arXiv subjects

Manish Bhattarai

Publications and source records attributed to Manish Bhattarai.

At least 19 recordsLinked to original sources

arXiv2026

Rubric-Grounded RL: Structured Judge Rewards for Generalizable Reasoning

We argue that decomposing reward into weighted, verifiable criteria and using an LLM judge to score them provides a partial-credit optimization signal: instead of a binary outcome or a single holistic score, each response is graded along multiple task-specific criteria. We formalize \emph{rubric-grounded reinforcement learning (RL)}: a framework in which the policy is optimized against a structured, multi-criterion reward produced by a frozen LLM judge that conditions on auxiliary grounding the policy never sees. We instantiate the framework by deriving rubrics from an Office of Scientific and Technical Information (OSTI)-derived corpus of roughly 100,000 scientific and technical documents and training Llama-3.1-8B-Instruct with Group Relative Policy Optimization (GRPO). With GRPO-based training, the model achieves $71.7\%$ normalized reward on held-out rubric evaluation. The GRPO-tuned policy also improves over the base model on four reasoning benchmarks not derived from the training corpus -- GSM8K, MATH, GPQA Main, and GPQA Diamond. These results provide evidence that structured, document-grounded rewards can improve held-out rubric performance and induce transferable reasoning behaviors beyond the corpus used to construct the training environment.

cs.AI↗

arXiv2026

Sanity Checks for Long-Form Hallucination Detection

Hallucination detection methods for large language models increasingly operate on chain-of-thought reasoning traces, yet it remains unclear whether they evaluate the reasoning itself or merely exploit surface correlates of the final answer. We introduce a controlled-invariance methodology that exposes this distinction through two oracle tests: \textsc{Force}, which replaces each response's final answer with the ground truth while preserving the reasoning trace, and \textsc{Remove}, which strips answer-announcement steps while leaving the trajectory intact. This reveals if their predictive power derives from answer-level artifacts rather than from the structure or validity of intermediate reasoning. We further show that once these artifacts are controlled for, effective detection does not necessarily require complex learned representations: TRACT, a lightweight scorer built on lexical trajectory features (hedging trends, step-length dynamics, and cross-response vocabulary convergence), achieves strong robustness while remaining competitive with or outperforming existing baselines on unperturbed traces. These findings suggest that the current central challenge in reasoning-aware hallucination detection is not the absence of signal in the trace, but the failure to isolate it from endpoint cues.

cs.CL↗

arXiv2026

Trustworthy Agentic AI Requires Deterministic Architectural Boundaries

Current agentic AI architectures are fundamentally incompatible with the security and epistemological requirements of high-stakes scientific workflows. The problem is not inadequate alignment or insufficient guardrails, it is architectural: autoregressive language models process all tokens uniformly, making deterministic command--data separation unattainable through training alone. We argue that deterministic, architectural enforcement, not probabilistic learned behavior, is a necessary condition for trustworthy AI-assisted science. We introduce the Trinity Defense Architecture, which enforces security through three mechanisms: action governance via a finite action calculus with reference-monitor enforcement, information-flow control via mandatory access labels preventing cross-scope leakage, and privilege separation isolating perception from execution. We show that without unforgeable provenance and deterministic mediation, the ``Lethal Trifecta'' (untrusted inputs, privileged data access, external action capability) turns authorization security into an exploit-discovery problem: training-based defenses may reduce empirical attack rates but cannot provide deterministic guarantees. The ML community must recognize that alignment is insufficient for authorization security, and that architectural mediation is required before agentic AI can be safely deployed in consequential scientific domains.

cs.CR↗

arXiv2026

Privacy Enhanced PEFT: Tensor Train Decomposition Improves Privacy Utility Tradeoffs under DP-SGD

Fine-tuning large language models on sensitive data poses significant privacy risks, as membership inference attacks can reveal whether individual records were used during training. While Differential Privacy (DP) provides formal protection, applying DP to conventional Parameter-Efficient Fine-Tuning (PEFT) methods such as Low-Rank Adaptation (LoRA) often incurs substantial utility loss. In this work, we show that a more structurally constrained PEFT architecture, Tensor Train Low-Rank Adaptation (TTLoRA), can improve the privacy-utility tradeoff by shrinking the effective parameter space while preserving expressivity. To this end, we develop TTLoRA-DP, a differentially private training framework for TTLoRA. Specifically, we extend the ghost clipping algorithm to Tensor Train cores via cached contraction states, enabling efficient Differentially Private Stochastic Gradient Descent (DP-SGD) with exact per-example gradient norm computation without materializing full per-example gradients. Experiments on GPT-2 fine-tuning over the Enron and Penn Treebank datasets show that TTLoRA-DP consistently strengthens privacy protection relative to LoRA-DP while maintaining comparable or better downstream utility. Moreover, TTLoRA exhibits lower membership leakage even without DP training, using substantially smaller adapters and requiring on average 7.6X fewer parameters than LoRA. Overall, our results demonstrate that TTLoRA offers a practical path to improving the privacy-utility tradeoff in parameter-efficient language model adaptation.

cs.CR↗

arXiv2025

PAS : Prelim Attention Score for Detecting Object Hallucinations in Large Vision--Language Models

Large vision-language models (LVLMs) are powerful, yet they remain unreliable due to object hallucinations. In this work, we show that in many hallucinatory predictions the LVLM effectively ignores the image and instead relies on previously generated output (prelim) tokens to infer new objects. We quantify this behavior via the mutual information between the image and the predicted object conditioned on the prelim, demonstrating that weak image dependence strongly correlates with hallucination. Building on this finding, we introduce the Prelim Attention Score (PAS), a lightweight, training-free signal computed from attention weights over prelim tokens. PAS requires no additional forward passes and can be computed on the fly during inference. Exploiting this previously overlooked signal, PAS achieves state-of-the-art object-hallucination detection across multiple models and datasets, enabling real-time filtering and intervention.

cs.CV↗

arXiv2025

HalluField: Detecting LLM Hallucinations via Field-Theoretic Modeling

Large Language Models (LLMs) exhibit impressive reasoning and question-answering capabilities. However, they often produce inaccurate or unreliable content known as hallucinations. This unreliability significantly limits their deployment in high-stakes applications. Thus, there is a growing need for a general-purpose method to detect hallucinations in LLMs. In this work, we introduce HalluField, a novel field-theoretic approach for hallucination detection based on a parametrized variational principle and thermodynamics. Inspired by thermodynamics, HalluField models an LLM's response to a given query and temperature setting as a collection of discrete likelihood token paths, each associated with a corresponding energy and entropy. By analyzing how energy and entropy distributions vary across token paths under changes in temperature and likelihood, HalluField quantifies the semantic stability of a response. Hallucinations are then detected by identifying unstable or erratic behavior in this energy landscape. HalluField is computationally efficient and highly practical: it operates directly on the model's output logits without requiring fine-tuning or auxiliary neural networks. Notably, the method is grounded in a principled physical interpretation, drawing analogies to the first law of thermodynamics. Remarkably, by modeling LLM behavior through this physical lens, HalluField achieves state-of-the-art hallucination detection performance across models and datasets.

cs.LG↗

arXiv2025

MTRE: Multi-Token Reliability Estimation for Hallucination Detection in VLMs

Vision-language models (VLMs) now rival human performance on many multimodal tasks, yet they still hallucinate objects or generate unsafe text. Current hallucination detectors, e.g., single-token linear probing (LP) and PTrue, typically analyze only the logit of the first generated token or just its highest-scoring component, overlooking richer signals embedded within earlier token distributions. We demonstrate that analyzing the complete sequence of early logits potentially provides substantially more diagnostic information. We emphasize that hallucinations may only emerge after several tokens, as subtle inconsistencies accumulate over time. By analyzing the Kullback-Leibler (KL) divergence between logits corresponding to hallucinated and non-hallucinated tokens, we underscore the importance of incorporating later-token logits to more accurately capture the reliability dynamics of VLMs. In response, we introduce Multi-Token Reliability Estimation (MTRE), a lightweight, white-box method that aggregates logits from the first ten tokens using multi-token log-likelihood ratios and self-attention. Despite the challenges posed by large vocabulary sizes and long logit sequences, MTRE remains efficient and tractable. Across MAD-Bench, MM-SafetyBench, MathVista, and four compositional-geometry benchmarks, MTRE achieves a 9.4% gain in accuracy and a 14.8% gain in AUROC over standard detection methods, establishing a new state of the art in hallucination detection for open-source VLMs.

cs.AI↗

arXiv2025

Lost in OCR Translation? Vision-Based Approaches to Robust Document Retrieval

Retrieval-Augmented Generation (RAG) has become a popular technique for enhancing the reliability and utility of Large Language Models (LLMs) by grounding responses in external documents. Traditional RAG systems rely on Optical Character Recognition (OCR) to first process scanned documents into text. However, even state-of-the-art OCRs can introduce errors, especially in degraded or complex documents. Recent vision-language approaches, such as ColPali, propose direct visual embedding of documents, eliminating the need for OCR. This study presents a systematic comparison between a vision-based RAG system (ColPali) and more traditional OCR-based pipelines utilizing Llama 3.2 (90B) and Nougat OCR across varying document qualities. Beyond conventional retrieval accuracy metrics, we introduce a semantic answer evaluation benchmark to assess end-to-end question-answering performance. Our findings indicate that while vision-based RAG performs well on documents it has been fine-tuned on, OCR-based RAG is better able to generalize to unseen documents of varying quality. We highlight the key trade-offs between computational efficiency and semantic accuracy, offering practical guidance for RAG practitioners in selecting between OCR-dependent and vision-based document retrieval systems in production environments.

cs.CV↗

arXiv2025

ARCS: Agentic Retrieval-Augmented Code Synthesis with Iterative Refinement

We present Agentic Retrieval-Augmented Code Synthesis (ARCS), a system that improves LLM-based code generation without fine-tuning. ARCS operates through a budgeted synthesize-execute-repair loop over a frozen model: it retrieves relevant code context before generation, proposes candidates, executes them against tests, and repairs based on execution feedback. This retrieval-before-generation design reduces hallucination and accelerates convergence. We formalize ARCS as a state-action process with provable guarantees on termination, monotonic improvement, and bounded cost. A tiered controller (Small/Medium/Large) trades latency for accuracy predictably. On HumanEval, ARCS achieves up to 87.2% pass@1 with Llama-3.1-405B, surpassing CodeAgent (82.3%) while using simpler control than tree-search methods. On TransCoder, it achieves >= 90% accuracy on most translation pairs. On a LANL scientific corpus, it improves CodeBLEU by +0.115 over baseline RAG. ARCS provides a practical, reproducible approach to reliable code synthesis using existing LLM checkpoints.

cs.SE↗

arXiv2025

TT-LoRA MoE: Unifying Parameter-Efficient Fine-Tuning and Sparse Mixture-of-Experts

We propose Tensor-Trained Low-Rank Adaptation Mixture of Experts (TT-LoRA MoE), a novel computational framework integrating Parameter-Efficient Fine-Tuning (PEFT) with sparse MoE routing to address scalability challenges in large model deployments. Unlike traditional MoE approaches, which face substantial computational overhead as expert counts grow, TT-LoRA MoE decomposes training into two distinct, optimized stages. First, we independently train lightweight, tensorized low-rank adapters (TT-LoRA experts), each specialized for specific tasks. Subsequently, these expert adapters remain frozen, eliminating inter-task interference and catastrophic forgetting in multi-task setting. A sparse MoE router, trained separately, dynamically leverages base model representations to select exactly one specialized adapter per input at inference time, automating expert selection without explicit task specification. Comprehensive experiments confirm our architecture retains the memory efficiency of low-rank adapters, seamlessly scales to large expert pools, and achieves robust task-level optimization. This structured decoupling significantly enhances computational efficiency and flexibility: uses only 2% of LoRA, 0.3% of Adapters and 0.03% of AdapterFusion parameters and outperforms AdapterFusion by 4 value in multi-tasking, enabling practical and scalable multi-task inference deployments.

cs.LG↗

arXiv2025

Topological Signatures of Adversaries in Multimodal Alignments

Multimodal Machine Learning systems, particularly those aligning text and image data like CLIP/BLIP models, have become increasingly prevalent, yet remain susceptible to adversarial attacks. While substantial research has addressed adversarial robustness in unimodal contexts, defense strategies for multimodal systems are underexplored. This work investigates the topological signatures that arise between image and text embeddings and shows how adversarial attacks disrupt their alignment, introducing distinctive signatures. We specifically leverage persistent homology and introduce two novel Topological-Contrastive losses based on Total Persistence and Multi-scale kernel methods to analyze the topological signatures introduced by adversarial perturbations. We observe a pattern of monotonic changes in the proposed topological losses emerging in a wide range of attacks on image-text alignments, as more adversarial samples are introduced in the data. By designing an algorithm to back-propagate these signatures to input samples, we are able to integrate these signatures into Maximum Mean Discrepancy tests, creating a novel class of tests that leverage topological signatures for better adversarial detection.

cs.LG↗

arXiv2024

Patchfinder: Leveraging Visual Language Models for Accurate Information Retrieval using Model Uncertainty

For decades, corporations and governments have relied on scanned documents to record vast amounts of information. However, extracting this information is a slow and tedious process due to the sheer volume and complexity of these records. The rise of Vision Language Models (VLMs) presents a way to efficiently and accurately extract the information out of these documents. The current automated workflow often requires a two-step approach involving the extraction of information using optical character recognition software and subsequent usage of large language models for processing this information. Unfortunately, these methods encounter significant challenges when dealing with noisy scanned documents, often requiring computationally expensive language models to handle high information density effectively. In this study, we propose PatchFinder, an algorithm that builds upon VLMs to improve information extraction. First, we devise a confidence-based score, called Patch Confidence, based on the Maximum Softmax Probability of the VLMs' output to measure the model's confidence in its predictions. Using this metric, PatchFinder determines a suitable patch size, partitions the input document into overlapping patches, and generates confidence-based predictions for the target information. Our experimental results show that PatchFinder, leveraging Phi-3v, a 4.2-billion-parameter VLM, achieves an accuracy of 94% on our dataset of 190 noisy scanned documents, outperforming ChatGPT-4o by 18.5 percentage points.

cs.CV↗

arXiv2024

Enhancing Cross-Language Code Translation via Task-Specific Embedding Alignment in Retrieval-Augmented Generation

We introduce a novel method to enhance cross-language code translation from Fortran to C++ by integrating task-specific embedding alignment into a Retrieval-Augmented Generation (RAG) framework. Unlike conventional retrieval approaches that utilize generic embeddings agnostic to the downstream task, our strategy aligns the retrieval model directly with the objective of maximizing translation quality, as quantified by the CodeBLEU metric. This alignment ensures that the embeddings are semantically and syntactically meaningful for the specific code translation task. Our methodology involves constructing a dataset of 25,000 Fortran code snippets sourced from Stack-V2 dataset and generating their corresponding C++ translations using the LLaMA 3.1-8B language model. We compute pairwise CodeBLEU scores between the generated translations and ground truth examples to capture fine-grained similarities. These scores serve as supervision signals in a contrastive learning framework, where we optimize the embedding model to retrieve Fortran-C++ pairs that are most beneficial for improving the language model's translation performance. By integrating these CodeBLEU-optimized embeddings into the RAG framework, our approach significantly enhances both retrieval accuracy and code generation quality over methods employing generic embeddings. On the HPC Fortran2C++ dataset, our method elevates the average CodeBLEU score from 0.64 to 0.73, achieving a 14% relative improvement. On the Numerical Recipes dataset, we observe an increase from 0.52 to 0.60, marking a 15% relative improvement. Importantly, these gains are realized without any fine-tuning of the language model, underscoring the efficiency and practicality of our approach.

cs.AI↗

arXiv2024

HEAL: Hierarchical Embedding Alignment Loss for Improved Retrieval and Representation Learning

Retrieval-Augmented Generation (RAG) enhances Large Language Models (LLMs) by integrating external document retrieval to provide domain-specific or up-to-date knowledge. The effectiveness of RAG depends on the relevance of retrieved documents, which is influenced by the semantic alignment of embeddings with the domain's specialized content. Although full fine-tuning can align language models to specific domains, it is computationally intensive and demands substantial data. This paper introduces Hierarchical Embedding Alignment Loss (HEAL), a novel method that leverages hierarchical fuzzy clustering with matrix factorization within contrastive learning to efficiently align LLM embeddings with domain-specific content. HEAL computes level/depth-wise contrastive losses and incorporates hierarchical penalties to align embeddings with the underlying relationships in label hierarchies. This approach enhances retrieval relevance and document classification, effectively reducing hallucinations in LLM outputs. In our experiments, we benchmark and evaluate HEAL across diverse domains, including Healthcare, Material Science, Cyber-security, and Applied Maths.

cs.IR↗

arXiv2024

Benchmarking Large Language Models with Integer Sequence Generation Tasks

We present a novel benchmark designed to rigorously evaluate the capabilities of large language models (LLMs) in mathematical reasoning and algorithmic code synthesis tasks. The benchmark comprises integer sequence generation tasks sourced from the Online Encyclopedia of Integer Sequences (OEIS), testing LLMs' abilities to accurately and efficiently generate Python code to compute these sequences without using lookup tables. Our comprehensive evaluation includes leading models from OpenAI (including the specialized reasoning-focused o-series), Anthropic, Meta, and Google across a carefully selected set of 1000 OEIS sequences categorized as ``easy'' or ``hard.'' Half of these sequences are classical sequences from the early days of OEIS and half were recently added to avoid contamination with the models' training data. To prevent models from exploiting memorized sequence values, we introduce an automated cheating detection mechanism that flags usage of lookup tables, validated by comparison with human expert evaluations. Experimental results demonstrate that reasoning-specialized models (o3, o3-mini, o4-mini from OpenAI, and Gemini 2.5-pro from Google) achieve substantial improvements in accuracy over non-reasoning models, especially on more complex tasks. However, overall model performance on the hard sequences is poor, highlighting persistent challenges in algorithmic reasoning. Our benchmark provides important insights into the strengths and limitations of state-of-the-art LLMs, particularly emphasizing the necessity for further advancements to reliably solve complex mathematical reasoning tasks algorithmically.

cs.LG↗

arXiv2024

Domain-Specific Retrieval-Augmented Generation Using Vector Stores, Knowledge Graphs, and Tensor Factorization

Large Language Models (LLMs) are pre-trained on large-scale corpora and excel in numerous general natural language processing (NLP) tasks, such as question answering (QA). Despite their advanced language capabilities, when it comes to domain-specific and knowledge-intensive tasks, LLMs suffer from hallucinations, knowledge cut-offs, and lack of knowledge attributions. Additionally, fine tuning LLMs' intrinsic knowledge to highly specific domains is an expensive and time consuming process. The retrieval-augmented generation (RAG) process has recently emerged as a method capable of optimization of LLM responses, by referencing them to a predetermined ontology. It was shown that using a Knowledge Graph (KG) ontology for RAG improves the QA accuracy, by taking into account relevant sub-graphs that preserve the information in a structured manner. In this paper, we introduce SMART-SLIC, a highly domain-specific LLM framework, that integrates RAG with KG and a vector store (VS) that store factual domain specific information. Importantly, to avoid hallucinations in the KG, we build these highly domain-specific KGs and VSs without the use of LLMs, but via NLP, data mining, and nonnegative tensor factorization with automatic model selection. Pairing our RAG with a domain-specific: (i) KG (containing structured information), and (ii) VS (containing unstructured information) enables the development of domain-specific chat-bots that attribute the source of information, mitigate hallucinations, lessen the need for fine-tuning, and excel in highly domain-specific question answering tasks. We pair SMART-SLIC with chain-of-thought prompting agents. The framework is designed to be generalizable to adapt to any specific or specialized domain. In this paper, we demonstrate the question answering capabilities of our framework on a corpus of scientific publications on malware analysis and anomaly detection.

cs.CL↗

arXiv2024

LoRID: Low-Rank Iterative Diffusion for Adversarial Purification

This work presents an information-theoretic examination of diffusion-based purification methods, the state-of-the-art adversarial defenses that utilize diffusion models to remove malicious perturbations in adversarial examples. By theoretically characterizing the inherent purification errors associated with the Markov-based diffusion purifications, we introduce LoRID, a novel Low-Rank Iterative Diffusion purification method designed to remove adversarial perturbation with low intrinsic purification errors. LoRID centers around a multi-stage purification process that leverages multiple rounds of diffusion-denoising loops at the early time-steps of the diffusion models, and the integration of Tucker decomposition, an extension of matrix factorization, to remove adversarial noise at high-noise regimes. Consequently, LoRID increases the effective diffusion time-steps and overcomes strong adversarial attacks, achieving superior robustness performance in CIFAR-10/100, CelebA-HQ, and ImageNet datasets under both white-box and black-box settings.

cs.LG↗

arXiv2024

LaFA: Latent Feature Attacks on Non-negative Matrix Factorization

As Machine Learning (ML) applications rapidly grow, concerns about adversarial attacks compromising their reliability have gained significant attention. One unsupervised ML method known for its resilience to such attacks is Non-negative Matrix Factorization (NMF), an algorithm that decomposes input data into lower-dimensional latent features. However, the introduction of powerful computational tools such as Pytorch enables the computation of gradients of the latent features with respect to the original data, raising concerns about NMF's reliability. Interestingly, naively deriving the adversarial loss for NMF as in the case of ML would result in the reconstruction loss, which can be shown theoretically to be an ineffective attacking objective. In this work, we introduce a novel class of attacks in NMF termed Latent Feature Attacks (LaFA), which aim to manipulate the latent features produced by the NMF process. Our method utilizes the Feature Error (FE) loss directly on the latent features. By employing FE loss, we generate perturbations in the original data that significantly affect the extracted latent features, revealing vulnerabilities akin to those found in other ML techniques. To handle large peak-memory overhead from gradient back-propagation in FE attacks, we develop a method based on implicit differentiation which enables their scaling to larger datasets. We validate NMF vulnerabilities and FE attacks effectiveness through extensive experiments on synthetic and real-world data.

cs.LG↗