arXiv subjects

Wenxuan Zhou

Publications and source records attributed to Wenxuan Zhou.

At least 37 records · Page 2Linked to original sources

arXiv2025

Instructional Segment Embedding: Improving LLM Safety with Instruction Hierarchy

Large Language Models (LLMs) are susceptible to security and safety threats, such as prompt injection, prompt extraction, and harmful requests. One major cause of these vulnerabilities is the lack of an instruction hierarchy. Modern LLM architectures treat all inputs equally, failing to distinguish between and prioritize various types of instructions, such as system messages, user prompts, and data. As a result, lower-priority user prompts may override more critical system instructions, including safety protocols. Existing approaches to achieving instruction hierarchy, such as delimiters and instruction-based training, do not address this issue at the architectural level. We introduce the Instructional Segment Embedding (ISE) technique, inspired by BERT, to modern large language models, which embeds instruction priority information directly into the model. This approach enables models to explicitly differentiate and prioritize various instruction types, significantly improving safety against malicious prompts that attempt to override priority rules. Our experiments on the Structured Query and Instruction Hierarchy benchmarks demonstrate an average robust accuracy increase of up to 15.75% and 18.68%, respectively. Furthermore, we observe an improvement in instruction-following capability of up to 4.1% evaluated on AlpacaEval. Overall, our approach offers a promising direction for enhancing the safety and effectiveness of LLM architectures.

cs.LG↗

arXiv2025

Think Smarter not Harder: Adaptive Reasoning with Inference Aware Optimization

Solving mathematics problems has been an intriguing capability of large language models, and many efforts have been made to improve reasoning by extending reasoning length, such as through self-correction and extensive long chain-of-thoughts. While promising in problem-solving, advanced long reasoning chain models exhibit an undesired single-modal behavior, where trivial questions require unnecessarily tedious long chains of thought. In this work, we propose a way to allow models to be aware of inference budgets by formulating it as utility maximization with respect to an inference budget constraint, hence naming our algorithm Inference Budget-Constrained Policy Optimization (IBPO). In a nutshell, models fine-tuned through IBPO learn to ``understand'' the difficulty of queries and allocate inference budgets to harder ones. With different inference budgets, our best models are able to have a $4.14$\% and $5.74$\% absolute improvement ($8.08$\% and $11.2$\% relative improvement) on MATH500 using $2.16$x and $4.32$x inference budgets respectively, relative to LLaMA3.1 8B Instruct. These improvements are approximately $2$x those of self-consistency under the same budgets.

cs.AI↗

arXiv2024

T-REG: Preference Optimization with Token-Level Reward Regularization

Reinforcement learning from human feedback (RLHF) has been crucial in aligning large language models (LLMs) with human values. Traditionally, RLHF involves generating responses to a query and using a reward model to assign a reward to the entire response. However, this approach faces challenges due to its reliance on a single, sparse reward, which makes it challenging for the model to identify which parts of the sequence contribute most significantly to the final reward. Recent methods have attempted to address this limitation by introducing token-level rewards. However, these methods often rely on either a trained credit assignment model or AI annotators, raising concerns about the quality and reliability of the rewards. In this paper, we propose token-level reward regularization (T-REG), a novel approach that leverages both sequence-level and token-level rewards for preference optimization. Harnessing the self-refinement capabilities of LLMs, our method uses contrastive prompting to enable LLMs to self-generate token-level rewards. These self-generated rewards then act as reward regularization, guiding the model to more effectively distribute sequence-level rewards across tokens. This facilitates better token-level credit assignment and enhances alignment performance. Experiments on the instruction following benchmarks, including Alpaca Eval 2 and Arena-Hard, show that our method consistently outperforms baseline methods by up to 3.8% and 4.4%, respectively. We will release the code and models at https://github.com/wzhouad/T-REG.

cs.CL↗

arXiv2024

Fine-grained Distributed Data Plane Verification with Intent-based Slicing

Data plane verification has grown into a powerful tool to ensure network correctness. However, existing methods with monolithic models have memory requirements tied to network sizes, and the existing method of scaling out is too limited in expressiveness to capture practical network features. In this paper, we describe Scylla, a general data plane verifier that provides fine-grained scale-out without the need for a monolithic network model. Scylla creates models for what we call intent-based slices, each of which is constructed at the rule-level granularity with only enough to verify a given set of intents. The sliced models are retained and incrementally updated in memory across a distributed compute cluster in response to network updates. Our experiments show that Scylla makes the scaling problem more granular -- tied to the size of the intent-based slices rather than that of the overall network. This enables Scylla to verify large, complex networks in minimum units of work that are significantly smaller (in both memory and time) than past techniques, enabling fast scale-out verification with minimal resource requirement.

cs.NI↗

arXiv2024

Visual Manipulation with Legs

Animals use limbs for both locomotion and manipulation. We aim to equip quadruped robots with similar versatility. This work introduces a system that enables quadruped robots to interact with objects using their legs, inspired by non-prehensile manipulation. The system has two main components: a visual manipulation policy module and a loco-manipulator module. The visual manipulation policy, trained with reinforcement learning (RL) using point cloud observations and object-centric actions, decides how the leg should interact with the object. The loco-manipulator controller manages leg movements and body pose adjustments, based on impedance control and Model Predictive Control (MPC). Besides manipulating objects with a single leg, the system can select from the left or right leg based on critic maps and move objects to distant goals through base adjustment. Experiments evaluate the system on object pose alignment tasks in both simulation and the real world, demonstrating more versatile object manipulation skills with legs than previous work. Videos can be found at https://legged-manipulation.github.io/

cs.RO↗

arXiv2024

mDPO: Conditional Preference Optimization for Multimodal Large Language Models

Direct preference optimization (DPO) has shown to be an effective method for large language model (LLM) alignment. Recent works have attempted to apply DPO to multimodal scenarios but have found it challenging to achieve consistent improvement. Through a comparative experiment, we identify the unconditional preference problem in multimodal preference optimization, where the model overlooks the image condition. To address this problem, we propose mDPO, a multimodal DPO objective that prevents the over-prioritization of language-only preferences by also optimizing image preference. Moreover, we introduce a reward anchor that forces the reward to be positive for chosen responses, thereby avoiding the decrease in their likelihood -- an intrinsic problem of relative preference optimization. Experiments on two multimodal LLMs of different sizes and three widely used benchmarks demonstrate that mDPO effectively addresses the unconditional preference problem in multimodal preference optimization and significantly improves model performance, particularly in reducing hallucination.

cs.CV↗

arXiv2024

WPO: Enhancing RLHF with Weighted Preference Optimization

Reinforcement learning from human feedback (RLHF) is a promising solution to align large language models (LLMs) more closely with human values. Off-policy preference optimization, where the preference data is obtained from other models, is widely adopted due to its cost efficiency and scalability. However, off-policy preference optimization often suffers from a distributional gap between the policy used for data collection and the target policy, leading to suboptimal optimization. In this paper, we propose a novel strategy to mitigate this problem by simulating on-policy learning with off-policy preference data. Our Weighted Preference Optimization (WPO) method adapts off-policy data to resemble on-policy data more closely by reweighting preference pairs according to their probability under the current policy. This method not only addresses the distributional gap problem but also enhances the optimization process without incurring additional costs. We validate our method on instruction following benchmarks including Alpaca Eval 2 and MT-bench. WPO not only outperforms Direct Preference Optimization (DPO) by up to 5.6% on Alpaca Eval 2 but also establishes a remarkable length-controlled winning rate against GPT-4-turbo of 76.7% based on Gemma-2-9b-it. We release the code and models at https://github.com/wzhouad/WPO.

cs.CL↗

arXiv2024

The Perfect Blend: Redefining RLHF with Mixture of Judges

Reinforcement learning from human feedback (RLHF) has become the leading approach for fine-tuning large language models (LLM). However, RLHF has limitations in multi-task learning (MTL) due to challenges of reward hacking and extreme multi-objective optimization (i.e., trade-off of multiple and/or sometimes conflicting objectives). Applying RLHF for MTL currently requires careful tuning of the weights for reward model and data combinations. This is often done via human intuition and does not generalize. In this work, we introduce a novel post-training paradigm which we called Constrained Generative Policy Optimization (CGPO). The core of CGPO is Mixture of Judges (MoJ) with cost-efficient constrained policy optimization with stratification, which can identify the perfect blend in RLHF in a principled manner. It shows strong empirical results with theoretical guarantees, does not require extensive hyper-parameter tuning, and is plug-and-play in common post-training pipelines. Together, this can detect and mitigate reward hacking behaviors while reaching a pareto-optimal point across an extremely large number of objectives. Our empirical evaluations demonstrate that CGPO significantly outperforms standard RLHF algorithms like PPO and DPO across various tasks including general chat, STEM questions, instruction following, and coding. Specifically, CGPO shows improvements of 7.4% in AlpacaEval-2 (general chat), 12.5% in Arena-Hard (STEM & reasoning), and consistent gains in other domains like math and coding. Notably, PPO, while commonly used, is prone to severe reward hacking in popular coding benchmarks, which CGPO successfully addresses. This breakthrough in RLHF not only tackles reward hacking and extreme multi-objective optimization challenges but also advances the state-of-the-art in aligning general-purpose LLMs for diverse applications.

cs.LG↗

arXiv2024

Mask-Encoded Sparsification: Mitigating Biased Gradients in Communication-Efficient Split Learning

This paper introduces a novel framework designed to achieve a high compression ratio in Split Learning (SL) scenarios where resource-constrained devices are involved in large-scale model training. Our investigations demonstrate that compressing feature maps within SL leads to biased gradients that can negatively impact the convergence rates and diminish the generalization capabilities of the resulting models. Our theoretical analysis provides insights into how compression errors critically hinder SL performance, which previous methodologies underestimate. To address these challenges, we employ a narrow bit-width encoded mask to compensate for the sparsification error without increasing the order of time complexity. Supported by rigorous theoretical analysis, our framework significantly reduces compression errors and accelerates the convergence. Extensive experiments also verify that our method outperforms existing solutions regarding training efficiency and communication complexity.

cs.LG↗

arXiv2024

HACMan: Learning Hybrid Actor-Critic Maps for 6D Non-Prehensile Manipulation

Manipulating objects without grasping them is an essential component of human dexterity, referred to as non-prehensile manipulation. Non-prehensile manipulation may enable more complex interactions with the objects, but also presents challenges in reasoning about gripper-object interactions. In this work, we introduce Hybrid Actor-Critic Maps for Manipulation (HACMan), a reinforcement learning approach for 6D non-prehensile manipulation of objects using point cloud observations. HACMan proposes a temporally-abstracted and spatially-grounded object-centric action representation that consists of selecting a contact location from the object point cloud and a set of motion parameters describing how the robot will move after making contact. We modify an existing off-policy RL algorithm to learn in this hybrid discrete-continuous action representation. We evaluate HACMan on a 6D object pose alignment task in both simulation and in the real world. On the hardest version of our task, with randomized initial poses, randomized 6D goals, and diverse object categories, our policy demonstrates strong generalization to unseen object categories without a performance drop, achieving an 89% success rate on unseen objects in simulation and 50% success rate with zero-shot transfer in the real world. Compared to alternative action representations, HACMan achieves a success rate more than three times higher than the best baseline. With zero-shot sim2real transfer, our policy can successfully manipulate unseen objects in the real world for challenging non-planar goals, using dynamic and contact-rich non-prehensile skills. Videos can be found on the project website: https://hacman-2023.github.io.

cs.RO↗

arXiv2024

Reinforcement Learning in a Safety-Embedded MDP with Trajectory Optimization

Safe Reinforcement Learning (RL) plays an important role in applying RL algorithms to safety-critical real-world applications, addressing the trade-off between maximizing rewards and adhering to safety constraints. This work introduces a novel approach that combines RL with trajectory optimization to manage this trade-off effectively. Our approach embeds safety constraints within the action space of a modified Markov Decision Process (MDP). The RL agent produces a sequence of actions that are transformed into safe trajectories by a trajectory optimizer, thereby effectively ensuring safety and increasing training stability. This novel approach excels in its performance on challenging Safety Gym tasks, achieving significantly higher rewards and near-zero safety violations during inference. The method's real-world applicability is demonstrated through a safe and effective deployment in a real robot task of box-pushing around obstacles.

cs.RO↗

arXiv2024

HACMan++: Spatially-Grounded Motion Primitives for Manipulation

Although end-to-end robot learning has shown some success for robot manipulation, the learned policies are often not sufficiently robust to variations in object pose or geometry. To improve the policy generalization, we introduce spatially-grounded parameterized motion primitives in our method HACMan++. Specifically, we propose an action representation consisting of three components: what primitive type (such as grasp or push) to execute, where the primitive will be grounded (e.g. where the gripper will make contact with the world), and how the primitive motion is executed, such as parameters specifying the push direction or grasp orientation. These three components define a novel discrete-continuous action space for reinforcement learning. Our framework enables robot agents to learn to chain diverse motion primitives together and select appropriate primitive parameters to complete long-horizon manipulation tasks. By grounding the primitives on a spatial location in the environment, our method is able to effectively generalize across object shape and pose variations. Our approach significantly outperforms existing methods, particularly in complex scenarios demanding both high-level sequential reasoning and object generalization. With zero-shot sim-to-real transfer, our policy succeeds in challenging real-world manipulation tasks, with generalization to unseen objects. Videos can be found on the project website: https://sgmp-rss2024.github.io.

cs.RO↗

arXiv2024

MuirBench: A Comprehensive Benchmark for Robust Multi-image Understanding

We introduce MuirBench, a comprehensive benchmark that focuses on robust multi-image understanding capabilities of multimodal LLMs. MuirBench consists of 12 diverse multi-image tasks (e.g., scene understanding, ordering) that involve 10 categories of multi-image relations (e.g., multiview, temporal relations). Comprising 11,264 images and 2,600 multiple-choice questions, MuirBench is created in a pairwise manner, where each standard instance is paired with an unanswerable variant that has minimal semantic differences, in order for a reliable assessment. Evaluated upon 20 recent multi-modal LLMs, our results reveal that even the best-performing models like GPT-4o and Gemini Pro find it challenging to solve MuirBench, achieving 68.0% and 49.3% in accuracy. Open-source multimodal LLMs trained on single images can hardly generalize to multi-image questions, hovering below 33.3% in accuracy. These results highlight the importance of MuirBench in encouraging the community to develop multimodal LLMs that can look beyond a single image, suggesting potential pathways for future improvements.

cs.CV↗

arXiv2024

Improving Multilingual Instruction Finetuning via Linguistically Natural and Diverse Datasets

Advancements in Large Language Models (LLMs) have significantly enhanced instruction-following capabilities. However, most Instruction Fine-Tuning (IFT) datasets are predominantly in English, limiting model performance in other languages. Traditional methods for creating multilingual IFT datasets such as translating existing English IFT datasets or converting existing NLP datasets into IFT datasets by templating, struggle to capture linguistic nuances and ensure prompt (instruction) diversity. To address this issue, we propose a novel method for collecting multilingual IFT datasets that preserves linguistic naturalness and ensures prompt diversity. This approach leverages English-focused LLMs, monolingual corpora, and a scoring function to create high-quality, diversified IFT datasets in multiple languages. Experiments demonstrate that LLMs finetuned using these IFT datasets show notable improvements in both generative and discriminative tasks, indicating enhanced language comprehension by LLMs in non-English contexts. Specifically, on the multilingual summarization task, LLMs using our IFT dataset achieved 17.57% and 15.23% improvements over LLMs fine-tuned with translation-based and template-based datasets, respectively.

cs.CL↗

arXiv2024

Contrastive Instruction Tuning

Instruction tuning has been used as a promising approach to improve the performance of large language models (LLMs) on unseen tasks. However, current LLMs exhibit limited robustness to unseen instructions, generating inconsistent outputs when the same instruction is phrased with slightly varied forms or language styles. This behavior indicates LLMs' lack of robustness to textual variations and generalizability to unseen instructions, potentially leading to trustworthiness issues. Accordingly, we propose Contrastive Instruction Tuning, which maximizes the similarity between the hidden representations of semantically equivalent instruction-instance pairs while minimizing the similarity between semantically different ones. To facilitate this approach, we augment the existing FLAN collection by paraphrasing task instructions. Experiments on the PromptBench benchmark show that CoIN consistently improves LLMs' robustness to unseen instructions with variations across character, word, sentence, and semantic levels by an average of +2.5% in accuracy. Code is available at https://github.com/luka-group/CoIN.

cs.CL↗

arXiv2024

How Fragile is Relation Extraction under Entity Replacements?

Relation extraction (RE) aims to extract the relations between entity names from the textual context. In principle, textual context determines the ground-truth relation and the RE models should be able to correctly identify the relations reflected by the textual context. However, existing work has found that the RE models memorize the entity name patterns to make RE predictions while ignoring the textual context. This motivates us to raise the question: ``are RE models robust to the entity replacements?'' In this work, we operate the random and type-constrained entity replacements over the RE instances in TACRED and evaluate the state-of-the-art RE models under the entity replacements. We observe the 30\% - 50\% F1 score drops on the state-of-the-art RE models under entity replacements. These results suggest that we need more efforts to develop effective RE models robust to entity replacements. We release the source code at https://github.com/wangywUST/RobustRE.

cs.CL↗

arXiv2024

Getting Sick After Seeing a Doctor? Diagnosing and Mitigating Knowledge Conflicts in Event Temporal Reasoning

Event temporal reasoning aims at identifying the temporal relations between two or more events from narratives. However, knowledge conflicts arise when there is a mismatch between the actual temporal relations of events in the context and the prior knowledge or biases learned by the model. In this paper, we propose to detect knowledge-conflict examples in event temporal reasoning using bias indicators, which include event relation prior bias, tense bias, narrative bias, and dependency bias. We define conflict examples as those where event relations are opposite to biased or prior relations. To mitigate event-related knowledge conflicts, we introduce a Counterfactual Data Augmentation (CDA) based method that can be applied to both Pre-trained Language Models (PLMs) and Large Language Models (LLMs) either as additional training data or demonstrations for In-Context Learning. Experiments suggest both PLMs and LLMs suffer from knowledge conflicts in event temporal reasoning, and CDA has the potential for reducing hallucination and improving model performance.

cs.CL↗

arXiv2024

Sim2Real Manipulation on Unknown Objects with Tactile-based Reinforcement Learning

Using tactile sensors for manipulation remains one of the most challenging problems in robotics. At the heart of these challenges is generalization: How can we train a tactile-based policy that can manipulate unseen and diverse objects? In this paper, we propose to perform Reinforcement Learning with only visual tactile sensing inputs on diverse objects in a physical simulator. By training with diverse objects in simulation, it enables the policy to generalize to unseen objects. However, leveraging simulation introduces the Sim2Real transfer problem. To mitigate this problem, we study different tactile representations and evaluate how each affects real-robot manipulation results after transfer. We conduct our experiments on diverse real-world objects and show significant improvements over baselines for the pivoting task. Our project page is available at https://tactilerl.github.io/.

cs.RO↗