et al.

Multi-modal Large Language Models (MLLMs) show promise in video understanding. However, their reasoning often suffers from thinking drift and weak temporal comprehension, even when enhanced by Reinforcement Learning (RL) techniques like Group Relative Policy Optimization (GRPO). Moreover, existing RL methods usually depend on Supervised Fine-Tuning (SFT), which requires costly Chain-of-Thought (CoT) annotation and multi-stage training, and enforces fixed reasoning paths, limiting MLLMs' ability to generalize and potentially inducing bias. To overcome these limitations, we introduce Summary-Driven Reinforcement Learning (SDRL), a novel single-stage RL framework that obviates the need for SFT by utilizing a Structured CoT format: Summarize -> Think -> Answer. SDRL introduces two self-supervised mechanisms integrated into the GRPO objective: 1) Consistency of Vision Knowledge (CVK) enforces factual grounding by reducing KL divergence among generated summaries; and 2) Dynamic Variety of Reasoning (DVR) promotes exploration by dynamically modulating thinking diversity based on group accuracy. This novel integration effectively balances alignment and exploration, supervising both the final answer and the reasoning process. Our method achieves state-of-the-art performance on seven public VideoQA datasets.

Hayeon Lee, Rui Hou, Jongpil Kim et al.

Knowledge Distillation (KD) compresses computationally expensive pre-trained language models (PLMs) by transferring their knowledge to smaller models, allowing their use in resource-constrained or real-time settings. However, most smaller models fail to surpass the performance of the original larger model, resulting in sacrificing performance to improve inference speed. To address this issue, we propose Co-Training and Co-Distillation (CTCD), a novel framework that improves performance and inference speed together by co-training two models while mutually distilling knowledge. The CTCD framework successfully achieves this based on two significant findings: 1) Distilling knowledge from the smaller model to the larger model during co-training improves the performance of the larger model. 2) The enhanced performance of the larger model further boosts the performance of the smaller model. The CTCD framework shows promise as it can be combined with existing techniques like architecture design or data augmentation, replacing one-way KD methods, to achieve further performance improvement. Extensive ablation studies demonstrate the effectiveness of CTCD, and the small model distilled by CTCD outperforms the original larger model by a significant margin of 1.66 on the GLUE benchmark.

Ciyan Ouyang, Peinan Li, Yubiao Huang et al.

We present Janus, a compiler-based security framework that mitigates transient execution attacks like Spectre and control-flow hijacking on ARM64 platforms. Janus integrates speculative execution and control flow dependencies with PA modifiers, using PA and BTI microarchitectural features to prevent control-flow speculation attacks and secure both control flow and speculative execution through existing control-flow integrity mechanisms. To optimize performance, Janus minimizes overhead by merging defense operations across different defense layers (modifier fusion) and reusing registers of protected variables (carrier reuse), while maintaining strong security guarantees. Evaluation on SPEC CPU2017 shows an average performance overhead of 3.85%, with real-world applications exhibiting overheads ranging from 2.97% to 7.80%. Janus offers effective speculative execution security and low performance and code size overhead, making it a robust solution for ARM-based systems.

John Yang, Kilian Lieret, Jeffrey Ma et al.

Turning ideas into full software projects from scratch has become a popular use case for language models. Agents are being deployed to seed, maintain, and grow codebases over extended periods with minimal human oversight. Such settings require models to make high-level software architecture decisions. However, existing benchmarks measure focused, limited tasks such as fixing a single bug or developing a single, specified feature. We therefore introduce ProgramBench to measure the ability of software engineering agents to develop software holisitically. In ProgramBench, given only a program and its documentation, agents must architect and implement a codebase that matches the reference executable's behavior. End-to-end behavioral tests are generated via agent-driven fuzzing, enabling evaluation without prescribing implementation structure. Our 200 tasks range from compact CLI tools to widely used software such as FFmpeg, SQLite, and the PHP interpreter. We evaluate 9 LMs and find that none fully resolve any task, with the best model passing 95\% of tests on only 3\% of tasks. Models favor monolithic, single-file implementations that diverge sharply from human-written code.

Tim Franzmeyer, Archie Sravankumar, Lijuan Liu et al.

Large Language Models (LLMs) currently respond to every prompt. However, they can produce incorrect answers when they lack knowledge or capability -- a problem known as hallucination. We instead propose post-training an LLM to generate content only when confident in its correctness and to otherwise (partially) abstain. Specifically, our method, HALT, produces capability-aligned post-training data that encodes what the model can and cannot reliably generate. We generate this data by splitting responses of the pretrained LLM into factual fragments (atomic statements or reasoning steps), and use ground truth information to identify incorrect fragments. We achieve capability-aligned finetuning responses by either removing incorrect fragments or replacing them with "Unsure from Here" -- according to a tunable threshold that allows practitioners to trade off response completeness and mean correctness of the response's fragments. We finetune four open-source models for biography writing, mathematics, coding, and medicine with HALT for three different trade-off thresholds. HALT effectively trades off response completeness for correctness, increasing the mean correctness of response fragments by 15% on average, while resulting in a 4% improvement in the F1 score (mean of completeness and correctness of the response) compared to the relevant baselines. By tuning HALT for highest correctness, we train a single reliable Llama3-70B model with correctness increased from 51% to 87% across all four domains while maintaining 53% of the response completeness achieved with standard finetuning.

Shraman Pramanick, Guangxing Han, Rui Hou et al.

The ability of large language models (LLMs) to process visual inputs has given rise to general-purpose vision systems, unifying various vision-language (VL) tasks by instruction tuning. However, due to the enormous diversity in input-output formats in the vision domain, existing general-purpose models fail to successfully integrate segmentation and multi-image inputs with coarse-level tasks into a single framework. In this work, we introduce VistaLLM, a powerful visual system that addresses coarse- and fine-grained VL tasks over single and multiple input images using a unified framework. VistaLLM utilizes an instruction-guided image tokenizer that filters global embeddings using task descriptions to extract compressed and refined features from numerous images. Moreover, VistaLLM employs a gradient-aware adaptive sampling technique to represent binary segmentation masks as sequences, significantly improving over previously used uniform sampling. To bolster the desired capability of VistaLLM, we curate CoinIt, a comprehensive coarse-to-fine instruction tuning dataset with 6.8M samples. We also address the lack of multi-image grounding datasets by introducing a novel task, AttCoSeg (Attribute-level Co-Segmentation), which boosts the model's reasoning and grounding capability over multiple input images. Extensive experiments on a wide range of V- and VL tasks demonstrate the effectiveness of VistaLLM by achieving consistent state-of-the-art performance over strong baselines across all downstream tasks. Our project page can be found at https://shramanpramanick.github.io/VistaLLM/.

Yue Yu, Zhengxing Chen, Aston Zhang et al. · allen-ai

Reward modeling is crucial for aligning large language models (LLMs) with human preferences, especially in reinforcement learning from human feedback (RLHF). However, current reward models mainly produce scalar scores and struggle to incorporate critiques in a natural language format. We hypothesize that predicting both critiques and the scalar reward would improve reward modeling ability. Motivated by this, we propose Critic-RM, a framework that improves reward models using self-generated critiques without extra supervision. Critic-RM employs a two-stage process: generating and filtering high-quality critiques, followed by joint fine-tuning on reward prediction and critique generation. Experiments across benchmarks show that Critic-RM improves reward modeling accuracy by 3.7%-7.3% compared to standard reward models and LLM judges, demonstrating strong performance and data efficiency. Additional studies further validate the effectiveness of generated critiques in rectifying flawed reasoning steps with 2.5%-3.2% gains in improving reasoning accuracy.

Bo Wu, Sid Wang, Yunhao Tang et al.

Reinforcement Learning (RL) has become the most effective post-training approach for improving the capabilities of Large Language Models (LLMs). In practice, because of the high demands on latency and memory, it is particularly challenging to develop an efficient RL framework that reliably manages policy models with hundreds to thousands of billions of parameters. In this paper, we present LlamaRL, a fully distributed, asynchronous RL framework optimized for efficient training of large-scale LLMs with various model sizes (8B, 70B, and 405B parameters) on GPU clusters ranging from a handful to thousands of devices. LlamaRL introduces a streamlined, single-controller architecture built entirely on native PyTorch, enabling modularity, ease of use, and seamless scalability to thousands of GPUs. We also provide a theoretical analysis of LlamaRL's efficiency, including a formal proof that its asynchronous design leads to strict RL speed-up. Empirically during the Llama 3 post-training, by leveraging best practices such as colocated model offloading, asynchronous off-policy training, and distributed direct memory access for weight synchronization, LlamaRL achieves significant efficiency gains -- up to 10.7x speed-up compared to DeepSpeed-Chat-like systems on a 405B-parameter policy model. Furthermore, the efficiency advantage continues to grow with increasing model scale, demonstrating the framework's suitability for future large-scale RL training.

Xianjun Yang, Shaoliang Nie, Lijuan Liu et al. · allen-ai

Instruction tuning data are often quantity-saturated due to the large volume of data collection and fast model iteration, leaving data selection important but underexplored. Existing quality-driven data selection methods, such as LIMA (NeurIPS 2023 \citep{zhou2024lima}) and AlpaGasus (ICLR 2024 \citep{chenalpagasus}) generally ignore the equal importance of data diversity and complexity. In this work, we aim to design a diversity-aware data selection strategy and creatively propose using sparse autoencoders (SAEs) to tackle the challenge of data diversity measure. In addition, SAEs can also provide more interpretability of model behavior and explain, e.g., the surprising effectiveness of selecting the longest response (ICML 2024 \citep{zhaolong}). Using effective data selection, we experimentally prove that models trained on our selected data can outperform other methods in terms of model capabilities, reduce training cost, and potentially gain more control over model behaviors. We prove that SAEs can serve as a good alternative to diversity measure and design our method to be scalable for potential industrial large-scale pruning, and we will also release our trained SAEs for use by the broader community.

Joongwon Kim, Anirudh Goyal, Aston Zhang et al.

Preference learning is a widely adopted post-training technique that aligns large language models (LLMs) to human preferences and improves specific downstream task capabilities. In this work we systematically investigate how specific attributes of preference datasets affect the alignment and downstream performance of LLMs in instruction-following tasks. We use a novel synthetic data generation pipeline to generate 48,000 unique instruction-following prompts with combinations of 23 verifiable constraints that enable fine-grained and automated quality assessments of model responses. With our synthetic prompts, we use two preference dataset curation methods - rejection sampling (RS) and Monte Carlo Tree Search (MCTS) - to obtain pairs of (chosen, rejected) responses. Then, we perform experiments investigating the effects of (1) the presence of shared prefixes between the chosen and rejected responses, (2) the contrast and quality of the chosen, rejected responses and (3) the complexity of the training prompts. Our experiments reveal that shared prefixes in preference pairs, as generated by MCTS, provide marginal but consistent improvements and greater stability across challenging training configurations. High-contrast preference pairs generally outperform low-contrast pairs; however, combining both often yields the best performance by balancing diversity and learning efficiency. Additionally, training on prompts of moderate difficulty leads to better generalization across tasks, even for more complex evaluation scenarios, compared to overly challenging prompts. Our findings provide actionable insights into optimizing preference data curation for instruction-following tasks, offering a scalable and effective framework for enhancing LLM training and alignment.

Jaehyung Kim, Yuning Mao, Rui Hou et al.

Fine-tuning pre-trained language models (LMs) has become the de facto standard in many NLP tasks. Nevertheless, fine-tuned LMs are still prone to robustness issues, such as adversarial robustness and model calibration. Several perspectives of robustness for LMs have been studied independently, but lacking a unified consideration in multiple perspectives. In this paper, we propose Robustifying LMs via Adversarial perturbation with Selective Training (RoAST), a simple yet effective fine-tuning technique to enhance the multi-perspective robustness of LMs in a unified way. RoAST effectively incorporates two important sources for the model robustness, robustness on the perturbed inputs and generalizable knowledge in pre-trained LMs. To be specific, RoAST introduces adversarial perturbation during fine-tuning while the model parameters are selectively updated upon their relative importance to minimize unnecessary deviation. Under a unified evaluation of fine-tuned LMs by incorporating four representative perspectives of model robustness, we demonstrate the effectiveness of RoAST compared to state-of-the-art fine-tuning methods on six different types of LMs, which indicates its usefulness in practice.

Guang Yan, Yuhui Zhang, Zimu Guo et al.

With the growing use of large language models (LLMs) hosted on cloud platforms to offer inference services, privacy concerns about the potential leakage of sensitive information are escalating. Secure multi-party computation (MPC) is a promising solution to protect the privacy in LLM inference. However, MPC requires frequent inter-server communication, causing high performance overhead. Inspired by the prevalent activation sparsity of LLMs, where most neuron are not activated after non-linear activation functions, we propose an efficient private inference system, Comet. This system employs an accurate and fast predictor to predict the sparsity distribution of activation function output. Additionally, we introduce a new private inference protocol. It efficiently and securely avoids computations involving zero values by exploiting the spatial locality of the predicted sparse distribution. While this computation-avoidance approach impacts the spatiotemporal continuity of KV cache entries, we address this challenge with a low-communication overhead cache refilling strategy that merges miss requests and incorporates a prefetching mechanism. Finally, we evaluate Comet on four common LLMs and compare it with six state-of-the-art private inference systems. Comet achieves a 1.87x-2.63x speedup and a 1.94x-2.64x communication reduction.

Xuanting Xie, Bingheng Li, Erlin Pan et al.

Attention mechanisms have become a cornerstone in modern neural networks, driving breakthroughs across diverse domains. However, their application to graph structured data, where capturing topological connections is essential, remains underexplored and underperforming compared to Graph Neural Networks (GNNs), particularly in the graph clustering task. GNN tends to overemphasize neighborhood aggregation, leading to a homogenization of node representations. Conversely, Transformer tends to over globalize, highlighting distant nodes at the expense of meaningful local patterns. This dichotomy raises a key question: Is attention inherently redundant for unsupervised graph learning? To address this, we conduct a comprehensive empirical analysis, uncovering the complementary weaknesses of GNN and Transformer in graph clustering. Motivated by these insights, we propose the Attentive Graph Clustering Network (AGCN) a novel architecture that reinterprets the notion that graph is attention. AGCN directly embeds the attention mechanism into the graph structure, enabling effective global information extraction while maintaining sensitivity to local topological cues. Our framework incorporates theoretical analysis to contrast AGCN behavior with GNN and Transformer and introduces two innovations: (1) a KV cache mechanism to improve computational efficiency, and (2) a pairwise margin contrastive loss to boost the discriminative capacity of the attention space. Extensive experimental results demonstrate that AGCN outperforms state-of-the-art methods.

Hayeon Lee, Rui Hou, Jongpil Kim et al.

Distillation from Weak Teacher (DWT) is a method of transferring knowledge from a smaller, weaker teacher model to a larger student model to improve its performance. Previous studies have shown that DWT can be effective in the vision domain and natural language processing (NLP) pre-training stage. Specifically, DWT shows promise in practical scenarios, such as enhancing new generation or larger models using pre-trained yet older or smaller models and lacking a resource budget. However, the optimal conditions for using DWT have yet to be fully investigated in NLP pre-training. Therefore, this study examines three key factors to optimize DWT, distinct from those used in the vision domain or traditional knowledge distillation. These factors are: (i) the impact of teacher model quality on DWT effectiveness, (ii) guidelines for adjusting the weighting value for DWT loss, and (iii) the impact of parameter remapping as a student model initialization technique for DWT.

Kuan-Hao Huang, Liang Tan, Rui Hou et al.

Many real-world applications require making multiple predictions from the same text. Fine-tuning a large pre-trained language model for each downstream task causes computational burdens in the inference time due to several times of forward passes. To amortize the computational cost, freezing the language model and building lightweight models for downstream tasks based on fixed text representations are common solutions. Accordingly, how to learn fixed but general text representations that can generalize well to unseen downstream tasks becomes a challenge. Previous works have shown that the generalizability of representations can be improved by fine-tuning the pre-trained language model with some source tasks in a multi-tasking way. In this work, we propose a prefix-based method to learn the fixed text representations with source tasks. We learn a task-specific prefix for each source task independently and combine them to get the final representations. Our experimental results show that prefix-based training performs better than multi-tasking training and can update the text representations at a smaller computational cost than multi-tasking training.

Anastasia Razdaibiedina, Yuning Mao, Rui Hou et al.

Prompt tuning is one of the successful approaches for parameter-efficient tuning of pre-trained language models. Despite being arguably the most parameter-efficient (tuned soft prompts constitute <0.1% of total parameters), it typically performs worse than other efficient tuning methods and is quite sensitive to hyper-parameters. In this work, we introduce Residual Prompt Tuning - a simple and efficient method that significantly improves the performance and stability of prompt tuning. We propose to reparameterize soft prompt embeddings using a shallow network with a residual connection. Our experiments show that Residual Prompt Tuning significantly outperforms prompt tuning on SuperGLUE benchmark. Notably, our method reaches +7 points improvement over prompt tuning with T5-Base and allows to reduce the prompt length by 10x without hurting performance. In addition, we show that our approach is robust to the choice of learning rate and prompt initialization, and is effective in few-shot settings.

Yuning Mao, Lambert Mathias, Rui Hou et al.

Recent parameter-efficient language model tuning (PELT) methods manage to match the performance of fine-tuning with much fewer trainable parameters and perform especially well when training data is limited. However, different PELT methods may perform rather differently on the same task, making it nontrivial to select the most appropriate method for a specific task, especially considering the fast-growing number of new PELT methods and tasks. In light of model diversity and the difficulty of model selection, we propose a unified framework, UniPELT, which incorporates different PELT methods as submodules and learns to activate the ones that best suit the current data or task setup via gating mechanism. On the GLUE benchmark, UniPELT consistently achieves 1~4% gains compared to the best individual PELT method that it incorporates and even outperforms fine-tuning under different setups. Moreover, UniPELT generally surpasses the upper bound that takes the best performance of all its submodules used individually on each task, indicating that a mixture of multiple PELT methods may be inherently more effective than single methods.

Fengkai Yuan, Kai Wang, Rui Hou et al.

Cache side channel attacks obtain victim cache line access footprint to infer security-critical information. Among them, cross-core attacks exploiting the shared last level cache are more threatening as their simplicity to set up and high capacity. Stateful approaches of detection-based mitigation observe precise cache behaviors and protect specific cache lines that are suspected of being attacked. However, their recording structures incur large storage overhead and are vulnerable to reverse engineering attacks. Exploring the intrinsic non-determinate layout of a traditional Cuckoo filter, this paper proposes a space efficient Auto-Cuckoo filter to record access footprints, which succeed to decrease storage overhead and resist reverse engineering attacks at the same time. With Auto-Cuckoo filter, we propose PiPoMonitor to detect \textit{Ping-Pong patterns} and prefetch specific cache line to interfere with adversaries' cache probes. Security analysis shows the PiPoMonitor can effectively mitigate cross-core attacks and the Auto-Cuckoo filter is immune to reverse engineering attacks. Evaluation results indicate PiPoMonitor has negligible impact on performance and the storage overhead is only 0.37$\%$, an order of magnitude lower than previous stateful approaches.

Anindo Saha, Fakrul I. Tushar, Khrystyna Faryna et al.

Weakly supervised disease classification of CT imaging suffers from poor localization owing to case-level annotations, where even a positive scan can hold hundreds to thousands of negative slices along multiple planes. Furthermore, although deep learning segmentation and classification models extract distinctly unique combinations of anatomical features from the same target class(es), they are typically seen as two independent processes in a computer-aided diagnosis (CAD) pipeline, with little to no feature reuse. In this research, we propose a medical classifier that leverages the semantic structural concepts learned via multi-resolution segmentation feature maps, to guide weakly supervised 3D classification of chest CT volumes. Additionally, a comparative analysis is drawn across two different types of feature aggregation to explore the vast possibilities surrounding feature fusion. Using a dataset of 1593 scans labeled on a case-level basis via rule-based model, we train a dual-stage convolutional neural network (CNN) to perform organ segmentation and binary classification of four representative diseases (emphysema, pneumonia/atelectasis, mass and nodules) in lungs. The baseline model, with separate stages for segmentation and classification, results in AUC of 0.791. Using identical hyperparameters, the connected architecture using static and dynamic feature aggregation improves performance to AUC of 0.832 and 0.851, respectively. This study advances the field in two key ways. First, case-level report data is used to weakly supervise a 3D CT classifier of multiple, simultaneous diseases for an organ. Second, segmentation and classification models are connected with two different feature aggregation strategies to enhance the classification performance.

Fakrul Islam Tushar, Vincent M. D'Anniballe, Rui Hou et al.

Purpose: To design multi-disease classifiers for body CT scans for three different organ systems using automatically extracted labels from radiology text reports.Materials & Methods: This retrospective study included a total of 12,092 patients (mean age 57 +- 18; 6,172 women) for model development and testing (from 2012-2017). Rule-based algorithms were used to extract 19,225 disease labels from 13,667 body CT scans from 12,092 patients. Using a three-dimensional DenseVNet, three organ systems were segmented: lungs and pleura; liver and gallbladder; and kidneys and ureters. For each organ, a three-dimensional convolutional neural network classified no apparent disease versus four common diseases for a total of 15 different labels across all three models. Testing was performed on a subset of 2,158 CT volumes relative to 2,875 manually derived reference labels from 2133 patients (mean age 58 +- 18;1079 women). Performance was reported as receiver operating characteristic area under the curve (AUC) with 95% confidence intervals by the DeLong method. Results: Manual validation of the extracted labels confirmed 91% to 99% accuracy across the 15 different labels. AUCs for lungs and pleura labels were: atelectasis 0.77 (95% CI: 0.74, 0.81), nodule 0.65 (0.61, 0.69), emphysema 0.89 (0.86, 0.92), effusion 0.97 (0.96, 0.98), and no apparent disease 0.89 (0.87, 0.91). AUCs for liver and gallbladder were: hepatobiliary calcification 0.62 (95% CI: 0.56, 0.67), lesion 0.73 (0.69, 0.77), dilation 0.87 (0.84, 0.90), fatty 0.89 (0.86, 0.92), and no apparent disease 0.82 (0.78, 0.85). AUCs for kidneys and ureters were: stone 0.83 (95% CI: 0.79, 0.87), atrophy 0.92 (0.89, 0.94), lesion 0.68 (0.64, 0.72), cyst 0.70 (0.66, 0.73), and no apparent disease 0.79 (0.75, 0.83). Conclusion: Weakly-supervised deep learning models were able to classify diverse diseases in multiple organ systems.

Lutan Zhao, Peinan Li, Rui Hou et al.

Recently exposed vulnerabilities reveal the necessity to improve the security of branch predictors. Branch predictors record history about the execution of different programs, and such information from different processes are stored in the same structure and thus accessible to each other. This leaves the attackers with the opportunities for malicious training and malicious perception. Instead of flush-based or physical isolation of hardware resources, we want to achieve isolation of the content in these hardware tables with some lightweight processing using randomization as follows. (1) Content encoding. We propose to use hardware-based thread-private random numbers to encode the contents of the branch predictor tables (both direction and destination histories) which we call XOR-BP. Specifically, the data is encoded by XOR operation with the key before written in the table and decoded after read from the table. Such a mechanism obfuscates the information adding difficulties to cross-process or cross-privilege level analysis and perception. It achieves a similar effect of logical isolation but adds little in terms of space or time overheads. (2) Index encoding. We propose a randomized index mechanism of the branch predictor (Noisy-XOR-BP). Similar to the XOR-BP, another thread-private random number is used together with the branch instruction address as the input to compute the index of the branch predictor. This randomized indexing mechanism disrupts the correspondence between the branch instruction address and the branch predictor entry, thus increases the noise for malicious perception attacks. Our analyses using an FPGA-based RISC-V processor prototype and additional auxiliary simulations suggest that the proposed mechanisms incur a very small performance cost while providing strong protection.

Vitor Guizilini, Rui Hou, Jie Li et al.

Self-supervised learning is showing great promise for monocular depth estimation, using geometry as the only source of supervision. Depth networks are indeed capable of learning representations that relate visual appearance to 3D properties by implicitly leveraging category-level patterns. In this work we investigate how to leverage more directly this semantic structure to guide geometric representation learning, while remaining in the self-supervised regime. Instead of using semantic labels and proxy losses in a multi-task approach, we propose a new architecture leveraging fixed pretrained semantic segmentation networks to guide self-supervised representation learning via pixel-adaptive convolutions. Furthermore, we propose a two-stage training process to overcome a common semantic bias on dynamic objects via resampling. Our method improves upon the state of the art for self-supervised monocular depth prediction over all pixels, fine-grained details, and per semantic categories.

Peng Sun, Rui Hou, Jerome Lynch

Physical activities and social interactions are essential activities that ensure a healthy lifestyle. Public open spaces (POS), such as parks, plazas and greenways, are key environments that encourage those activities. To evaluate a POS, there is a need to study how humans use the facilities within it. However, traditional approaches to studying use of POS are manual and therefore time and labor intensive. They also may only provide qualitative insights. It is appealing to make use of surveillance cameras and to extract user-related information through computer vision. This paper proposes a proof-of-concept deep learning computer vision framework for measuring human activities quantitatively in POS and demonstrates a case study of the proposed framework using the Detroit Riverfront Conservancy (DRFC) surveillance camera network. A custom image dataset is presented to train the framework; the dataset includes 7826 fully annotated images collected from 18 cameras across the DRFC park space under various illumination conditions. Dataset analysis is also provided as well as a baseline model for one-step user localization and activity recognition. The mAP results are 77.5\% for {\it pedestrian} detection and 81.6\% for {\it cyclist} detection. Behavioral maps are autonomously generated by the framework to locate different POS users and the average error for behavioral localization is within 10 cm.

Rui Hou, Jie Li, Arjun Bhargava et al.

Panoptic segmentation is a complex full scene parsing task requiring simultaneous instance and semantic segmentation at high resolution. Current state-of-the-art approaches cannot run in real-time, and simplifying these architectures to improve efficiency severely degrades their accuracy. In this paper, we propose a new single-shot panoptic segmentation network that leverages dense detections and a global self-attention mechanism to operate in real-time with performance approaching the state of the art. We introduce a novel parameter-free mask construction method that substantially reduces computational complexity by efficiently reusing information from the object detection and semantic segmentation sub-tasks. The resulting network has a simple data flow that does not require feature map re-sampling or clustering post-processing, enabling significant hardware acceleration. Our experiments on the Cityscapes and COCO benchmarks show that our network works at 30 FPS on 1024x2048 resolution, trading a 3% relative performance degradation from the current state of the art for up to 440% faster inference.

Rui Hou, Verónica Pérez-Rosas, Stacy Loeb et al.

Recent years have witnessed a significant increase in the online sharing of medical information, with videos representing a large fraction of such online sources. Previous studies have however shown that more than half of the health-related videos on platforms such as YouTube contain misleading information and biases. Hence, it is crucial to build computational tools that can help evaluate the quality of these videos so that users can obtain accurate information to help inform their decisions. In this study, we focus on the automatic detection of misinformation in YouTube videos. We select prostate cancer videos as our entry point to tackle this problem. The contribution of this paper is twofold. First, we introduce a new dataset consisting of 250 videos related to prostate cancer manually annotated for misinformation. Second, we explore the use of linguistic, acoustic, and user engagement features for the development of classification models to identify misinformation. Using a series of ablation experiments, we show that we can build automatic models with accuracies of up to 74%, corresponding to a 76.5% precision and 73.2% recall for misinformative instances.

Rui Hou, Chen Chen, Rahul Sukthankar et al.

Convolutional Neural Network (CNN) based image segmentation has made great progress in recent years. However, video object segmentation remains a challenging task due to its high computational complexity. Most of the previous methods employ a two-stream CNN framework to handle spatial and motion features separately. In this paper, we propose an end-to-end encoder-decoder style 3D CNN to aggregate spatial and temporal information simultaneously for video object segmentation. To efficiently process video, we propose 3D separable convolution for the pyramid pooling module and decoder, which dramatically reduces the number of operations while maintaining the performance. Moreover, we also extend our framework to video action segmentation by adding an extra classifier to predict the action label for actors in videos. Extensive experiments on several video datasets demonstrate the superior performance of the proposed approach for action and object segmentation compared to the state-of-the-art.

Jianping Zhu, Rui Hou, XiaoFeng Wang et al.

There is an urgent demand for privacy-preserving techniques capable of supporting compute and data intensive (CDI) computing in the era of big data. However, none of existing TEEs can truly support CDI computing tasks, as CDI requires high throughput accelerators like GPU and TPU but TEEs do not offer security protection of such accelerators. This paper present HETEE (Heterogeneous TEE), the first design of TEE capable of strongly protecting heterogeneous computing with unsecure accelerators. HETEE is uniquely constructed to work with today's servers, and does not require any changes for existing commercial CPUs or accelerators. The key idea of our design runs security controller as a stand-alone computing system to dynamically adjust the boundary of between secure and insecure worlds through the PCIe switches, rendering the control of an accelerator to the host OS when it is not needed for secure computing, and shifting it back when it is. The controller is the only trust unit in the system and it runs the custom OS and accelerator runtimes, together with the encryption, authentication and remote attestation components. The host server and other computing systems communicate with controller through an in memory task queue that accommodates the computing tasks offloaded to HETEE, in the form of encrypted and signed code and data. Also, HETEE offers a generic and efficient programming model to the host CPU. We have implemented the HETEE design on a hardware prototype system, and evaluated it with large-scale Neural Networks inference and training tasks. Our evaluations show that HETEE can easily support such secure computing tasks and only incurs a 12.34% throughput overhead for inference and 9.87% overhead for training on average.

Rui Hou, Chen Chen, Mubarak Shah

In this paper, we propose an end-to-end 3D CNN for action detection and segmentation in videos. The proposed architecture is a unified deep network that is able to recognize and localize action based on 3D convolution features. A video is first divided into equal length clips and next for each clip a set of tube proposals are generated based on 3D CNN features. Finally, the tube proposals of different clips are linked together and spatio-temporal action detection is performed using these linked video proposals. This top-down action detection approach explicitly relies on a set of good tube proposals to perform well and training the bounding box regression usually requires a large number of annotated samples. To remedy this, we further extend the 3D CNN to an encoder-decoder structure and formulate the localization problem as action segmentation. The foreground regions (i.e. action regions) for each frame are segmented first then the segmented foreground maps are used to generate the bounding boxes. This bottom-up approach effectively avoids tube proposal generation by leveraging the pixel-wise annotations of segmentation. The segmentation framework also can be readily applied to a general problem of video object segmentation. Extensive experiments on several video datasets demonstrate the superior performance of our approach for action detection and video object segmentation compared to the state-of-the-arts.

Rui Hou, Chen Chen, Mubarak Shah

Deep learning has been demonstrated to achieve excellent results for image classification and object detection. However, the impact of deep learning on video analysis (e.g. action detection and recognition) has been limited due to complexity of video data and lack of annotations. Previous convolutional neural networks (CNN) based video action detection approaches usually consist of two major steps: frame-level action proposal detection and association of proposals across frames. Also, these methods employ two-stream CNN framework to handle spatial and temporal feature separately. In this paper, we propose an end-to-end deep network called Tube Convolutional Neural Network (T-CNN) for action detection in videos. The proposed architecture is a unified network that is able to recognize and localize action based on 3D convolution features. A video is first divided into equal length clips and for each clip a set of tube proposals are generated next based on 3D Convolutional Network (ConvNet) features. Finally, the tube proposals of different clips are linked together employing network flow and spatio-temporal action detection is performed using these linked video proposals. Extensive experiments on several video datasets demonstrate the superior performance of T-CNN for classifying and localizing actions in both trimmed and untrimmed videos compared to state-of-the-arts.