Dinesh Raghu

Sonam Gupta, Yatin Nandwani, Asaf Yehudai et al. · ibm-research

Fine-tuning Large Language Models (LLMs) on specific datasets is a common practice to improve performance on target tasks. However, this performance gain often leads to overfitting, where the model becomes too specialized in either the task or the characteristics of the training data, resulting in a loss of generalization. This paper introduces Selective Self-Rehearsal (SSR), a fine-tuning approach that achieves performance comparable to the standard supervised fine-tuning (SFT) while improving generalization. SSR leverages the fact that there can be multiple valid responses to a query. By utilizing the model's correct responses, SSR reduces model specialization during the fine-tuning stage. SSR first identifies the correct model responses from the training set by deploying an appropriate LLM as a judge. Then, it fine-tunes the model using the correct model responses and the gold response for the remaining samples. The effectiveness of SSR is demonstrated through experiments on the task of identifying unanswerable queries across various datasets. The results show that standard SFT can lead to an average performance drop of up to $16.7\%$ on multiple benchmarks, such as MMLU and TruthfulQA. In contrast, SSR results in close to $2\%$ drop on average, indicating better generalization capabilities compared to standard SFT.

Mayank Mishra, Danish Contractor, Dinesh Raghu · ibm-research

Traditional systems designed for task oriented dialog utilize knowledge present only in structured knowledge sources to generate responses. However, relevant information required to generate responses may also reside in unstructured sources, such as documents. Recent state of the art models such as HyKnow and SeKnow aimed at overcoming these challenges make limiting assumptions about the knowledge sources. For instance, these systems assume that certain types of information, such as a phone number, is always present in a structured knowledge base (KB) while information about aspects such as entrance ticket prices, would always be available in documents. In this paper, we create a modified version of the MutliWOZ-based dataset prepared by SeKnow to demonstrate how current methods have significant degradation in performance when strict assumptions about the source of information are removed. Then, in line with recent work exploiting pre-trained language models, we fine-tune a BART based model using prompts for the tasks of querying knowledge sources, as well as, for response generation, without making assumptions about the information present in each knowledge source. Through a series of experiments, we demonstrate that our model is robust to perturbations to knowledge modality (source of information), and that it can fuse information from structured as well as unstructured knowledge to generate responses.

Saeel Sandeep Nachane, Ojas Gramopadhye, Prateek Chanda et al. · ibm-research

In this paper, we propose a modified version of the MedQA-USMLE dataset, named MEDQA-OPEN, which contains open-ended medical questions without options to mimic clinical scenarios, along with clinician-approved reasoned answers. Additionally, we implement a prompt driven by Chain of Thought (CoT) reasoning, CLINICR, to mirror the prospective process of incremental reasoning, reaching a correct response to medical questions. We empirically demonstrate how CLINICR outperforms the state-of-the-art 5-shot CoT-based prompt (Liévin et al., 2022). We also present an approach that mirrors real-life clinical practice by first exploring multiple differential diagnoses through MCQ-CLINICR and subsequently narrowing down to a final diagnosis using MCQ-ELIMINATIVE. Finally, emphasizing the importance of response verification in medical settings, we utilize a reward model mechanism, replacing the elimination process performed by MCQ-ELIMINATIVE.

Kushagra Bhushan, Yatin Nandwani, Dinesh Khandelwal et al. · ibm-research

Retrieval-Augmented Generation (RAG) has emerged as a prominent method for incorporating domain knowledge into Large Language Models (LLMs). While RAG enhances response relevance by incorporating retrieved domain knowledge in the context, retrieval errors can still lead to hallucinations and incorrect answers. To recover from retriever failures, domain knowledge is injected by fine-tuning the model to generate the correct response, even in the case of retrieval errors. However, we observe that without systematic knowledge augmentation, fine-tuned LLMs may memorize new information but still fail to extract relevant domain knowledge, leading to poor performance. In this work, we present a novel framework that significantly enhances the fine-tuning process by augmenting the training data in two ways -- context augmentation and knowledge paraphrasing. In context augmentation, we create multiple training samples for a given QA pair by varying the relevance of the retrieved information, teaching the model when to ignore and when to rely on retrieved content. In knowledge paraphrasing, we fine-tune with multiple answers to the same question, enabling LLMs to better internalize specialized knowledge. To mitigate catastrophic forgetting due to fine-tuning, we add a domain-specific identifier to a question and also utilize a replay buffer containing general QA pairs. Experimental results demonstrate the efficacy of our method over existing techniques, achieving up to 10\% relative gain in token-level recall while preserving the LLM's generalization capabilities.

Gaurav Pandey, Yatin Nandwani, Tahira Naseem et al. · ibm-research

Distribution matching methods for language model alignment such as Generation with Distributional Control (GDC) and Distributional Policy Gradient (DPG) have not received the same level of attention in reinforcement learning from human feedback (RLHF) as contrastive methods such as Sequence Likelihood Calibration (SLiC), Direct Preference Optimization (DPO) and its variants. We identify high variance of the gradient estimate as the primary reason for the lack of success of these methods and propose a self-normalized baseline to reduce the variance. We further generalize the target distribution in DPG, GDC and DPO by using Bayes' rule to define the reward-conditioned posterior. The resulting approach, referred to as BRAIn - Bayesian Reward-conditioned Amortized Inference acts as a bridge between distribution matching methods and DPO and significantly outperforms prior art in summarization and Antropic HH tasks.

Vishal Vivek Saley, Rocktim Jyoti Das, Dinesh Raghu et al. · ibm-research

End-to-end Task-Oriented Dialog (TOD) systems typically require extensive training datasets to perform well. In contrast, large language model (LLM) based TOD systems can excel even with limited data due to their ability to learn tasks through in-context exemplars. However, these models lack alignment with the style of responses in training data and often generate comprehensive responses, making it difficult for users to grasp the information quickly. In response, we propose SyncTOD that synergizes LLMs with task-specific hints to improve alignment in low-data settings. SyncTOD employs small auxiliary models to provide hints and select exemplars for in-context prompts. With ChatGPT, SyncTOD achieves superior performance compared to LLM-based baselines and SoTA models in low-data settings, while retaining competitive performance in full-data settings.

Sonam Gupta, Yatin Nandwani, Asaf Yehudai et al. · ibm-research

Fine-tuning Large Language Models (LLMs) on specific datasets is a common practice to improve performance on target tasks. However, this performance gain often leads to overfitting, where the model becomes too specialized in either the task or the characteristics of the training data, resulting in a loss of generalization. This paper introduces Selective Self-to-Supervised Fine-Tuning (S3FT), a fine-tuning approach that achieves better performance than the standard supervised fine-tuning (SFT) while improving generalization. S3FT leverages the existence of multiple valid responses to a query. By utilizing the model's correct responses, S3FT reduces model specialization during the fine-tuning stage. S3FT first identifies the correct model responses from the training set by deploying an appropriate judge. Then, it fine-tunes the model using the correct model responses and the gold response (or its paraphrase) for the remaining samples. The effectiveness of S3FT is demonstrated through experiments on mathematical reasoning, Python programming and reading comprehension tasks. The results show that standard SFT can lead to an average performance drop of up to $4.4$ on multiple benchmarks, such as MMLU and TruthfulQA. In contrast, S3FT reduces this drop by half, i.e. $2.5$, indicating better generalization capabilities than SFT while performing significantly better on the fine-tuning tasks.

Vishal Vivek Saley, Goonjan Saha, Rocktim Jyoti Das et al. · ibm-research

Medical task-oriented dialogue systems can assist doctors by collecting patient medical history, aiding in diagnosis, or guiding treatment selection, thereby reducing doctor burnout and expanding access to medical services. However, doctor-patient dialogue datasets are not readily available, primarily due to privacy regulations. Moreover, existing datasets lack comprehensive annotations involving medical slots and their different attributes, such as symptoms and their onset, progression, and severity. These comprehensive annotations are crucial for accurate diagnosis. Finally, most existing datasets are non-English, limiting their utility for the larger research community. In response, we introduce MediTOD, a new dataset of doctor-patient dialogues in English for the medical history-taking task. Collaborating with doctors, we devise a questionnaire-based labeling scheme tailored to the medical domain. Then, medical professionals create the dataset with high-quality comprehensive annotations, capturing medical slots and their attributes. We establish benchmarks in supervised and few-shot settings on MediTOD for natural language understanding, policy learning, and natural language generation subtasks, evaluating models from both TOD and biomedical domains. We make MediTOD publicly available for future research.

Dinesh Khandelwal, Gnana Prakash Punnavajhala, GPS Bhargav et al.

Multi-turn tool calling is essential for LLMs to function as autonomous agents, yet synthesizing the training data required for these capabilities remains a fundamental challenge. Existing synthetic data generation pipelines often produce unrealistic dialogues for two reasons: they chain tools that are only superficially compatible rather than aligned with meaningful user tasks, and they generate dialogues in one shot, which often introduces arguments that were neither provided by the user nor produced by prior tool calls. These issues also lead to a severe underrepresentation of multi-step tool interactions. We introduce ToolWeave, a structured framework for synthesizing realistic multi-turn tool-calling dialogues. ToolWeave support realistic multi-step workflows (or tool sequences) by constructing tools with built-in dependencies and filters the workflows based on alignment with user goals. It reduces parameter hallucination by using a fine-grained planning stage that explicitly tracks parameter provenance. As a result, ToolWeave-generated synthetic dialogues contain more multi-step tool interactions (45%) and fewer hallucinations in parameters and tool names. Consequently, LLMs fine-tuned on ToolWeave consistently outperform those fine-tuned on prior datasets across three public benchmarks. Notably, Llama-3.1-70B fine-tuned on ToolWeave achieves 39.75% on BFCL-V3 multi-turn, compared to 23.50% when fine-tuned on SOTA ToolFlow data.

Marina Danilevsky, Kristjan Greenewald, Chulaka Gunasekara et al. · ibm-research

In the developer community for large language models (LLMs), there is not yet a clean pattern analogous to a software library, to support very large scale collaboration. Even for the commonplace use case of Retrieval-Augmented Generation (RAG), it is not currently possible to write a RAG application against a well-defined set of APIs that are agreed upon by different LLM providers. Inspired by the idea of compiler intrinsics, we propose some elements of such a concept through introducing a library of LLM Intrinsics for RAG. An LLM intrinsic is defined as a capability that can be invoked through a well-defined API that is reasonably stable and independent of how the LLM intrinsic itself is implemented. The intrinsics in our library are released as LoRA adapters on HuggingFace, and through a software interface with clear structured input/output characteristics on top of vLLM as an inference platform, accompanied in both places with documentation and code. This article describes the intended usage, training details, and evaluations for each intrinsic, as well as compositions of multiple intrinsics.

Ibrahim Abdelaziz, Kinjal Basu, Mayank Agarwal et al.

Large language models (LLMs) have recently shown tremendous promise in serving as the backbone to agentic systems, as demonstrated by their performance in multi-faceted, challenging benchmarks like SWE-Bench and Agent-Bench. However, to realize the true potential of LLMs as autonomous agents, they must learn to identify, call, and interact with external tools and application program interfaces (APIs) to complete complex tasks. These tasks together are termed function calling. Endowing LLMs with function calling abilities leads to a myriad of advantages, such as access to current and domain-specific information in databases and knowledge sources, and the ability to outsource tasks that can be reliably performed by tools, e.g., a Python interpreter or calculator. While there has been significant progress in function calling with LLMs, there is still a dearth of open models that perform on par with proprietary LLMs like GPT, Claude, and Gemini. Therefore, in this work, we introduce the GRANITE-20B-FUNCTIONCALLING model under an Apache 2.0 license. The model is trained using a multi-task training approach on seven fundamental tasks encompassed in function calling, those being Nested Function Calling, Function Chaining, Parallel Functions, Function Name Detection, Parameter-Value Pair Detection, Next-Best Function, and Response Generation. We present a comprehensive evaluation on multiple out-of-domain datasets comparing GRANITE-20B-FUNCTIONCALLING to more than 15 other best proprietary and open models. GRANITE-20B-FUNCTIONCALLING provides the best performance among all open models on the Berkeley Function Calling Leaderboard and fourth overall. As a result of the diverse tasks and datasets used for training our model, we show that GRANITE-20B-FUNCTIONCALLING has better generalizability on multiple tasks in seven different evaluation datasets.

Vishal Vivek Saley, Rocktim Jyoti Das, Dinesh Raghu et al.

Task-oriented dialog (TOD) agents often ground their responses on external knowledge bases (KBs). These KBs can be dynamic and may be updated frequently. Existing approaches for learning TOD agents assume the KB snapshot contemporary to each individual dialog is available during training. However, in real-world scenarios, only the latest KB snapshot is available during training and as a result, the train dialogs may contain facts conflicting with the latest KB. These dialog-KB inconsistencies in the training data may potentially confuse the TOD agent learning algorithm. In this work, we define the novel problem of learning a TOD agent with dialog-KB inconsistencies in the training data. We propose a Dialog-KB Arbitration Framework (DKAF) which reduces the dialog-KB inconsistencies by predicting the contemporary KB snapshot for each train dialog. These predicted KB snapshots are then used for training downstream TOD agents. As there are no existing datasets with dialog-KB inconsistencies, we systematically introduce inconsistencies in two publicly available dialog datasets. We show that TOD agents trained with DKAF perform better than existing baselines on both these datasets

Yatin Nandwani, Vineet Kumar, Dinesh Raghu et al.

A major concern in using deep learning based generative models for document-grounded dialogs is the potential generation of responses that are not \textit{faithful} to the underlying document. Existing automated metrics used for evaluating the faithfulness of response with respect to the grounding document measure the degree of similarity between the generated response and the document's content. However, these automated metrics are far from being well aligned with human judgments. Therefore, to improve the measurement of faithfulness, we propose a new metric that utilizes (Conditional) Point-wise Mutual Information (PMI) between the generated response and the source document, conditioned on the dialogue. PMI quantifies the extent to which the document influences the generated response -- with a higher PMI indicating a more faithful response. We build upon this idea to create a new decoding technique that incorporates PMI into the response generation process to predict more faithful responses. Our experiments on the BEGIN benchmark demonstrate an improved correlation of our metric with human evaluation. We also show that our decoding technique is effective in generating more faithful responses when compared to standard decoding techniques on a set of publicly available document-grounded dialog datasets.

Kevin Leyton-Brown, Mausam, Yatin Nandwani et al.

Peer-reviewed conferences, the main publication venues in CS, rely critically on matching highly qualified reviewers for each paper. Because of the growing scale of these conferences, the tight timelines on which they operate, and a recent surge in explicitly dishonest behavior, there is now no alternative to performing this matching in an automated way. This paper studies a novel reviewer-paper matching approach that was recently deployed in the 35th AAAI Conference on Artificial Intelligence (AAAI 2021), and has since been adopted (wholly or partially) by other conferences including ICML 2022, AAAI 2022, and IJCAI 2022. This approach has three main elements: (1) collecting and processing input data to identify problematic matches and generate reviewer-paper scores; (2) formulating and solving an optimization problem to find good reviewer-paper matchings; and (3) a two-phase reviewing process that shifts reviewing resources away from papers likely to be rejected and towards papers closer to the decision boundary. This paper also describes an evaluation of these innovations based on an extensive post-hoc analysis on real data -- including a comparison with the matching algorithm used in AAAI's previous (2020) iteration -- and supplements this with additional numerical experimentation.

Dinesh Raghu, Atishya Jain, Mausam et al.

End-to-End task-oriented dialogue systems generate responses based on dialog history and an accompanying knowledge base (KB). Inferring those KB entities that are most relevant for an utterance is crucial for response generation. Existing state of the art scales to large KBs by softly filtering over irrelevant KB information. In this paper, we propose a novel filtering technique that consists of (1) a pairwise similarity based filter that identifies relevant information by respecting the n-ary structure in a KB record. and, (2) an auxiliary loss that helps in separating contextually unrelated KB information. We also propose a new metric -- multiset entity F1 which fixes a correctness issue in the existing entity F1 metric. Experimental results on three publicly available task-oriented dialog datasets show that our proposed approach outperforms existing state-of-the-art models.

Dinesh Raghu, Shantanu Agarwal, Sachindra Joshi et al.

We propose a novel problem within end-to-end learning of task-oriented dialogs (TOD), in which the dialog system mimics a troubleshooting agent who helps a user by diagnosing their problem (e.g., car not starting). Such dialogs are grounded in domain-specific flowcharts, which the agent is supposed to follow during the conversation. Our task exposes novel technical challenges for neural TOD, such as grounding an utterance to the flowchart without explicit annotation, referring to additional manual pages when user asks a clarification question, and ability to follow unseen flowcharts at test time. We release a dataset (FloDial) consisting of 2,738 dialogs grounded on 12 different troubleshooting flowcharts. We also design a neural model, FloNet, which uses a retrieval-augmented generation architecture to train the dialog agent. Our experiments find that FloNet can do zero-shot transfer to unseen flowcharts, and sets a strong baseline for future research.

Dinesh Raghu, Nikhil Gupta, Mausam

Task-oriented dialog (TOD) systems often need to formulate knowledge base (KB) queries corresponding to the user intent and use the query results to generate system responses. Existing approaches require dialog datasets to explicitly annotate these KB queries -- these annotations can be time consuming, and expensive. In response, we define the novel problems of predicting the KB query and training the dialog agent, without explicit KB query annotation. For query prediction, we propose a reinforcement learning (RL) baseline, which rewards the generation of those queries whose KB results cover the entities mentioned in subsequent dialog. Further analysis reveals that correlation among query attributes in KB can significantly confuse memory augmented policy optimization (MAPO), an existing state of the art RL agent. To address this, we improve the MAPO baseline with simple but important modifications suited to our task. To train the full TOD system for our setting, we propose a pipelined approach: it independently predicts when to make a KB query (query position predictor), then predicts a KB query at the predicted position (query predictor), and uses the results of predicted query in subsequent dialog (next response predictor). Overall, our work proposes first solutions to our novel problem, and our analysis highlights the research challenges in training TOD systems without query annotation.

Gaurav Pandey, Dinesh Raghu, Sachindra Joshi

Sequence to sequence models attempt to capture the correlation between all the words in the input and output sequences. While this is quite useful for machine translation where the correlation among the words is indeed quite strong, it becomes problematic for conversation modelling where the correlation is often at a much abstract level. In contrast, humans tend to focus on the essential concepts discussed in the conversation context and generate responses accordingly. In this paper, we attempt to mimic this response generating mechanism by learning the essential concepts in the context and response in an unsupervised manner. The proposed model, referred to as Mask \& Focus maps the input context to a sequence of concepts which are then used to generate the response concepts. Together, the context and the response concepts generate the final response. In order to learn context concepts from the training data automatically, we \emph{mask} words in the input and observe the effect of masking on response generation. We train our model to learn those response concepts that have high mutual information with respect to the context concepts, thereby guiding the model to \emph{focus} on the context concepts. Mask \& Focus achieves significant improvement over the existing baselines in several established metrics for dialogues.

Dhiraj Madan, Dinesh Raghu, Gaurav Pandey et al.

Recently several deep learning based models have been proposed for end-to-end learning of dialogs. While these models can be trained from data without the need for any additional annotations, it is hard to interpret them. On the other hand, there exist traditional state based dialog systems, where the states of the dialog are discrete and hence easy to interpret. However these states need to be handcrafted and annotated in the data. To achieve the best of both worlds, we propose Latent State Tracking Network (LSTN) using which we learn an interpretable model in unsupervised manner. The model defines a discrete latent variable at each turn of the conversation which can take a finite set of values. Since these discrete variables are not present in the training data, we use EM algorithm to train our model in unsupervised manner. In the experiments, we show that LSTN can help achieve interpretability in dialog models without much decrease in performance compared to end-to-end approaches.

Revanth Reddy, Danish Contractor, Dinesh Raghu et al.

Recent end-to-end task oriented dialog systems use memory architectures to incorporate external knowledge in their dialogs. Current work makes simplifying assumptions about the structure of the knowledge base, such as the use of triples to represent knowledge, and combines dialog utterances (context) as well as knowledge base (KB) results as part of the same memory. This causes an explosion in the memory size, and makes the reasoning over memory harder. In addition, such a memory design forces hierarchical properties of the data to be fit into a triple structure of memory. This requires the memory reader to infer relationships across otherwise connected attributes. In this paper we relax the strong assumptions made by existing architectures and separate memories used for modeling dialog context and KB results. Instead of using triples to store KB results, we introduce a novel multi-level memory architecture consisting of cells for each query and their corresponding results. The multi-level memory first addresses queries, followed by results and finally each key-value pair within a result. We conduct detailed experiments on three publicly available task oriented dialog data sets and we find that our method conclusively outperforms current state-of-the-art models. We report a 15-25% increase in both entity F1 and BLEU scores.

Dinesh Raghu, Nikhil Gupta, Mausam

The Knowledge Base (KB) used for real-world applications, such as booking a movie or restaurant reservation, keeps changing over time. End-to-end neural networks trained for these task-oriented dialogs are expected to be immune to any changes in the KB. However, existing approaches breakdown when asked to handle such changes. We propose an encoder-decoder architecture (BoSsNet) with a novel Bag-of-Sequences (BoSs) memory, which facilitates the disentangled learning of the response's language model and its knowledge incorporation. Consequently, the KB can be modified with new knowledge without a drop in interpretability. We find that BoSsNet outperforms state-of-the-art models, with considerable improvements (> 10\%) on bAbI OOV test sets and other human-human datasets. We also systematically modify existing datasets to measure disentanglement and show BoSsNet to be robust to KB modifications.