Rashmi Gangadharaiah

Vinayshekhar Bannihatti Kumar, Manoj Ghuhan Arivazhagan, Disha Makhija et al.

Prior KV cache compression schemes empirically demonstrate that the prompt cache is partially redundant during decoding, dropping or summarising entries with little accuracy loss. We ask when and what kind of redundancy: at which layers, after how many decoding steps, and in what form can the prompt span KV cache be replaced without breaking the task. A controlled splice intervention swept over layer cutoff and decoding steps shows this redundancy is about form (chat template scaffolding) rather than content. Replacing the upper layer prompt span KV cache with KV cache from a chat template scaffold whose user content is a neutral filler recovers near clean accuracy, while zeroing the same slots collapses accuracy. The dissociation replicates across the Qwen3, Gemma 3, and Llama 3 families on multiple datasets.

Amazon AGI, Aaron Langford, Aayush Shah et al. · amazon-science

We present Amazon Nova, a new generation of state-of-the-art foundation models that deliver frontier intelligence and industry-leading price performance. Amazon Nova Pro is a highly-capable multimodal model with the best combination of accuracy, speed, and cost for a wide range of tasks. Amazon Nova Lite is a low-cost multimodal model that is lightning fast for processing images, video, documents and text. Amazon Nova Micro is a text-only model that delivers our lowest-latency responses at very low cost. Amazon Nova Canvas is an image generation model that creates professional grade images with rich customization controls. Amazon Nova Reel is a video generation model offering high-quality outputs, customization, and motion control. Our models were built responsibly and with a commitment to customer trust, security, and reliability. We report benchmarking results for core capabilities, agentic performance, long context, functional adaptation, runtime performance, and human evaluation.

Vinayshekhar Bannihatti Kumar, Rashmi Gangadharaiah, Dan Roth

Regulations introduced by General Data Protection Regulation (GDPR) in the EU or California Consumer Privacy Act (CCPA) in the US have included provisions on the \textit{right to be forgotten} that mandates industry applications to remove data related to an individual from their systems. In several real world industry applications that use Machine Learning to build models on user data, such mandates require significant effort both in terms of data cleansing as well as model retraining while ensuring the models do not deteriorate in prediction quality due to removal of data. As a result, continuous removal of data and model retraining steps do not scale if these applications receive such requests at a very high frequency. Recently, a few researchers proposed the idea of \textit{Machine Unlearning} to tackle this challenge. Despite the significant importance of this task, the area of Machine Unlearning is under-explored in Natural Language Processing (NLP) tasks. In this paper, we explore the Unlearning framework on various GLUE tasks \cite{Wang:18}, such as, QQP, SST and MNLI. We propose computationally efficient approaches (SISA-FC and SISA-A) to perform \textit{guaranteed} Unlearning that provides significant reduction in terms of both memory (90-95\%), time (100x) and space consumption (99\%) in comparison to the baselines while keeping model performance constant.

Vinayshekhar Bannihatti Kumar, Rashmi Gangadharaiah, Dan Roth

Research has shown that personality is a key driver to improve engagement and user experience in conversational systems. Conversational agents should also maintain a consistent persona to have an engaging conversation with a user. However, text generation datasets are often crowd sourced and thereby have an averaging effect where the style of the generation model is an average style of all the crowd workers that have contributed to the dataset. While one can collect persona-specific datasets for each task, it would be an expensive and time consuming annotation effort. In this work, we propose a novel transfer learning framework which updates only $0.3\%$ of model parameters to learn style specific attributes for response generation. For the purpose of this study, we tackle the problem of stylistic story ending generation using the ROC stories Corpus. We learn style specific attributes from the PERSONALITY-CAPTIONS dataset. Through extensive experiments and evaluation metrics we show that our novel training procedure can improve the style generation by 200 over Encoder-Decoder baselines while maintaining on-par content relevance metrics with

Dhruv Agarwal, Rajarshi Das, Sopan Khosla et al.

We present BYOKG, a universal question-answering (QA) system that can operate on any knowledge graph (KG), requires no human-annotated training data, and can be ready to use within a day -- attributes that are out-of-scope for current KGQA systems. BYOKG draws inspiration from the remarkable ability of humans to comprehend information present in an unseen KG through exploration -- starting at random nodes, inspecting the labels of adjacent nodes and edges, and combining them with their prior world knowledge. In BYOKG, exploration leverages an LLM-backed symbolic agent that generates a diverse set of query-program exemplars, which are then used to ground a retrieval-augmented reasoning procedure to predict programs for arbitrary questions. BYOKG is effective over both small- and large-scale graphs, showing dramatic gains in QA accuracy over a zero-shot baseline of 27.89 and 58.02 F1 on GrailQA and MetaQA, respectively. On GrailQA, we further show that our unsupervised BYOKG outperforms a supervised in-context learning method, demonstrating the effectiveness of exploration. Lastly, we find that performance of BYOKG reliably improves with continued exploration as well as improvements in the base LLM, notably outperforming a state-of-the-art fine-tuned model by 7.08 F1 on a sub-sampled zero-shot split of GrailQA.

Rashmi Gangadharaiah, Balakrishnan Narayanaswamy

It is expensive and difficult to obtain the large number of sentence-level intent and token-level slot label annotations required to train neural network (NN)-based Natural Language Understanding (NLU) components of task-oriented dialog systems, especially for the many real world tasks that have a large and growing number of intents and slot types. While zero shot learning approaches that require no labeled examples -- only features and auxiliary information -- have been proposed only for slot labeling, we show that one can profitably perform joint zero-shot intent classification and slot labeling. We demonstrate the value of capturing dependencies between intents and slots, and between different slots in an utterance in the zero shot setting. We describe NN architectures that translate between word and sentence embedding spaces, and demonstrate that these modifications are required to enable zero shot learning for this task. We show a substantial improvement over strong baselines and explain the intuition behind each architectural modification through visualizations and ablation studies.

Sandesh Swamy, Narges Tabari, Chacha Chen et al.

Response generation is one of the critical components in task-oriented dialog systems. Existing studies have shown that large pre-trained language models can be adapted to this task. The typical paradigm of adapting such extremely large language models would be by fine-tuning on the downstream tasks which is not only time-consuming but also involves significant resources and access to fine-tuning data. Prompting (Schick and Schütze, 2020) has been an alternative to fine-tuning in many NLP tasks. In our work, we explore the idea of using prompting for response generation in task-oriented dialog systems. Specifically, we propose an approach that performs contextual dynamic prompting where the prompts are learnt from dialog contexts. We aim to distill useful prompting signals from the dialog context. On experiments with MultiWOZ 2.2 dataset (Zang et al., 2020), we show that contextual dynamic prompts improve response generation in terms of combined score (Mehri et al., 2019) by 3 absolute points, and a massive 20 points when dialog states are incorporated. Furthermore, human annotation on these conversations found that agents which incorporate context were preferred over agents with vanilla prefix-tuning.

Vinayshekhar Bannihatti Kumar, Disha Makhija, Manoj Ghuhan Arivazhagan et al.

Large language models (LLMs) achieve high pass rates on code generation benchmarks, yet whether they can transfer this ability to languages absent from pretraining remains poorly understood. We introduce PyLang, a minimal imperative language absent from all pretraining corpora, and evaluate frontier models zero-shot and fine-tuned Qwen3 (4B, 8B, 32B) on 352 problems. We find that fine-tuning quickly teaches syntax but fails to transfer semantic competence: Python outperforms PyLang by up to 19% across all configurations, and no intervention (multi-task learning, preference tuning, code infilling, or latent-space objectives) closes the gap. An LLM judge reveals that frontier models select an identical algorithm to Python 80% of the time, yet cannot translate it into a working PyLang implementation., and CKA analysis confirms that fine-tuned models converge to nearly identical internal representations across languages (CKA > 0.97) while diverging at the output stage. We term this the implementation fidelity gap: models possess language-agnostic algorithmic understanding but cannot express it in an unfamiliar language. Our findings highlight the need for training methods that decouple reasoning from language-specific realization.

Narges Tabari, Sandesh Swamy, Rashmi Gangadharaiah

Providing a personalized user experience on information dense webpages helps users in reaching their end-goals sooner. We explore an automated approach to identifying user personas by leveraging high dimensional trajectory information from user sessions on webpages. While neural collaborative filtering (NCF) approaches pay little attention to token semantics, our method introduces SessionBERT, a Transformer-backed language model trained from scratch on the masked language modeling (mlm) objective for user trajectories (pages, metadata, billing in a session) aiming to capture semantics within them. Our results show that representations learned through SessionBERT are able to consistently outperform a BERT-base model providing a 3% and 1% relative improvement in F1-score for predicting page links and next services. We leverage SessionBERT and extend it to provide recommendations (top-5) for the next most-relevant services that a user would be likely to use. We achieve a HIT@5 of 58% from our recommendation model.

Kasturi Bhattacharjee, Rashmi Gangadharaiah

Aspect-based sentiment analysis (ABSA) is a widely studied topic, most often trained through supervision from human annotations of opinionated texts. These fine-grained annotations include identifying aspects towards which a user expresses their sentiment, and their associated polarities (aspect-based sentiments). Such fine-grained annotations can be expensive and often infeasible to obtain in real-world settings. There is, however, an abundance of scenarios where user-generated text contains an overall sentiment, such as a rating of 1-5 in user reviews or user-generated feedback, which may be leveraged for this task. In this paper, we propose a VAE-based topic modeling approach that performs ABSA using document-level supervision and without requiring fine-grained labels for either aspects or sentiments. Our approach allows for the detection of multiple aspects in a document, thereby allowing for the possibility of reasoning about how sentiment expressed through multiple aspects comes together to form an observable overall document-level sentiment. We demonstrate results on two benchmark datasets from two different domains, significantly outperforming a state-of-the-art baseline.

Hanhan Zhou, Shamik Roy, Rashmi Gangadharaiah

Discrete diffusion language models (DLMs) generate text by iteratively denoising all positions in parallel, offering an alternative to autoregressive models. Controlled generation methods for DLMs, imported from autoregressive models, apply uniform intervention at every denoising steps. We show this uniform schedule degrades quality, and the damage compounds when multiple attributes are steered jointly. To diagnose the failure, we train sparse autoencoders on four DLMs (124M-8B parameters) and find that different attributes commit on distinct schedules, varying in timing, sharpness, and magnitude. For instance, topic commits within the first 2\% of denoising, whereas sentiment emerges gradually over 20\% of the process. Consequently, uniform intervention wastes steering capacity on steps where the target attribute has already solidified or has yet to emerge. We propose a novel adaptive scheduler that concentrates interventions on the steps where an attribute is actively forming and leaves the rest of generation untouched. The cost-control trade-off admits a closed-form characterization: the advantage of adaptive over uniform scheduling is governed by a single dispersion statistic of the commitment distribution. Across four DLMs and seven steering tasks, our method achieves precise control without the degradation typical of uniform interventions. Especially on challenging simultaneous three-attribute control, it reaches up to 93\% steering strength, beating the strongest baseline by up to 15\% points while preserving generation quality.

Nishanth Nakshatri, Shamik Roy, Rajarshi Das et al.

Constrained decoding with lookahead heuristics (CDLH) is a highly effective method for aligning LLM generations to human preferences. However, the extensive lookahead roll-out operations for each generated token makes CDLH prohibitively expensive, resulting in low adoption in practice. In contrast, common decoding strategies such as greedy decoding are extremely efficient, but achieve very low constraint satisfaction. We propose constrained decoding with speculative lookaheads (CDSL), a technique that significantly improves upon the inference efficiency of CDLH without experiencing the drastic performance reduction seen with greedy decoding. CDSL is motivated by the recently proposed idea of speculative decoding that uses a much smaller draft LLM for generation and a larger target LLM for verification. In CDSL, the draft model is used to generate lookaheads which is verified by a combination of target LLM and task-specific reward functions. This process accelerates decoding by reducing the computational burden while maintaining strong performance. We evaluate CDSL in two constraint decoding tasks with three LLM families and achieve 2.2x to 12.15x speedup over CDLH without significant performance reduction.

Mintong Kang, Vinayshekhar Bannihatti Kumar, Shamik Roy et al.

Text-to-image diffusion models often exhibit biases toward specific demographic groups, such as generating more males than females when prompted to generate images of engineers, raising ethical concerns and limiting their adoption. In this paper, we tackle the challenge of mitigating generation bias towards any target attribute value (e.g., "male" for "gender") in diffusion models while preserving generation quality. We propose FairGen, an adaptive latent guidance mechanism which controls the generation distribution during inference. In FairGen, a latent guidance module dynamically adjusts the diffusion process to enforce specific attributes, while a memory module tracks the generation statistics and steers latent guidance to align with the targeted fair distribution of the attribute values. Furthermore, we address the limitations of existing datasets by introducing the Holistic Bias Evaluation (HBE) benchmark, which covers diverse domains and incorporates complex prompts to assess bias more comprehensively. Extensive evaluations on HBE and Stable Bias datasets demonstrate that FairGen outperforms existing bias mitigation approaches, achieving substantial bias reduction (e.g., 68.5% gender bias reduction on Stable Diffusion 2). Ablation studies highlight FairGen's ability to flexibly control the output distribution at any user-specified granularity, ensuring adaptive and targeted bias mitigation.

Zhaoyi Joey Hou, Tanya Shourya, Yingfan Wang et al.

Evaluating conversational AI systems that use external tools is challenging, as errors can arise from complex interactions among user, agent, and tools. While existing evaluation methods assess either user satisfaction or agents' tool-calling capabilities, they fail to capture critical errors in multi-turn tool-augmented dialogues-such as when agents misinterpret tool results yet appear satisfactory to users. We introduce TRACE, a benchmark of systematically synthesized tool-augmented conversations covering diverse error cases, and SCOPE, an evaluation framework that automatically discovers diverse error patterns and evaluation rubrics in tool-augmented dialogues. Experiments show SCOPE significantly outperforms the baseline, particularly on challenging cases where user satisfaction signals are misleading.

Nishanth Sridhar Nakshatri, Shamik Roy, Manoj Ghuhan Arivazhagan et al.

LLMs often fail to handle temporal knowledge conflicts--contradictions arising when facts evolve over time within their training data. Existing studies evaluate this phenomenon through benchmarks built on structured knowledge bases like Wikidata, but they focus on widely-covered, easily-memorized popular entities and lack the dynamic structure needed to fairly evaluate LLMs with different knowledge cut-off dates. We introduce evolveQA, a benchmark specifically designed to evaluate LLMs on temporally evolving knowledge, constructed from 3 real-world, time-stamped corpora: AWS updates, Azure changes, and WHO disease outbreak reports. Our framework identifies naturally occurring knowledge evolution and generates questions with gold answers tailored to different LLM knowledge cut-off dates. Through extensive evaluation of 12 open and closed-source LLMs across 3 knowledge probing formats, we demonstrate significant performance drops of up to 31% on evolveQA compared to static knowledge questions.

Disha Makhija, Manoj Ghuhan Arivazhagan, Vinayshekhar Bannihatti Kumar et al.

Membership inference attacks (MIAs) reveal whether specific data was used to train machine learning models, serving as important tools for privacy auditing and compliance assessment. Recent studies have reported that MIAs perform only marginally better than random guessing against large language models, suggesting that modern pre-training approaches with massive datasets may be free from privacy leakage risks. Our work offers a complementary perspective to these findings by exploring how examining LLMs' internal representations, rather than just their outputs, may provide additional insights into potential membership inference signals. Our framework, \emph{memTrace}, follows what we call \enquote{neural breadcrumbs} extracting informative signals from transformer hidden states and attention patterns as they process candidate sequences. By analyzing layer-wise representation dynamics, attention distribution characteristics, and cross-layer transition patterns, we detect potential memorization fingerprints that traditional loss-based approaches may not capture. This approach yields strong membership detection across several model families achieving average AUC scores of 0.85 on popular MIA benchmarks. Our findings suggest that internal model behaviors can reveal aspects of training data exposure even when output-based signals appear protected, highlighting the need for further research into membership privacy and the development of more robust privacy-preserving training techniques for large language models.

Rashmi Gangadharaiah, Balakrishnan Narayanaswamy, Charles Elkan

We consider real world task-oriented dialog settings, where agents need to generate both fluent natural language responses and correct external actions like database queries and updates. We demonstrate that, when applied to customer support chat transcripts, Sequence to Sequence (Seq2Seq) models often generate short, incoherent and ungrammatical natural language responses that are dominated by words that occur with high frequency in the training data. These phenomena do not arise in synthetic datasets such as bAbI, where we show Seq2Seq models are nearly perfect. We develop techniques to learn embeddings that succinctly capture relevant information from the dialog history, and demonstrate that nearest neighbor based approaches in this learned neural embedding space generate more fluent responses. However, we see that these methods are not able to accurately predict when to execute an external action. We show how to combine nearest neighbor and Seq2Seq methods in a hybrid model, where nearest neighbor is used to generate fluent responses and Seq2Seq type models ensure dialog coherency and generate accurate external actions. We show that this approach is well suited for customer support scenarios, where agents' responses are typically script-driven, and correct external actions are critically important. The hybrid model on the customer support data achieves a 78% relative improvement in fluency scores, and a 130% improvement in accuracy of external calls.

Niraj Kumar, Rashmi Gangadharaiah, Kannan Srinathan et al.

In this paper, we show that certain phrases although not present in a given question/query, play a very important role in answering the question. Exploring the role of such phrases in answering questions not only reduces the dependency on matching question phrases for extracting answers, but also improves the quality of the extracted answers. Here matching question phrases means phrases which co-occur in given question and candidate answers. To achieve the above discussed goal, we introduce a bigram-based word graph model populated with semantic and topical relatedness of terms in the given document. Next, we apply an improved version of ranking with a prior-based approach, which ranks all words in the candidate document with respect to a set of root words (i.e. non-stopwords present in the question and in the candidate document). As a result, terms logically related to the root words are scored higher than terms that are not related to the root words. Experimental results show that our devised system performs better than state-of-the-art for the task of answering Why-questions.