Ateret Anaby-Tavor

George Kour, Marcel Zalmanovici, Naama Zwerdling et al. · ibm-research

As large language models become more prevalent, their possible harmful or inappropriate responses are a cause for concern. This paper introduces a unique dataset containing adversarial examples in the form of questions, which we call AttaQ, designed to provoke such harmful or inappropriate responses. We assess the efficacy of our dataset by analyzing the vulnerabilities of various models when subjected to it. Additionally, we introduce a novel automatic approach for identifying and naming vulnerable semantic regions - input semantic areas for which the model is likely to produce harmful outputs. This is achieved through the application of specialized clustering techniques that consider both the semantic similarity of the input attacks and the harmfulness of the model's responses. Automatically identifying vulnerable semantic regions enhances the evaluation of model weaknesses, facilitating targeted improvements to its safety mechanisms and overall reliability.

George Kour, Samuel Ackerman, Orna Raz et al. · ibm-research

The ability to compare the semantic similarity between text corpora is important in a variety of natural language processing applications. However, standard methods for evaluating these metrics have yet to be established. We propose a set of automatic and interpretable measures for assessing the characteristics of corpus-level semantic similarity metrics, allowing sensible comparison of their behavior. We demonstrate the effectiveness of our evaluation measures in capturing fundamental characteristics by evaluating them on a collection of classical and state-of-the-art metrics. Our measures revealed that recently-developed metrics are becoming better in identifying semantic distributional mismatch while classical metrics are more sensitive to perturbations in the surface text levels.

Ella Rabinovich, Matan Vetzler, David Boaz et al. · ibm-research

The rapidly growing market demand for automatic dialogue agents capable of goal-oriented behavior has caused many tech-industry leaders to invest considerable efforts into task-oriented dialog systems. The success of these systems is highly dependent on the accuracy of their intent identification -- the process of deducing the goal or meaning of the user's request and mapping it to one of the known intents for further processing. Gaining insights into unrecognized utterances -- user requests the systems fail to attribute to a known intent -- is therefore a key process in continuous improvement of goal-oriented dialog systems. We present an end-to-end pipeline for processing unrecognized user utterances, deployed in a real-world, commercial task-oriented dialog system, including a specifically-tailored clustering algorithm, a novel approach to cluster representative extraction, and cluster naming. We evaluated the proposed components, demonstrating their benefits in the analysis of unrecognized user requests.

Itay Nakash, George Kour, Ateret Anaby-Tavor · ibm-research

Large language models (LLMs) are increasingly deployed as customer-facing agents, yet evaluating their reliability remains challenging due to stochastic, multi-turn interactions. Current evaluation protocols rely on linear Monte Carlo rollouts of complete agent-user conversations to estimate success. However, this approach is computationally inefficient, repeatedly regenerating identical early prefixes, and often fails to uncover deep failure modes that arise from rare user behaviors. We introduce DIVERT (Diversity-Induced Evaluation via Branching of Trajectories), an efficient, snapshot-based, coverage-guided user simulation framework for systematic exploration of agent-user interactions. DIVERT captures the full agent-environment state at critical decision points and resumes execution from these snapshots, enabling reuse of shared conversation prefixes and reducing redundant computation. From each junction, the framework branches using targeted, diversity-inducing user responses, allowing directed exploration of alternative interaction paths. By focusing evaluation on semantically diverse and underexplored trajectories, DIVERT improves both efficiency and coverage. Empirical results show that it discovers more failures per token compared to standard linear rollout protocols, while expanding the set of tasks on which failures are identified.

Ella Rabinovich, Samuel Ackerman, Orna Raz et al.

Semantic consistency of a language model is broadly defined as the model's ability to produce semantically-equivalent outputs, given semantically-equivalent inputs. We address the task of assessing question-answering (QA) semantic consistency of contemporary large language models (LLMs) by manually creating a benchmark dataset with high-quality paraphrases for factual questions, and release the dataset to the community. We further combine the semantic consistency metric with additional measurements suggested in prior work as correlating with LLM QA accuracy, for building and evaluating a framework for factual QA reference-less performance prediction -- predicting the likelihood of a language model to accurately answer a question. Evaluating the framework on five contemporary LLMs, we demonstrate encouraging, significantly outperforming baselines, results.

George Kour, Naama Zwerdling, Marcel Zalmanovici et al. · ibm-research

Large language models (LLMs) are increasingly used in business dialogue systems but they pose security and ethical risks. Multi-turn conversations, where context influences the model's behavior, can be exploited to produce undesired responses. In this paper, we examine the effectiveness of utilizing off-the-shelf LLMs in straightforward red-teaming approaches, where an attacker LLM aims to elicit undesired output from a target LLM, comparing both single-turn and conversational red-teaming tactics. Our experiments offer insights into various usage strategies that significantly affect their performance as red teamers. They suggest that off-the-shelf models can act as effective red teamers and even adjust their attack strategy based on past attempts, although their effectiveness decreases with greater alignment.

Naama Zwerdling, Segev Shlomov, Esther Goldbraich et al. · ibm-research

Models for text generation have become focal for many research tasks and especially for the generation of sentence corpora. However, understanding the properties of an automatically generated text corpus remains challenging. We propose a set of tools that examine the properties of generated text corpora. Applying these tools on various generated corpora allowed us to gain new insights into the properties of the generative models. As part of our characterization process, we found remarkable differences in the corpora generated by two leading generative technologies.

Samuel Ackerman, Ella Rabinovich, Eitan Farchi et al.

We evaluate the robustness of several large language models on multiple datasets. Robustness here refers to the relative insensitivity of the model's answers to meaning-preserving variants of their input. Benchmark datasets are constructed by introducing naturally-occurring, non-malicious perturbations, or by generating semantically equivalent paraphrases of input questions or statements. We further propose a novel metric for assessing a model robustness, and demonstrate its benefits in the non-adversarial scenario by empirical evaluation of several models on the created datasets.

Swapnaja Achintalwar, Adriana Alvarado Garcia, Ateret Anaby-Tavor et al. · ibm-research

Large language models (LLMs) are susceptible to a variety of risks, from non-faithful output to biased and toxic generations. Due to several limiting factors surrounding LLMs (training cost, API access, data availability, etc.), it may not always be feasible to impose direct safety constraints on a deployed model. Therefore, an efficient and reliable alternative is required. To this end, we present our ongoing efforts to create and deploy a library of detectors: compact and easy-to-build classification models that provide labels for various harms. In addition to the detectors themselves, we discuss a wide range of uses for these detector models - from acting as guardrails to enabling effective AI governance. We also deep dive into inherent challenges in their development and discuss future work aimed at making the detectors more reliable and broadening their scope.

Itay Nakash, George Kour, Guy Uziel et al. · ibm-research

Following the advancement of large language models (LLMs), the development of LLM-based autonomous agents has become increasingly prevalent. As a result, the need to understand the security vulnerabilities of these agents has become a critical task. We examine how ReAct agents can be exploited using a straightforward yet effective method we refer to as the foot-in-the-door attack. Our experiments show that indirect prompt injection attacks, prompted by harmless and unrelated requests (such as basic calculations) can significantly increase the likelihood of the agent performing subsequent malicious actions. Our results show that once a ReAct agents thought includes a specific tool or action, the likelihood of executing this tool in the subsequent steps increases significantly, as the agent seldom re-evaluates its actions. Consequently, even random, harmless requests can establish a foot-in-the-door, allowing an attacker to embed malicious instructions into the agents thought process, making it more susceptible to harmful directives. To mitigate this vulnerability, we propose implementing a simple reflection mechanism that prompts the agent to reassess the safety of its actions during execution, which can help reduce the success of such attacks.

Eran Hirsch, Guy Uziel, Ateret Anaby-Tavor

Planning is a fundamental task in artificial intelligence that involves finding a sequence of actions that achieve a specified goal in a given environment. Large language models (LLMs) are increasingly used for applications that require planning capabilities, such as web or embodied agents. In line with recent studies, we demonstrate through experimentation that LLMs lack necessary skills required for planning. Based on these observations, we advocate for the potential of a hybrid approach that combines LLMs with classical planning methodology. Then, we introduce SimPlan, a novel hybrid-method, and evaluate its performance in a new challenging setup. Our extensive experiments across various planning domains demonstrate that SimPlan significantly outperforms existing LLM-based planners.

Koren Lazar, Matan Vetzler, Guy Uziel et al.

In the digital era, the widespread use of APIs is evident. However, scalable utilization of APIs poses a challenge due to structure divergence observed in online API documentation. This underscores the need for automatic tools to facilitate API consumption. A viable approach involves the conversion of documentation into an API Specification format. While previous attempts have been made using rule-based methods, these approaches encountered difficulties in generalizing across diverse documentation. In this paper we introduce SpeCrawler, a comprehensive system that utilizes large language models (LLMs) to generate OpenAPI Specifications from diverse API documentation through a carefully crafted pipeline. By creating a standardized format for numerous APIs, SpeCrawler aids in streamlining integration processes within API orchestrating systems and facilitating the incorporation of tools into LLMs. The paper explores SpeCrawler's methodology, supported by empirical evidence and case studies, demonstrating its efficacy through LLM capabilities.

Ella Rabinovich, Ateret Anaby-Tavor

Large Language Models (LLMs) are increasingly acting as autonomous agents, with function calling (FC) capabilities enabling them to invoke specific tools for tasks. While prior research has primarily focused on improving FC accuracy, little attention has been given to the robustness of these agents to perturbations in their input. We introduce a benchmark assessing FC robustness in two key areas: resilience to naturalistic query variations, and stability in function calling when the toolkit expands with semantically related tools. Evaluating best-performing FC models on a carefully expanded subset of the Berkeley function calling leaderboard (BFCL), we identify critical weaknesses in existing evaluation methodologies, and highlight areas for improvement in real-world agentic deployments.

Itay Nakash, George Kour, Koren Lazar et al. · ibm-research

Task-oriented LLM-based agents are increasingly used in domains with strict policies, such as refund eligibility or cancellation rules. The challenge lies in ensuring that the agent consistently adheres to these rules and policies, appropriately refusing any request that would violate them, while still maintaining a helpful and natural interaction. This calls for the development of tailored design and evaluation methodologies to ensure agent resilience against malicious user behavior. We propose a novel threat model that focuses on adversarial users aiming to exploit policy-adherent agents for personal benefit. To address this, we present CRAFT, a multi-agent red-teaming system that leverages policy-aware persuasive strategies to undermine a policy-adherent agent in a customer-service scenario, outperforming conventional jailbreak methods such as DAN prompts, emotional manipulation, and coercive. Building upon the existing tau-bench benchmark, we introduce tau-break, a complementary benchmark designed to rigorously assess the agent's robustness against manipulative user behavior. Finally, we evaluate several straightforward yet effective defense strategies. While these measures provide some protection, they fall short, highlighting the need for stronger, research-driven safeguards to protect policy-adherent agents from adversarial attacks

Ella Rabinovich, David Boaz, Naama Zwerdling et al.

Agentic systems for business process automation often require compliance with policies governing conditional updates to the system state. Evaluation of policy adherence in LLM-based agentic workflows is typically performed by comparing the final system state against a predefined ground truth. While this approach detects explicit policy violations, it may overlook a more subtle class of issues in which agents bypass required policy checks, yet reach a correct outcome due to favorable circumstances. We refer to such cases as $\textit{near-misses}$ or $\textit{latent failures}$. In this work, we introduce a novel metric for detecting latent policy failures in agent conversations traces. Building on the ToolGuard framework, which converts natural-language policies into executable guard code, our method analyzes agent trajectories to determine whether agent's tool-calling decisions where sufficiently informed. We evaluate our approach on the $τ^2$-verified Airlines benchmark across several contemporary open and proprietary LLMs acting as agents. Our results show that latent failures occur in 8-17% of trajectories involving mutating tool calls, even when the final outcome matches the expected ground-truth state. These findings reveal a blind spot in current evaluation methodologies and highlight the need for metrics that assess not only final outcomes but also the decision process leading to them.

Naama Zwerdling, David Boaz, Ella Rabinovich et al.

Large Language Model (LLM) agents hold promise for a flexible and scalable alternative to traditional business process automation, but struggle to reliably follow complex company policies. In this study we introduce a deterministic, transparent, and modular framework for enforcing business policy adherence in agentic workflows. Our method operates in two phases: (1) an offline buildtime stage that compiles policy documents into verifiable guard code associated with tool use, and (2) a runtime integration where these guards ensure compliance before each agent action. We demonstrate our approach on the challenging $τ$-bench Airlines domain, showing encouraging preliminary results in policy enforcement, and further outline key challenges for real-world deployments.

Matan Vetzler, Koren Lazar, Guy Uziel et al.

Recent advancements in large language models (LLMs) underscore the need for stronger reasoning capabilities to solve complex problems effectively. While Chain-of-Thought (CoT) reasoning has been a step forward, it remains insufficient for many domains. A promising alternative is explicit high-level plan generation, but existing approaches largely assume that LLMs can produce effective plans through few-shot prompting alone, without additional training. In this work, we challenge this assumption and introduce CRISP (Complex Reasoning with Interpretable Step-based Plans), a multi-domain dataset of high-level plans for mathematical reasoning and code generation. The plans in CRISP are automatically generated and rigorously validated--both intrinsically, using an LLM as a judge, and extrinsically, by evaluating their impact on downstream task performance. We demonstrate that fine-tuning a small model on CRISP enables it to generate higher-quality plans than much larger models using few-shot prompting, while significantly outperforming Chain-of-Thought reasoning. Furthermore, our out-of-domain evaluation reveals that fine-tuning on one domain improves plan generation in the other, highlighting the generalizability of learned planning capabilities.

George Kour, Itay Nakash, Ateret Anaby-Tavor et al. · ibm-research

As Large Language Models (LLMs) become deeply integrated into human life and increasingly influence decision-making, it's crucial to evaluate whether and to what extent they exhibit subjective preferences, opinions, and beliefs. These tendencies may stem from biases within the models, which may shape their behavior, influence the advice and recommendations they offer to users, and potentially reinforce certain viewpoints. This paper presents the Preference, Opinion, and Belief survey (POBs), a benchmark developed to assess LLMs' subjective inclinations across societal, cultural, ethical, and personal domains. We applied our benchmark to evaluate leading open- and closed-source LLMs, measuring desired properties such as reliability, neutrality, and consistency. In addition, we investigated the effect of increasing the test-time compute, through reasoning and self-reflection mechanisms, on those metrics. While effective in other tasks, our results show that these mechanisms offer only limited gains in our domain. Furthermore, we reveal that newer model versions are becoming less consistent and more biased toward specific viewpoints, highlighting a blind spot and a concerning trend. POBS: https://ibm.github.io/POBS

Ella Rabinovich, Matan Vetzler, Samuel Ackerman et al.

Data drift is the change in model input data that is one of the key factors leading to machine learning models performance degradation over time. Monitoring drift helps detecting these issues and preventing their harmful consequences. Meaningful drift interpretation is a fundamental step towards effective re-training of the model. In this study we propose an end-to-end framework for reliable model-agnostic change-point detection and interpretation in large task-oriented dialog systems, proven effective in multiple customer deployments. We evaluate our approach and demonstrate its benefits with a novel variant of intent classification training dataset, simulating customer requests to a dialog system. We make the data publicly available.

George Kour, Marcel Zalmanovici, Orna Raz et al.

Testing Machine Learning (ML) models and AI-Infused Applications (AIIAs), or systems that contain ML models, is highly challenging. In addition to the challenges of testing classical software, it is acceptable and expected that statistical ML models sometimes output incorrect results. A major challenge is to determine when the level of incorrectness, e.g., model accuracy or F1 score for classifiers, is acceptable and when it is not. In addition to business requirements that should provide a threshold, it is a best practice to require any proposed ML solution to out-perform simple baseline models, such as a decision tree. We have developed complexity measures, which quantify how difficult given observations are to assign to their true class label; these measures can then be used to automatically determine a baseline performance threshold. These measures are superior to the best practice baseline in that, for a linear computation cost, they also quantify each observation' classification complexity in an explainable form, regardless of the classifier model used. Our experiments with both numeric synthetic data and real natural language chatbot data demonstrate that the complexity measures effectively highlight data regions and observations that are likely to be misclassified.

Eyal Ben-David, Boaz Carmeli, Ateret Anaby-Tavor

Goal oriented dialogue systems have become a prominent customer-care interaction channel for most businesses. However, not all interactions are smooth, and customer intent misunderstanding is a major cause of dialogue failure. We show that intent prediction can be improved by training a deep text-to-text neural model to generate successive user utterances from unlabeled dialogue data. For that, we define a multi-task training regime that utilizes successive user-utterance generation to improve the intent prediction. Our approach achieves the reported improvement due to two complementary factors: First, it uses a large amount of unlabeled dialogue data for an auxiliary generation task. Second, it uses the generated user utterance as an additional signal for the intent prediction model. Lastly, we present a novel look-ahead approach that uses user utterance generation to improve intent prediction in inference time. Specifically, we generate counterfactual successive user utterances for conversations with ambiguous predicted intents, and disambiguate the prediction by reassessing the concatenated sequence of available and generated utterances.

Ofer Lavi, Ella Rabinovich, Segev Shlomov et al.

Dialog is a core building block of human natural language interactions. It contains multi-party utterances used to convey information from one party to another in a dynamic and evolving manner. The ability to compare dialogs is beneficial in many real world use cases, such as conversation analytics for contact center calls and virtual agent design. We propose a novel adaptation of the edit distance metric to the scenario of dialog similarity. Our approach takes into account various conversation aspects such as utterance semantics, conversation flow, and the participants. We evaluate this new approach and compare it to existing document similarity measures on two publicly available datasets. The results demonstrate that our method outperforms the other approaches in capturing dialog flow, and is better aligned with the human perception of conversation similarity.

Ateret Anaby-Tavor, Boaz Carmeli, Esther Goldbraich et al.

Based on recent advances in natural language modeling and those in text generation capabilities, we propose a novel data augmentation method for text classification tasks. We use a powerful pre-trained neural network model to artificially synthesize new labeled data for supervised learning. We mainly focus on cases with scarce labeled data. Our method, referred to as language-model-based data augmentation (LAMBADA), involves fine-tuning a state-of-the-art language generator to a specific task through an initial training phase on the existing (usually small) labeled data. Using the fine-tuned model and given a class label, new sentences for the class are generated. Our process then filters these new sentences by using a classifier trained on the original data. In a series of experiments, we show that LAMBADA improves classifiers' performance on a variety of datasets. Moreover, LAMBADA significantly improves upon the state-of-the-art techniques for data augmentation, specifically those applicable to text classification tasks with little data.