Improving Reinforcement Learning from Human Feedback Using Contrastive Rewards

Reinforcement learning from human feedback (RLHF) is the mainstream paradigm used to align large language models (LLMs) with human preferences. Yet existing RLHF heavily relies on accurate and informative reward models, which are vulnerable and sensitive to noise from various sources, e.g. human labeling errors, making the pipeline fragile. In this work, we improve the effectiveness of the reward model by introducing a penalty term on the reward, named as \textit{contrastive rewards}. %Contrastive rewards Our approach involves two steps: (1) an offline sampling step to obtain responses to prompts that serve as baseline calculation and (2) a contrastive reward calculated using the baseline responses and used in the Proximal Policy Optimization (PPO) step. We show that contrastive rewards enable the LLM to penalize reward uncertainty, improve robustness, encourage improvement over baselines, calibrate according to task difficulty, and reduce variance in PPO. We show empirically contrastive rewards can improve RLHF substantially, evaluated by both GPTs and humans, and our method consistently outperforms strong baselines.

Steering LLMs Towards Unbiased Responses: A Causality-Guided Debiasing Framework

Large language models (LLMs) can easily generate biased and discriminative responses. As LLMs tap into consequential decision-making (e.g., hiring and healthcare), it is of crucial importance to develop strategies to mitigate these biases. This paper focuses on social bias, tackling the association between demographic information and LLM outputs. We propose a causality-guided debiasing framework that utilizes causal understandings of (1) the data-generating process of the training corpus fed to LLMs, and (2) the internal reasoning process of LLM inference, to guide the design of prompts for debiasing LLM outputs through selection mechanisms. Our framework unifies existing de-biasing prompting approaches such as inhibitive instructions and in-context contrastive examples, and sheds light on new ways of debiasing by encouraging bias-free reasoning. Our strong empirical performance on real-world datasets demonstrates that our framework provides principled guidelines on debiasing LLM outputs even with only the black-box access.

StableToolBench: Towards Stable Large-Scale Benchmarking on Tool Learning of Large Language Models

Large Language Models (LLMs) have witnessed remarkable advancements in recent years, prompting the exploration of tool learning, which integrates LLMs with external tools to address diverse real-world challenges. Assessing the capability of LLMs to utilise tools necessitates large-scale and stable benchmarks. However, previous works relied on either hand-crafted online tools with limited scale, or large-scale real online APIs suffering from instability of API status. To address this problem, we introduce StableToolBench, a benchmark evolving from ToolBench, proposing a virtual API server and stable evaluation system. The virtual API server contains a caching system and API simulators which are complementary to alleviate the change in API status. Meanwhile, the stable evaluation system designs solvable pass and win rates using GPT-4 as the automatic evaluator to eliminate the randomness during evaluation. Experimental results demonstrate the stability of StableToolBench, and further discuss the effectiveness of API simulators, the caching system, and the evaluator system.

Learning to Watermark LLM-generated Text via Reinforcement Learning

We study how to watermark LLM outputs, i.e. embedding algorithmically detectable signals into LLM-generated text to track misuse. Unlike the current mainstream methods that work with a fixed LLM, we expand the watermark design space by including the LLM tuning stage in the watermark pipeline. While prior works focus on token-level watermark that embeds signals into the output, we design a model-level watermark that embeds signals into the LLM weights, and such signals can be detected by a paired detector. We propose a co-training framework based on reinforcement learning that iteratively (1) trains a detector to detect the generated watermarked text and (2) tunes the LLM to generate text easily detectable by the detector while keeping its normal utility. We empirically show that our watermarks are more accurate, robust, and adaptable (to new attacks). It also allows watermarked model open-sourcing. In addition, if used together with alignment, the extra overhead introduced is low - only training an extra reward model (i.e. our detector). We hope our work can bring more effort into studying a broader watermark design that is not limited to working with a fixed LLM. We open-source the code: https://github.com/xiaojunxu/learning-to-watermark-llm .

ToolRerank: Adaptive and Hierarchy-Aware Reranking for Tool Retrieval

Tool learning aims to extend the capabilities of large language models (LLMs) with external tools. A major challenge in tool learning is how to support a large number of tools, including unseen tools. To address this challenge, previous studies have proposed retrieving suitable tools for the LLM based on the user query. However, previously proposed methods do not consider the differences between seen and unseen tools, nor do they take the hierarchy of the tool library into account, which may lead to suboptimal performance for tool retrieval. Therefore, to address the aforementioned issues, we propose ToolRerank, an adaptive and hierarchy-aware reranking method for tool retrieval to further refine the retrieval results. Specifically, our proposed ToolRerank includes Adaptive Truncation, which truncates the retrieval results related to seen and unseen tools at different positions, and Hierarchy-Aware Reranking, which makes retrieval results more concentrated for single-tool queries and more diverse for multi-tool queries. Experimental results show that ToolRerank can improve the quality of the retrieval results, leading to better execution results generated by the LLM.

SuPRA: Surgical Phase Recognition and Anticipation for Intra-Operative Planning

Intra-operative recognition of surgical phases holds significant potential for enhancing real-time contextual awareness in the operating room. However, we argue that online recognition, while beneficial, primarily lends itself to post-operative video analysis due to its limited direct impact on the actual surgical decisions and actions during ongoing procedures. In contrast, we contend that the prediction and anticipation of surgical phases are inherently more valuable for intra-operative assistance, as they can meaningfully influence a surgeon's immediate and long-term planning by providing foresight into future steps. To address this gap, we propose a dual approach that simultaneously recognises the current surgical phase and predicts upcoming ones, thus offering comprehensive intra-operative assistance and guidance on the expected remaining workflow. Our novel method, Surgical Phase Recognition and Anticipation (SuPRA), leverages past and current information for accurate intra-operative phase recognition while using future segments for phase prediction. This unified approach challenges conventional frameworks that treat these objectives separately. We have validated SuPRA on two reputed datasets, Cholec80 and AutoLaparo21, where it demonstrated state-of-the-art performance with recognition accuracies of 91.8% and 79.3%, respectively. Additionally, we introduce and evaluate our model using new segment-level evaluation metrics, namely Edit and F1 Overlap scores, for a more temporal assessment of segment classification. In conclusion, SuPRA presents a new multi-task approach that paves the way for improved intra-operative assistance through surgical phase recognition and prediction of future events.

Overcoming Reward Overoptimization via Adversarial Policy Optimization with Lightweight Uncertainty Estimation

We introduce Adversarial Policy Optimization (AdvPO), a novel solution to the pervasive issue of reward over-optimization in Reinforcement Learning from Human Feedback (RLHF) for Large Language Models (LLMs). Over-optimization occurs when a reward model serves as an imperfect proxy for human preference, and RL-driven policy optimization erroneously exploits reward inaccuracies. In this paper, we begin by introducing a lightweight way to quantify uncertainties in rewards, relying solely on the last layer embeddings of the reward model, without the need for computationally expensive reward ensembles. AdvPO then addresses a distributionally robust optimization problem centred around the confidence interval of the reward model's predictions for policy improvement. Through comprehensive experiments on the Anthropic HH and TL;DR summarization datasets, we illustrate the efficacy of AdvPO in mitigating the overoptimization issue, consequently resulting in enhanced performance as evaluated through human-assisted evaluation.

A Concept-based Interpretable Model for the Diagnosis of Choroid Neoplasias using Multimodal Data

Diagnosing rare diseases presents a common challenge in clinical practice, necessitating the expertise of specialists for accurate identification. The advent of machine learning offers a promising solution, while the development of such technologies is hindered by the scarcity of data on rare conditions and the demand for models that are both interpretable and trustworthy in a clinical context. Interpretable AI, with its capacity for human-readable outputs, can facilitate validation by clinicians and contribute to medical education. In the current work, we focus on choroid neoplasias, the most prevalent form of eye cancer in adults, albeit rare with 5.1 per million. We built the so-far largest dataset consisting of 750 patients, incorporating three distinct imaging modalities collected from 2004 to 2022. Our work introduces a concept-based interpretable model that distinguishes between three types of choroidal tumors, integrating insights from domain experts via radiological reports. Remarkably, this model not only achieves an F1 score of 0.91, rivaling that of black-box models, but also boosts the diagnostic accuracy of junior doctors by 42%. This study highlights the significant potential of interpretable machine learning in improving the diagnosis of rare diseases, laying a groundwork for future breakthroughs in medical AI that could tackle a wider array of complex health scenarios.

Towards Multimodal Sentiment Analysis Debiasing via Bias Purification

Multimodal Sentiment Analysis (MSA) aims to understand human intentions by integrating emotion-related clues from diverse modalities, such as visual, language, and audio. Unfortunately, the current MSA task invariably suffers from unplanned dataset biases, particularly multimodal utterance-level label bias and word-level context bias. These harmful biases potentially mislead models to focus on statistical shortcuts and spurious correlations, causing severe performance bottlenecks. To alleviate these issues, we present a Multimodal Counterfactual Inference Sentiment (MCIS) analysis framework based on causality rather than conventional likelihood. Concretely, we first formulate a causal graph to discover harmful biases from already-trained vanilla models. In the inference phase, given a factual multimodal input, MCIS imagines two counterfactual scenarios to purify and mitigate these biases. Then, MCIS can make unbiased decisions from biased observations by comparing factual and counterfactual outcomes. We conduct extensive experiments on several standard MSA benchmarks. Qualitative and quantitative results show the effectiveness of the proposed framework.

On the Effectiveness of Distillation in Mitigating Backdoors in Pre-trained Encoder

In this paper, we study a defense against poisoned encoders in SSL called distillation, which is a defense used in supervised learning originally. Distillation aims to distill knowledge from a given model (a.k.a the teacher net) and transfer it to another (a.k.a the student net). Now, we use it to distill benign knowledge from poisoned pre-trained encoders and transfer it to a new encoder, resulting in a clean pre-trained encoder. In particular, we conduct an empirical study on the effectiveness and performance of distillation against poisoned encoders. Using two state-of-the-art backdoor attacks against pre-trained image encoders and four commonly used image classification datasets, our experimental results show that distillation can reduce attack success rate from 80.87% to 27.51% while suffering a 6.35% loss in accuracy. Moreover, we investigate the impact of three core components of distillation on performance: teacher net, student net, and distillation loss. By comparing 4 different teacher nets, 3 student nets, and 6 distillation losses, we find that fine-tuned teacher nets, warm-up-training-based student nets, and attention-based distillation loss perform best, respectively.