Searcharxiv⌕ Search

arXiv subjects

Soumyadeep Roy

Publications and source records attributed to Soumyadeep Roy.

15 recordsLinked to original sources

Beyond Binary Audits: Does Guideline Grounding Mitigate Decision-Space Framing in LLM Opioid Prescribing?

Audits of LLMs for bias in opioid prescribing have largely relied on binary treatment decisions, a signal that current models have saturated. We introduce Q-Pain++, a full-factorial stress test that holds clinical content fixed while independently varying clinical risk factors and the granularity of the available dosage options. Adding an intermediate dosage option to otherwise unchanged vignettes substantially lowers agreement with expert-annotated dosages for most models, while demographic differences in dosage decisions remain small in all but one configuration once clinical risk factors are added. Grounding models in perioperative guidelines through retrieval reduces this framing effect on average but does not eliminate it, and one model's treatment decisions change when the retrieved passages are reordered. These results suggest that audits should vary the answer format, not only patient identity. We will release our codebase as well as Guideline-RAG-associated resources on GitHub (https://github.com/su-boussard-lab/llm-guideline-grounding-opioids).

cs.CL↗

Zero-Shot Temporal Localisation of Audio Deepfakes in Multi-Speaker Conversations

Voice-cloning fraud increasingly relies on surgical injection: a genuine conversation in which only one or two sentences are replaced by synthetic speech. Utterance-level deepfake detectors emit a single real/fake label per clip and cannot report where the synthetic speech lies. We formalise this as Temporal Deepfake Localisation in Multi-Speaker Conversations (TDLMC), show that equal error rate and min-DCF are ill-posed once a file contains both classes, and propose temporal metrics for this regime. Our contribution is a training-free five-stage pipeline that wraps a frozen binary detector and adds segment-level output with no retraining, using a two-threshold hysteresis finitestate-machine decoder to turn noisy window scores into coherent intervals. On 180 constructed multi-speaker conversations from ASVspoof 5, the system attains temporal intersection-over-union 0.90, temporal detection rate 0.95, and MS-DCF 0.26 with a strong backbone, and its false-alarm rate on genuine speech is below 6%, falling under 2% on genuine real multi-speaker dialogue (AMI). Under an identical pipeline, a trained localiser improves temporal IoU by only about 0.04, bounding the cost of forgoing supervision. Evaluated across three frozen detectors under one decoder whose constants are selected on a held-out calibration split, and with a controlled analysis attributing the residual false-alarm rate to a backbone domain gap rather than to the decoder, this provides the first zero-shot baseline and a reusable benchmark for TDLMC.

cs.SD↗

The Hidden Costs of 99% Accuracy: A Trustworthiness Audit of the Telco Customer Churn Benchmark

Customer churn prediction on the IBM Telco Customer Churn benchmark (n = 7,043) routinely reports test accuracies above 95%, with the most cited published study reporting 99.01%. We audit this benchmark for four trustworthiness failures invisible to the accuracy- and F1-centred reporting that dominates the literature. First, pre-split SMOTE inflates churn-class F1 by 13.1 percentage points across ten classifiers and fifteen seeds (Wilcoxon p < 10^-4 per classifier); the same leaky pipeline ordering paired with class weighting yields no inflation, isolating the effect to SMOTE's geometric construction. We measure the mechanism directly: approximately 36% of synthetic training points are nearest-neighbour interpolations of test-set instances. Second, the TotalCharges field is approximately determined by tenure multiplied by MonthlyCharges (R2 = 0.999); removing it changes accuracy by less than 0.2 percentage points, yet TreeSHAP ranks it ninth in mean absolute attribution - a pattern that materially corrupts SHAP-based interpretation. We propose an R2 > 0.95 pre-modelling diagnostic. Third, in a 15-seed calibration audit, isotonic regression is the strongest default; temperature scaling fails on class-weighted tree ensembles whose predicted-probability distribution is bimodal. Fourth, the cost-optimal decision threshold (under a 50 USD retention offer and 24-month CLV proxy) is approximately 5-10 times lower than the F1-optimal threshold, saving approximately 77,000 USD per 1,000 customers. We replicate F1 and F2 on Iranian Telecom Churn (within domain) and Bank Customer Churn (across domain): F1 generalises; F2 generalises only within telecom. We synthesise these findings into a four-component reporting checklist - pipeline disclosure, redundancy diagnostic, calibration audit, and cost-sensitive thresholds - and release a reproducible implementation.

cs.LG↗

A Fairness Audit of the Duckworth-Lewis-Stern Method: Format-Specific and Gender-Differential Bias, with an Interpretable Calibration Layer for Cricket Target Revision

The Duckworth-Lewis-Stern (DLS) method has been the international standard for revising target scores in rain-interrupted limited-overs cricket since 1999. Despite over two decades of operational use, no large-scale empirical audit of its prediction bias has been published. We conduct such an audit on 8,150 international matches (3,095 ODIs, 5,055 T20Is) from Cricsheet, generating 233,550 synthetic interruption scenarios with temporal splits. We document two structured biases. First, DLS prediction error spans a 137-run range across (overs-remaining, wickets-lost) match-state buckets. Second, DLS exhibits a gender-differential bias on ODIs that has not previously been quantified: on the training split, mean over-prediction is +1.51 runs for men but +7.63 runs for women, a gap of +6.13 runs (F = 195.16, p < 10^-43). We benchmark DLS against five modern alternatives: Bi-LSTM, XGBoost, an enriched XGBoost variant, a deep context-aware model, and a stacking ensemble, and propose DLS-Cal, a lightweight interpretable calibration layer (27K parameters) outputting a state-conditioned correction added to DLS. DLS-Cal reduces absolute bias by 31% on ODI and 19% on T20I, and a gender-aware variant reduces women's ODI residual bias from +6.19 to +0.65 runs while leaving men's calibration unchanged. We release code, models, and data.

cs.LG↗

Learning Faster with Better Tokens: Parameter-Efficient Vocabulary Adaptation for Specialized Text Summarization

Large language models pretrained on general-domain corpora often exhibit tokenization inefficiencies when applied to specialized domains. Although continual pretraining for domain adaptation partially alleviate performance degradation, it does not resolve the fundamental vocabulary mismatch. To address this gap, we introduce a targeted parameter-efficient domain adaptation approach that combines vocabulary adaptation with pretraining for LLM-based text summarization. Our unified framework augments pretrained tokenizers with domain-specific tokens while selectively replacing under-trained and unreachable tokens to limit parameter growth. We evaluate our approach on Llama-3.1-8B and Qwen2.5-7B across legal and medical summarization tasks on a challenge-oriented evaluation protocol focused on expert-driven text and summaries which typically has higher concentration of over-fragmented Out-of-Vocabulary (OOV) words. The vocabulary adaptation algorithm enhances the overall quality of the summarization model by improving semantic similarity between the generated summaries and their references. In addition, the adapted model produces summaries that incorporate more appropriate novel and domain-specific words, leading to improved coherence, relevance, and faithfulness. We further observe that our proposed approach significantly reduce training time by $35-55\%$ over continual pretraining and reduce parameter counts up to $37\%$ w.r.t expansion-only methods. We make the codebase publicly available at https://github.com/gb-kgp/VocabReplace-Then-Expand.

cs.CL↗

Sample Efficient Active Algorithms for Offline Reinforcement Learning

Offline reinforcement learning (RL) enables policy learning from static data but often suffers from poor coverage of the state-action space and distributional shift problems. This problem can be addressed by allowing limited online interactions to selectively refine uncertain regions of the learned value function, which is referred to as Active Reinforcement Learning (ActiveRL). While there has been good empirical success, no theoretical analysis is available in the literature. We fill this gap by developing a rigorous sample-complexity analysis of ActiveRL through the lens of Gaussian Process (GP) uncertainty modeling. In this respect, we propose an algorithm and using GP concentration inequalities and information-gain bounds, we derive high-probability guarantees showing that an $ε$-optimal policy can be learned with ${\mathcal{O}}(1/ε^2)$ active transitions, improving upon the $Ω(1/ε^2(1-γ)^4)$ rate of purely offline methods. Our results reveal that ActiveRL achieves near-optimal information efficiency, that is, guided uncertainty reduction leads to accelerated value-function convergence with minimal online data. Our analysis builds on GP concentration inequalities and information-gain bounds, bridging Bayesian nonparametric regression and reinforcement learning theories. We conduct several experiments to validate the algorithm and theoretical findings.

cs.LG↗

"Where does it hurt?" -- Dataset and Study on Physician Intent Trajectories in Doctor Patient Dialogues

In a doctor-patient dialogue, the primary objective of physicians is to diagnose patients and propose a treatment plan. Medical doctors guide these conversations through targeted questioning to efficiently gather the information required to provide the best possible outcomes for patients. To the best of our knowledge, this is the first work that studies physician intent trajectories in doctor-patient dialogues. We use the `Ambient Clinical Intelligence Benchmark' (Aci-bench) dataset for our study. We collaborate with medical professionals to develop a fine-grained taxonomy of physician intents based on the SOAP framework (Subjective, Objective, Assessment, and Plan). We then conduct a large-scale annotation effort to label over 5000 doctor-patient turns with the help of a large number of medical experts recruited using Prolific, a popular crowd-sourcing platform. This large labeled dataset is an important resource contribution that we use for benchmarking the state-of-the-art generative and encoder models for medical intent classification tasks. Our findings show that our models understand the general structure of medical dialogues with high accuracy, but often fail to identify transitions between SOAP categories. We also report for the first time common trajectories in medical dialogue structures that provide valuable insights for designing `differential diagnosis' systems. Finally, we extensively study the impact of intent filtering for medical dialogue summarization and observe a significant boost in performance. We make the codes and data, including annotation guidelines, publicly available at https://github.com/DATEXIS/medical-intent-classification.

cs.CL↗

Evaluation of LLMs in Medical Text Summarization: The Role of Vocabulary Adaptation in High OOV Settings

Large Language Models (LLMs) recently achieved great success in medical text summarization by simply using in-context learning. However, these recent efforts do not perform fine-grained evaluations under difficult settings where LLMs might fail. They typically report performance scores over the entire dataset. Through our benchmarking study, we show that LLMs show a significant performance drop for data points with high concentration of out-of-vocabulary (OOV) words or with high novelty. Vocabulary adaptation is an intuitive solution to this vocabulary mismatch issue where the LLM vocabulary gets updated with certain expert domain (here, medical) words or subwords. An interesting finding from our study is that Llama-3.1, even with a vocabulary size of around 128K tokens, still faces over-fragmentation issue with medical words. To that end, we show vocabulary adaptation helps improve the LLM summarization performance even in difficult settings. Through extensive experimentation of multiple vocabulary adaptation strategies, two continual pretraining strategies, and three benchmark medical summarization datasets, we gain valuable insights into the role of vocabulary adaptation strategies for customizing LLMs to the medical domain. We also performed a human evaluation study with medical experts where they found that vocabulary adaptation results in more relevant and faithful summaries. Our codebase is made publicly available at https://github.com/gb-kgp/LLM-MedicalSummarization-Benchmark.

cs.CL↗

Adaptive BPE Tokenization for Enhanced Vocabulary Adaptation in Finetuning Pretrained Language Models

In this work, we show a fundamental limitation in vocabulary adaptation approaches that use Byte-Pair Encoding (BPE) tokenization scheme for fine-tuning pretrained language models (PLMs) to expert domains. Current approaches trivially append the target domain-specific vocabulary at the end of the PLM vocabulary. This approach leads to a lower priority score and causes sub-optimal tokenization in BPE that iteratively uses merge rules to tokenize a given text. To mitigate this issue, we propose AdaptBPE where the BPE tokenization initialization phase is modified to first perform the longest string matching on the added (target) vocabulary before tokenizing at the character level. We perform an extensive evaluation of AdaptBPE versus the standard BPE over various classification and summarization tasks; AdaptBPE improves by 3.57% (in terms of accuracy) and 1.87% (in terms of Rouge-L), respectively. AdaptBPE for MEDVOC works particularly well when reference summaries have high OOV concentration or are longer in length. We also conduct a human evaluation, revealing that AdaptBPE generates more relevant and more faithful summaries as compared to MEDVOC. We make our codebase publicly available at https://github.com/gb-kgp/adaptbpe.

cs.CL↗

MEDVOC: Vocabulary Adaptation for Fine-tuning Pre-trained Language Models on Medical Text Summarization

This work presents a dynamic vocabulary adaptation strategy, MEDVOC, for fine-tuning pre-trained language models (PLMs) like BertSumAbs, BART, and PEGASUS for improved medical text summarization. In contrast to existing domain adaptation approaches in summarization, MEDVOC treats vocabulary as an optimizable parameter and optimizes the PLM vocabulary based on fragment score conditioned only on the downstream task's reference summaries. Unlike previous works on vocabulary adaptation (limited only to classification tasks), optimizing vocabulary based on summarization tasks requires an extremely costly intermediate fine-tuning step on large summarization datasets. To that end, our novel fragment score-based hyperparameter search very significantly reduces this fine-tuning time -- from 450 days to less than 2 days on average. Furthermore, while previous works on vocabulary adaptation are often primarily tied to single PLMs, MEDVOC is designed to be deployable across multiple PLMs (with varying model vocabulary sizes, pre-training objectives, and model sizes) -- bridging the limited vocabulary overlap between the biomedical literature domain and PLMs. MEDVOC outperforms baselines by 15.74% in terms of Rouge-L in zero-shot setting and shows gains of 17.29% in high Out-Of-Vocabulary (OOV) concentrations. Our human evaluation shows MEDVOC generates more faithful medical summaries (88% compared to 59% in baselines). We make the codebase publicly available at https://github.com/gb-kgp/MEDVOC.

cs.CL↗

Unlocking Efficiency: Adaptive Masking for Gene Transformer Models

Gene transformer models such as Nucleotide Transformer, DNABert, and LOGO are trained to learn optimal gene sequence representations by using the Masked Language Modeling (MLM) training objective over the complete Human Reference Genome. However, the typical tokenization methods employ a basic sliding window of tokens, such as k-mers, that fail to utilize gene-centric semantics. This could result in the (trivial) masking of easily predictable sequences, leading to inefficient MLM training. Time-variant training strategies are known to improve pretraining efficiency in both language and vision tasks. In this work, we focus on using curriculum masking where we systematically increase the difficulty of masked token prediction task by using a Pointwise Mutual Information-based difficulty criterion, as gene sequences lack well-defined semantic units similar to words or sentences of NLP domain. Our proposed Curriculum Masking-based Gene Masking Strategy (CM-GEMS) demonstrates superior representation learning capabilities compared to baseline masking approaches when evaluated on downstream gene sequence classification tasks. We perform extensive evaluation in both few-shot (five datasets) and full dataset settings (Genomic Understanding Evaluation benchmark consisting of 27 tasks). Our findings reveal that CM-GEMS outperforms state-of-the-art models (DNABert-2, Nucleotide transformer, DNABert) trained at 120K steps, achieving similar results in just 10K and 1K steps. We also demonstrate that Curriculum-Learned LOGO (a 2-layer DNABert-like model) can achieve nearly 90% of the state-of-the-art model performance of 120K steps. We will make the models and codes publicly available at https://github.com/roysoumya/curriculum-GeneMask.

cs.CL↗

Beyond Accuracy: Investigating Error Types in GPT-4 Responses to USMLE Questions

GPT-4 demonstrates high accuracy in medical QA tasks, leading with an accuracy of 86.70%, followed by Med-PaLM 2 at 86.50%. However, around 14% of errors remain. Additionally, current works use GPT-4 to only predict the correct option without providing any explanation and thus do not provide any insight into the thinking process and reasoning used by GPT-4 or other LLMs. Therefore, we introduce a new domain-specific error taxonomy derived from collaboration with medical students. Our GPT-4 USMLE Error (G4UE) dataset comprises 4153 GPT-4 correct responses and 919 incorrect responses to the United States Medical Licensing Examination (USMLE) respectively. These responses are quite long (258 words on average), containing detailed explanations from GPT-4 justifying the selected option. We then launch a large-scale annotation study using the Potato annotation platform and recruit 44 medical experts through Prolific, a well-known crowdsourcing platform. We annotated 300 out of these 919 incorrect data points at a granular level for different classes and created a multi-label span to identify the reasons behind the error. In our annotated dataset, a substantial portion of GPT-4's incorrect responses is categorized as a "Reasonable response by GPT-4," by annotators. This sheds light on the challenge of discerning explanations that may lead to incorrect options, even among trained medical professionals. We also provide medical concepts and medical semantic predications extracted using the SemRep tool for every data point. We believe that it will aid in evaluating the ability of LLMs to answer complex medical questions. We make the resources available at https://github.com/roysoumya/usmle-gpt4-error-taxonomy .

cs.CL↗

GeneMask: Fast Pretraining of Gene Sequences to Enable Few-Shot Learning

Large-scale language models such as DNABert and LOGO aim to learn optimal gene representations and are trained on the entire Human Reference Genome. However, standard tokenization schemes involve a simple sliding window of tokens like k-mers that do not leverage any gene-based semantics and thus may lead to (trivial) masking of easily predictable sequences and subsequently inefficient Masked Language Modeling (MLM) training. Therefore, we propose a novel masking algorithm, GeneMask, for MLM training of gene sequences, where we randomly identify positions in a gene sequence as mask centers and locally select the span around the mask center with the highest Normalized Pointwise Mutual Information (NPMI) to mask. We observe that in the absence of human-understandable semantics in the genomics domain (in contrast, semantic units like words and phrases are inherently available in NLP), GeneMask-based models substantially outperform the SOTA models (DNABert and LOGO) over four benchmark gene sequence classification datasets in five few-shot settings (10 to 1000-shot). More significantly, the GeneMask-based DNABert model is trained for less than one-tenth of the number of epochs of the original SOTA model. We also observe a strong correlation between top-ranked PMI tokens and conserved DNA sequence motifs, which may indicate the incorporation of latent genomic information. The codes (including trained models) and datasets are made publicly available at https://github.com/roysoumya/GeneMask.

cs.CL↗

Knowledge-Aware Neural Networks for Medical Forum Question Classification

Online medical forums have become a predominant platform for answering health-related information needs of consumers. However, with a significant rise in the number of queries and the limited availability of experts, it is necessary to automatically classify medical queries based on a consumer's intention, so that these questions may be directed to the right set of medical experts. Here, we develop a novel medical knowledge-aware BERT-based model (MedBERT) that explicitly gives more weightage to medical concept-bearing words, and utilize domain-specific side information obtained from a popular medical knowledge base. We also contribute a multi-label dataset for the Medical Forum Question Classification (MFQC) task. MedBERT achieves state-of-the-art performance on two benchmark datasets and performs very well in low resource settings.

cs.CL↗

An Integrated Approach for Improving Brand Consistency of Web Content: Modeling, Analysis and Recommendation

A consumer-dependent (business-to-consumer) organization tends to present itself as possessing a set of human qualities, which is termed as the brand personality of the company. The perception is impressed upon the consumer through the content, be it in the form of advertisement, blogs or magazines, produced by the organization. A consistent brand will generate trust and retain customers over time as they develop an affinity towards regularity and common patterns. However, maintaining a consistent messaging tone for a brand has become more challenging with the virtual explosion in the amount of content which needs to be authored and pushed to the Internet to maintain an edge in the era of digital marketing. To understand the depth of the problem, we collect around 300K web page content from around 650 companies. We develop trait-specific classification models by considering the linguistic features of the content. The classifier automatically identifies the web articles which are not consistent with the mission and vision of a company and further helps us to discover the conditions under which the consistency cannot be maintained. To address the brand inconsistency issue, we then develop a sentence ranking system that outputs the top three sentences that need to be changed for making a web article more consistent with the company's brand personality.

cs.CL↗