Alfy Samuel

Duygu Nur Yaldiz, Yavuz Faruk Bakman, Sungmin Kang et al.

Generative Large Language Models (LLMs)inevitably produce untruthful responses. Accurately predicting the truthfulness of these outputs is critical, especially in high-stakes settings. To accelerate research in this domain and make truthfulness prediction methods more accessible, we introduce TruthTorchLM an open-source, comprehensive Python library featuring over 30 truthfulness prediction methods, which we refer to as Truth Methods. Unlike existing toolkits such as Guardrails, which focus solely on document-grounded verification, or LM-Polygraph, which is limited to uncertainty-based methods, TruthTorchLM offers a broad and extensible collection of techniques. These methods span diverse tradeoffs in computational cost, access level (e.g., black-box vs white-box), grounding document requirements, and supervision type (self-supervised or supervised). TruthTorchLM is seamlessly compatible with both HuggingFace and LiteLLM, enabling support for locally hosted and API-based models. It also provides a unified interface for generation, evaluation, calibration, and long-form truthfulness prediction, along with a flexible framework for extending the library with new methods. We conduct an evaluation of representative truth methods on three datasets, TriviaQA, GSM8K, and FactScore-Bio. The code is available at https://github.com/Ybakman/TruthTorchLM

Mengxuan Hu, Vivek V. Datla, Anoop Kumar et al.

Recent advances in alignment techniques such as Supervised Fine-Tuning (SFT), Reinforcement Learning from Human Feedback (RLHF), and Direct Preference Optimization (DPO) have improved the safety of large language models (LLMs). However, these LLMs remain vulnerable to jailbreak attacks that disguise harmful intent through indirect or deceptive phrasing. Using causal intervention, we empirically demonstrate that this vulnerability stems from shallow alignment mechanisms that lack deep reasoning, often rejecting harmful prompts without truly understanding why they are harmful. To mitigate this vulnerability, we propose enhancing alignment through reasoning-aware post-training. We construct and release a novel Chain-of-Thought (CoT) fine-tuning dataset that includes both utility-oriented and safety-critical prompts with step-by-step rationales. Fine-tuning on this dataset encourages models to produce principled refusals grounded in reasoning, outperforming standard SFT baselines. Furthermore, inspired by failure patterns in CoT fine-tuning, we introduce Alignment-Weighted DPO, which targets the most problematic parts of an output by assigning different preference weights to the reasoning and final-answer segments. This produces finer-grained, targeted updates than vanilla DPO and improves robustness to diverse jailbreak strategies. Extensive experiments across multiple safety and utility benchmarks show that our method consistently improves alignment robustness while maintaining overall model utility.

Parker Glenn, Alfy Samuel, Daben Liu

Integrating LLM powered operators in declarative query languages allows for the combination of cheap and interpretable functions with powerful, generalizable language model reasoning. However, in order to benefit from the optimized execution of a database query language like SQL, generated outputs must align with the rules enforced by both type checkers and database contents. Current approaches address this challenge with orchestrations consisting of many LLM-based post-processing calls to ensure alignment between generated outputs and database values, introducing performance bottlenecks. We perform a study on the ability of various sized open-source language models to both parse and execute functions within a query language based on SQL, showing that small language models can excel as function executors over hybrid data sources. Then, we propose an efficient solution to enforce the well-typedness of LLM functions, demonstrating 7% accuracy improvement on a multi-hop question answering dataset with 53% improvement in latency over comparable solutions. We make our implementation available at https://github.com/parkervg/blendsql

Nikhil Maddikunta, Huijun Zhao, Sumit Keswani et al.

In the past, computer vision systems for digitized documents could rely on systematically captured, high-quality scans. Today, transactions involving digital documents are more likely to start as mobile phone photo uploads taken by non-professionals. As such, computer vision for document automation must now account for documents captured in natural scene contexts. An additional challenge is that task objectives for document processing can be highly use-case specific, which makes publicly-available datasets limited in their utility, while manual data labeling is also costly and poorly translates between use cases. To address these issues we created Sim2Real Docs - a framework for synthesizing datasets and performing domain randomization of documents in natural scenes. Sim2Real Docs enables programmatic 3D rendering of documents using Blender, an open source tool for 3D modeling and ray-traced rendering. By using rendering that simulates physical interactions of light, geometry, camera, and background, we synthesize datasets of documents in a natural scene context. Each render is paired with use-case specific ground truth data specifying latent characteristics of interest, producing unlimited fit-for-task training data. The role of machine learning models is then to solve the inverse problem posed by the rendering pipeline. Such models can be further iterated upon with real-world data by either fine tuning or making adjustments to domain randomization parameters.

Amin Banayeeanzade, Qingchuan Yang, Dhruv Tarsadiya et al.

Diversity is essential for language-model applications ranging from creative generation to scientific discovery, yet modern LLMs often collapse into a narrow subset of plausible outputs. While prior work has developed benchmarks for measuring this lack of diversity, less is known about how the step-by-step probability distributions at inference time cause the problem. We introduce a validity--diversity framework that attributes diversity collapse to how an LLM allocates probability mass across valid and invalid continuations during decoding. This framework decomposes the bottleneck into two complementary forms of miscalibration. First, order calibration: valid tokens are not reliably ranked above invalid tokens, so rank-based cutoff rules must trade off between recovering valid continuations and admitting invalid ones. Second, shape calibration: probability mass is overly concentrated only on few valid continuations while having a heavy-tail of mixed valid and invalid tokens, so maintaining high validity limits diversity. We formalize both mechanisms and show that local failures compound across decoding steps, producing strong sequence-level losses in diversity. Empirically, we develop controlled diagnostics for probing these bottlenecks, including tasks with exactly known valid sets and oracle cutoff baselines. Across 14 language models spanning multiple families and scales, we find that diversity collapse is not merely a limitation of particular sampling heuristics, but a consequence of order and shape miscalibration in the LLM distribution.

Kushal Chawla, Chenyang Zhu, Pengshan Cai et al.

Summarization of multi-party dialogues is a critical capability in industry, enhancing knowledge transfer and operational effectiveness across many domains. However, automatically generating high-quality summaries is challenging, as the ideal summary must satisfy a set of complex, multi-faceted requirements. While summarization has received immense attention in research, prior work has primarily utilized static datasets and benchmarks, a condition rare in practical scenarios where requirements inevitably evolve. In this work, we present an industry case study on developing an agentic system to summarize multi-party interactions. We share practical insights spanning the full development lifecycle to guide practitioners in building reliable, adaptable summarization systems, as well as to inform future research, covering: 1) robust methods for evaluation despite evolving requirements and task subjectivity, 2) component-wise optimization enabled by the task decomposition inherent in an agentic architecture, 3) the impact of upstream data bottlenecks, and 4) the realities of vendor lock-in due to the poor transferability of LLM prompts.

Jingyu Wu, Aditya Shrivastava, Jing Zhu et al.

Efficiently reranking documents retrieved from information retrieval (IR) pipelines to enhance overall quality of Retrieval-Augmented Generation (RAG) system remains an important yet challenging problem. Recent studies have highlighted the importance of Large Language Models (LLMs) in reranking tasks. In particular, Pairwise Reranking Prompting (PRP) has emerged as a promising plug-and-play approach due to its usability and effectiveness. However, the inherent complexity of the algorithm, coupled with the high computational demands and latency incurred due to LLMs, raises concerns about its feasibility in real-time applications. To address these challenges, this paper presents a focused study on pairwise reranking, demonstrating that carefully applied optimization methods can significantly mitigate these issues. By implementing these methods, we achieve a remarkable latency reduction of up to 166 times, from 61.36 seconds to 0.37 seconds per query, with an insignificant drop in performance measured by Recall@k. Our study highlights the importance of design choices that were previously overlooked, such as using smaller models, limiting the reranked set, using lower precision, reducing positional bias with one-directional order inference, and restricting output tokens. These optimizations make LLM-based reranking substantially more efficient and feasible for latency-sensitive, real-world deployments.

Neel Jain, Aditya Shrivastava, Chenyang Zhu et al.

A key component of building safe and reliable language models is enabling the models to appropriately refuse to follow certain instructions or answer certain questions. We may want models to output refusal messages for various categories of user queries, for example, ill-posed questions, instructions for committing illegal acts, or queries which require information past the model's knowledge horizon. Engineering models that refuse to answer such questions is complicated by the fact that an individual may want their model to exhibit varying levels of sensitivity for refusing queries of various categories, and different users may want different refusal rates. The current default approach involves training multiple models with varying proportions of refusal messages from each category to achieve the desired refusal rates, which is computationally expensive and may require training a new model to accommodate each user's desired preference over refusal rates. To address these challenges, we propose refusal tokens, one such token for each refusal category or a single refusal token, which are prepended to the model's responses during training. We then show how to increase or decrease the probability of generating the refusal token for each category during inference to steer the model's refusal behavior. Refusal tokens enable controlling a single model's refusal rates without the need of any further fine-tuning, but only by selectively intervening during generation.

Yavuz Bakman, Sungmin Kang, Zhiqi Huang et al.

Uncertainty Quantification (UQ) research has primarily focused on closed-book factual question answering (QA), while contextual QA remains unexplored, despite its importance in real-world applications. In this work, we focus on UQ for the contextual QA task and propose a theoretically grounded approach to quantify epistemic uncertainty. We begin by introducing a task-agnostic, token-level uncertainty measure defined as the cross-entropy between the predictive distribution of the given model and the unknown true distribution. By decomposing this measure, we isolate the epistemic component and approximate the true distribution by a perfectly prompted, idealized model. We then derive an upper bound for epistemic uncertainty and show that it can be interpreted as semantic feature gaps in the given model's hidden representations relative to the ideal model. We further apply this generic framework to the contextual QA task and hypothesize that three features approximate this gap: context-reliance (using the provided context rather than parametric knowledge), context comprehension (extracting relevant information from context), and honesty (avoiding intentional lies). Using a top-down interpretability approach, we extract these features by using only a small number of labeled samples and ensemble them to form a robust uncertainty score. Experiments on multiple QA benchmarks in both in-distribution and out-of-distribution settings show that our method substantially outperforms state-of-the-art unsupervised (sampling-free and sampling-based) and supervised UQ methods, achieving up to a 13-point PRR improvement while incurring a negligible inference overhead.

Catarina G Belem, Parker Glenn, Alfy Samuel et al.

Automatic readability assessment plays a key role in ensuring effective and accessible written communication. Despite significant progress, the field is hindered by inconsistent definitions of readability and measurements that rely on surface-level text properties. In this work, we investigate the factors shaping human perceptions of readability through the analysis of 897 judgments, finding that, beyond surface-level cues, information content and topic strongly shape text comprehensibility. Furthermore, we evaluate 15 popular readability metrics across five English datasets, contrasting them with six more nuanced, model-based metrics. Our results show that four model-based metrics consistently place among the top four in rank correlations with human judgments, while the best performing traditional metric achieves an average rank of 8.6. These findings highlight a mismatch between current readability metrics and human perceptions, pointing to model-based approaches as a more promising direction.

Zhiqi Huang, Vivek Datla, Chenyang Zhu et al.

We propose a method for confidence estimation in retrieval-augmented generation (RAG) systems that aligns closely with the correctness of large language model (LLM) outputs. Confidence estimation is especially critical in high-stakes domains such as finance and healthcare, where the cost of an incorrect answer outweighs that of not answering the question. Our approach extends prior uncertainty quantification methods by leveraging raw feed-forward network (FFN) activations as auto-regressive signals, avoiding the information loss inherent in token logits and probabilities after projection and softmax normalization. We model confidence prediction as a sequence classification task, and regularize training with a Huber loss term to improve robustness against noisy supervision. Applied in a real-world financial industry customer-support setting with complex knowledge bases, our method outperforms strong baselines and maintains high accuracy under strict latency constraints. Experiments on Llama 3.1 8B model show that using activations from only the 16th layer preserves accuracy while reducing response latency. Our results demonstrate that activation-based confidence modeling offers a scalable, architecture-aware path toward trustworthy RAG deployment.

Faisal Hamman, Chenyang Zhu, Anoop Kumar et al.

RAG systems are increasingly deployed in high-stakes domains where users expect outputs to be consistent across semantically equivalent queries. However, existing systems often exhibit significant inconsistencies due to variability in both the retriever and generator (LLM), undermining trust and reliability. In this work, we focus on information consistency, i.e., the requirement that outputs convey the same core content across semantically equivalent inputs. We introduce a principled evaluation framework that decomposes RAG consistency into retriever-level, generator-level, and end-to-end components, helping identify inconsistency sources. To improve consistency, we propose Paraphrased Set Group Relative Policy Optimization (PS-GRPO), an RL approach that leverages multiple rollouts across paraphrased set to assign group similarity rewards. We leverage PS-GRPO to achieve Information Consistent RAG (Con-RAG), training the generator to produce consistent outputs across paraphrased queries and remain robust to retrieval-induced variability. Because exact reward computation over paraphrase sets is computationally expensive, we also introduce a scalable approximation method that retains effectiveness while enabling efficient, large-scale training. Empirical evaluations across short-form, multi-hop, and long-form QA benchmarks demonstrate that Con-RAG significantly improves both consistency and accuracy over strong baselines, even in the absence of explicit ground-truth supervision. Our work provides practical solutions for evaluating and building reliable RAG systems for safety-critical deployments.

Neal Gregory Lawton, Alfy Samuel, Anoop Kumar et al.

A Comparison of Independent and Joint Fine-tuning Strategies for Retrieval-Augmented Generation Download PDF Neal Gregory Lawton, Alfy Samuel, Anoop Kumar, Daben Liu Published: 20 Aug 2025, Retrieval augmented generation (RAG) is a popular framework for question answering that is powered by two large language models (LLMs): an embedding model that retrieves context documents from a database that are relevant to a given question, and a generator model that uses the retrieved context to generate an answer to the question. Both the embedding and generator models can be fine-tuned to increase performance of a RAG pipeline on a new task, but multiple fine-tuning strategies exist with different costs and benefits. In this paper, we evaluate and compare several RAG fine-tuning strategies, including independent, joint, and two-phase fine-tuning. In our experiments, we observe that all of these strategies achieve about equal improvement in EM and F1 generation quality metrics, although they have significantly different computational costs. We conclude the optimal fine-tuning strategy to use depends on whether the training dataset includes context labels and whether a grid search over the learning rates for the embedding and generator models is required.

Kushal Chawla, Alfy Samuel, Anoop Kumar et al.

Traditional Retrieval-Augmented Generation (RAG) struggles with complex queries that lack strong signals to retrieve the most relevant context, forcing a trade-off between choosing a small context that misses key information and a large context that confuses the LLM. To address this, we propose Forward-Backward RAG (FB-RAG), a new training-free framework based on a simple yet powerful forward-looking strategy. FB-RAG employs a light-weight LLM to peek into potential future generations, using evidence from multiple sampled outputs to precisely identify the most relevant context for a final, more powerful generator. This improves performance without complex finetuning or Reinforcement Learning common in prior work. Across $9$ datasets from LongBench and $\infty$Bench, FB-RAG consistently delivers strong results. Further, the performance gains can be achieved with reduced latency due to a shorter, more focused prompt for the powerful generator. On EN.QA dataset, FB-RAG matches the leading baseline with over $48$% latency reduction or achieves an $8$% performance improvement with a $10$% latency reduction. Our analysis finds cases where even when the forward-looking LLM fails to generate correct answers, its attempts are sufficient to guide the final model to an accurate response, demonstrating how smaller LLMs can systematically improve the performance and efficiency of larger ones.