AgentRewardBench: Evaluating Automatic Evaluations of Web Agent Trajectories

Web agents enable users to perform tasks on web browsers through natural language interaction. Evaluating web agents trajectories is an important problem, since it helps us determine whether the agent successfully completed the tasks. Rule-based methods are widely used for this purpose, but they are challenging to extend to new tasks and may not always recognize successful trajectories. We may achieve higher accuracy through human evaluation, but the process would be substantially slower and more expensive. Automatic evaluations with LLMs may avoid the challenges of designing new rules and manually annotating trajectories, enabling faster and cost-effective evaluation. However, it is unclear how effective they are at evaluating web agents. To this end, we propose AgentRewardBench, the first benchmark to assess the effectiveness of LLM judges for evaluating web agents. AgentRewardBench contains 1302 trajectories across 5 benchmarks and 4 LLMs. Each trajectory in AgentRewardBench is reviewed by an expert, who answers questions pertaining to the success, side effects, and repetitiveness of the agent. Using our benchmark, we evaluate 12 LLM judges and find that no single LLM excels across all benchmarks. We also find that the rule-based evaluation used by common benchmarks tends to underreport the success rate of web agents, highlighting a key weakness of rule-based evaluation and the need to develop more flexible automatic evaluations. We release the benchmark at: https://agent-reward-bench.github.io

Exploiting Instruction-Following Retrievers for Malicious Information Retrieval

Instruction-following retrievers have been widely adopted alongside LLMs in real-world applications, but little work has investigated the safety risks surrounding their increasing search capabilities. We empirically study the ability of retrievers to satisfy malicious queries, both when used directly and when used in a retrieval augmented generation-based setup. Concretely, we investigate six leading retrievers, including NV-Embed and LLM2Vec, and find that given malicious requests, most retrievers can (for >50% of queries) select relevant harmful passages. For example, LLM2Vec correctly selects passages for 61.35% of our malicious queries. We further uncover an emerging risk with instruction-following retrievers, where highly relevant harmful information can be surfaced by exploiting their instruction-following capabilities. Finally, we show that even safety-aligned LLMs, such as Llama3, can satisfy malicious requests when provided with harmful retrieved passages in-context. In summary, our findings underscore the malicious misuse risks associated with increasing retriever capability.

SafeArena: Evaluating the Safety of Autonomous Web Agents

LLM-based agents are becoming increasingly proficient at solving web-based tasks. With this capability comes a greater risk of misuse for malicious purposes, such as posting misinformation in an online forum or selling illicit substances on a website. To evaluate these risks, we propose SafeArena, the first benchmark to focus on the deliberate misuse of web agents. SafeArena comprises 250 safe and 250 harmful tasks across four websites. We classify the harmful tasks into five harm categories -- misinformation, illegal activity, harassment, cybercrime, and social bias, designed to assess realistic misuses of web agents. We evaluate leading LLM-based web agents, including GPT-4o, Claude-3.5 Sonnet, Qwen-2-VL 72B, and Llama-3.2 90B, on our benchmark. To systematically assess their susceptibility to harmful tasks, we introduce the Agent Risk Assessment framework that categorizes agent behavior across four risk levels. We find agents are surprisingly compliant with malicious requests, with GPT-4o and Qwen-2 completing 34.7% and 27.3% of harmful requests, respectively. Our findings highlight the urgent need for safety alignment procedures for web agents. Our benchmark is available here: https://safearena.github.io

Universal Adversarial Triggers Are Not Universal

Recent work has developed optimization procedures to find token sequences, called adversarial triggers, which can elicit unsafe responses from aligned language models. These triggers are believed to be universally transferable, i.e., a trigger optimized on one model can jailbreak other models. In this paper, we concretely show that such adversarial triggers are not universal. We extensively investigate trigger transfer amongst 13 open models and observe inconsistent transfer. Our experiments further reveal a significant difference in robustness to adversarial triggers between models Aligned by Preference Optimization (APO) and models Aligned by Fine-Tuning (AFT). We find that APO models are extremely hard to jailbreak even when the trigger is optimized directly on the model. On the other hand, while AFT models may appear safe on the surface, exhibiting refusals to a range of unsafe instructions, we show that they are highly susceptible to adversarial triggers. Lastly, we observe that most triggers optimized on AFT models also generalize to new unsafe instructions from five diverse domains, further emphasizing their vulnerability. Overall, our work highlights the need for more comprehensive safety evaluations for aligned language models.

Evaluating Correctness and Faithfulness of Instruction-Following Models for Question Answering

Retriever-augmented instruction-following models are attractive alternatives to fine-tuned approaches for information-seeking tasks such as question answering (QA). By simply prepending retrieved documents in its input along with an instruction, these models can be adapted to various information domains and tasks without additional fine-tuning. While the model responses tend to be natural and fluent, the additional verbosity makes traditional QA evaluation metrics such as exact match (EM) and F1 unreliable for accurately quantifying model performance. In this work, we investigate the performance of instruction-following models across three information-seeking QA tasks. We use both automatic and human evaluation to evaluate these models along two dimensions: 1) how well they satisfy the user's information need (correctness), and 2) whether they produce a response based on the provided knowledge (faithfulness). Guided by human evaluation and analysis, we highlight the shortcomings of traditional metrics for both correctness and faithfulness. We then propose simple token-overlap based and model-based metrics that reflect the true performance of these models. Our analysis reveals that instruction-following models are competitive, and sometimes even outperform fine-tuned models for correctness. However, these models struggle to stick to the provided knowledge and often hallucinate in their responses. We hope our work encourages a more holistic evaluation of instruction-following models for QA. Our code and data is available at https://github.com/McGill-NLP/instruct-qa

StarCoder: may the source be with you!

The BigCode community, an open-scientific collaboration working on the responsible development of Large Language Models for Code (Code LLMs), introduces StarCoder and StarCoderBase: 15.5B parameter models with 8K context length, infilling capabilities and fast large-batch inference enabled by multi-query attention. StarCoderBase is trained on 1 trillion tokens sourced from The Stack, a large collection of permissively licensed GitHub repositories with inspection tools and an opt-out process. We fine-tuned StarCoderBase on 35B Python tokens, resulting in the creation of StarCoder. We perform the most comprehensive evaluation of Code LLMs to date and show that StarCoderBase outperforms every open Code LLM that supports multiple programming languages and matches or outperforms the OpenAI code-cushman-001 model. Furthermore, StarCoder outperforms every model that is fine-tuned on Python, can be prompted to achieve 40\% pass@1 on HumanEval, and still retains its performance on other programming languages. We take several important steps towards a safe open-access model release, including an improved PII redaction pipeline and a novel attribution tracing tool, and make the StarCoder models publicly available under a more commercially viable version of the Open Responsible AI Model license.

Using In-Context Learning to Improve Dialogue Safety

While large neural-based conversational models have become increasingly proficient as dialogue agents, recent work has highlighted safety issues with these systems. For example, these systems can be goaded into generating toxic content, which often perpetuates social biases or stereotypes. We investigate a retrieval-based framework for reducing bias and toxicity in responses generated from neural-based chatbots. It uses in-context learning to steer a model towards safer generations. Concretely, to generate a response to an unsafe dialogue context, we retrieve demonstrations of safe model responses to similar dialogue contexts. We find our proposed approach performs competitively with strong baselines which use fine-tuning. For instance, using automatic evaluation, we find our best fine-tuned baseline only generates safe responses to unsafe dialogue contexts from DiaSafety 2.92% more than our approach. Finally, we also propose a straightforward re-ranking procedure which can further improve response safeness.

An Empirical Survey of the Effectiveness of Debiasing Techniques for Pre-Trained Language Models

Recent work has shown that pre-trained language models capture social biases from the text corpora they are trained on. This has attracted attention to developing techniques that mitigate such biases. In this work, we perform a empirical survey of five recently proposed debiasing techniques: Counterfactual Data Augmentation (CDA), Dropout, Iterative Nullspace Projection, Self-Debias, and SentenceDebias. We quantify the effectiveness of each technique using three different bias benchmarks while also measuring the impact of these techniques on a model's language modeling ability, as well as its performance on downstream NLU tasks. We experimentally find that: (1) CDA and Self-Debias are the strongest of the debiasing techniques, obtaining improved scores on most of the bias benchmarks (2) Current debiasing techniques do not generalize well beyond gender bias; And (3) improvements on bias benchmarks such as StereoSet and CrowS-Pairs by using debiasing strategies are usually accompanied by a decrease in language modeling ability, making it difficult to determine whether the bias mitigation is effective.

Evaluating the Faithfulness of Importance Measures in NLP by Recursively Masking Allegedly Important Tokens and Retraining

To explain NLP models, many methods inform which inputs tokens are important for a prediction. However, an open question is if these methods accurately reflect the model's logic, a property often called faithfulness. In this work, we adapt and improve a recently proposed faithfulness benchmark from computer vision called ROAR (RemOve And Retrain), by Hooker et al. (2019). We improve ROAR by recursively removing dataset redundancies, which otherwise interfere with ROAR. We adapt and apply ROAR, to popular NLP importance measures, namely attention, gradient, and integrated gradients. Additionally, we use mutual information as an additional baseline. Evaluation is done on a suite of classification tasks often used in the faithfulness of attention literature. Finally, we propose a scalar faithfulness metric, which makes it easy to compare results across papers. We find that, importance measures considered to be unfaithful for computer vision tasks perform favorably for NLP tasks, the faithfulness of an importance measure is task-dependent, and the computational overhead of integrated gradient is rarely justified.

Exploring Conditioning for Generative Music Systems with Human-Interpretable Controls

Performance RNN is a machine-learning system designed primarily for the generation of solo piano performances using an event-based (rather than audio) representation. More specifically, Performance RNN is a long short-term memory (LSTM) based recurrent neural network that models polyphonic music with expressive timing and dynamics (Oore et al., 2018). The neural network uses a simple language model based on the Musical Instrument Digital Interface (MIDI) file format. Performance RNN is trained on the e-Piano Junior Competition Dataset (International Piano e-Competition, 2018), a collection of solo piano performances by expert pianists. As an artistic tool, one of the limitations of the original model has been the lack of useable controls. The standard form of Performance RNN can generate interesting pieces, but little control is provided over what specifically is generated. This paper explores a set of conditioning-based controls used to influence the generation process.