arXiv subjects

Pouya Pezeshkpour

Publications and source records attributed to Pouya Pezeshkpour.

At least 19 recordsLinked to original sources

arXiv2026

MAWILE: Multi-Axis Workbench for Inspecting LLM Evaluators

Large language model (LLM) judges provide a flexible and scalable method for evaluating model and agent outputs, but their verdicts can be sensitive to incidental changes in the evaluated response, judge instructions, and scoring rubric. Existing systems examine important subsets of these failure modes, but auditing a configured judge requires testing both the judge instrument and the items it evaluates. We introduce MAWILE, a developer-facing workbench for auditing judge sensitivity across four surfaces: the judge prompt, judge rubric, target-system input, and target-system output. Given a user-supplied judge and representative evaluation items, MAWILE constructs and validates controlled perturbations, re-executes the judge, and localizes the resulting sensitivity. Each perturbation declares whether the verdict should remain invariant or change in a specified direction, allowing the same system to measure both robustness to irrelevant variations and sensitivity to meaningful changes. MAWILE audits binary, ordinal, and pairwise judges without requiring gold labels. The code for this tool is available at: github.com/megagonlabs/mawile-judge.

cs.AI↗

arXiv2026

Align Then Adapt: Label-Efficient Adapter Learning for Asymmetric Dense Retrieval

Dense retrieval systems increasingly face an asymmetry between complex instruction-like queries and relatively simple, static document collections. While stronger embedders can better understand such queries, re-embedding large corpora or fine-tuning large models is often impractical. We propose Efficient Retrieval Adapter (ERA), a query-side adapter learning framework for re-index-free retrieval adaptation. ERA first aligns the embedding spaces of a strong query embedder and a lightweight document embedder using unlabeled corpus documents, and then adapts the aligned query representation with a small number of labeled query-document pairs. Across 126 MAIR retrieval tasks from six domains, ERA improves average nDCG@10 by up to 8.2 points in symmetric settings and by more than 12 points in asymmetric settings, while using substantially fewer labels than supervised adapter training. These results show that retrieval systems can benefit from stronger query understanding without updating backbone embedders or rebuilding document indexes.

cs.IR↗

arXiv2026

Reflective Prompt Tuning through Language Model Function-Calling

Large language models (LLMs) have become increasingly capable of following instructions and complex reasoning, making prompting a flexible interface for adapting models without parameter updates. Yet prompt design remains labor-intensive and highly sensitive to formatting, phrasing, and instruction order, motivating automated prompt optimization methods that reduce manual effort while preserving inference-time flexibility. However, existing methods often search over prompt candidates or use fixed critique-refine pipelines driven by individual examples or small batches, limiting their ability to capture systematic error patterns and make targeted edits grounded in failure history. We propose Reflective Prompt Tuning (RPT), a framework that uses LLM function calling to simulate the iterative workflow of human prompt engineers. An LLM optimizer calls a diagnostic function that evaluates the target model over an entire optimization set, summarizes recurring failure modes, and returns a structured diagnostic report. The optimizer uses this report, together with an accumulated memory of prior reports, to revise the prompt for the next iteration. RPT further supports confidence-aware optimization by using calibration signals in diagnostic feedback and final prompt selection. Across three reasoning tasks, RPT improves over initial prompts by up to 12.9 points, remains competitive with state of the art, and improves confidence calibration. Our analyses show that RPT is especially effective on multi-hop and mathematical reasoning, producing targeted prompt revisions that align with diagnosed failure patterns and lead to gains in task performance and calibration.

cs.CL↗

arXiv2026

Geometry-Aware Decoding with Wasserstein-Regularized Truncation and Mass Penalties for Large Language Models

Large language models (LLMs) must balance diversity and creativity against logical coherence in open-ended generation. Existing truncation-based samplers are effective but largely heuristic, relying mainly on probability mass and entropy while ignoring semantic geometry of the token space. We present Top-W, a geometry-aware truncation rule that uses Wasserstein distance-defined over token-embedding geometry-to keep the cropped distribution close to the original, while explicitly balancing retained probability mass against the entropy of the kept set. Our theory yields a simple closed-form structure for the fixed-potential subset update: depending on the mass-entropy trade-off, the optimal crop either collapses to a single token or takes the form of a one-dimensional prefix that can be found efficiently with a linear scan. We implement Top-W using efficient geometry-based potentials (nearest-set or k-NN) and pair it with an alternating decoding routine that keeps the standard truncation-and-sampling interface unchanged. Extensive experiments on four benchmarks (GSM8K, GPQA, AlpacaEval, and MT-Bench) across three instruction-tuned models show that Top-W consistently outperforms prior state-of-the-art decoding approaches achieving up to 33.7% improvement. Moreover, we find that Top-W not only improves accuracy-focused performance, but also boosts creativity under judge-based open-ended evaluation.

cs.CL↗

arXiv2026

AutoPyVerifier: Learning Compact Executable Verifiers for Large Language Model Outputs

Verification is becoming central to both reinforcement-learning-based training and inference-time control of large language models (LLMs). Yet current verifiers face a fundamental trade-off: LLM-based verifiers are expressive but hard to control and prone to error, while deterministic executable verifiers are reliable and interpretable but often limited in capability. We study the following question: given a development set of LLM outputs and labels for a target objective, such as correctness, can we automatically induce a minimal set of Python verifiers whose joint satisfaction closely matches that objective? We propose AutoPyVerifier, a framework that uses an LLM to synthesize candidate verifier functions and then refines them through search over a directed acyclic graph (DAG). By navigating the DAG, AutoPyVerifier systematically explores the space of deterministic executable verifiers and selects a compact verifier set whose joint satisfaction best approximates the target objective. Across mathematical reasoning, coding, function calling, and instruction-following benchmarks for several state-of-the-art LLMs, AutoPyVerifier improves target-objective prediction by up to 55.0 F1 points over the initial LLM-generated verifier sets. Additional analyses show that the most useful verification targets vary by benchmark and model, and that the DAG-based search shifts the learned verifier sets toward more structural and semantically grounded checks. We further show that exposing the discovered verifier set to an LLM as an external tool improves downstream accuracy by up to 17.0 points. We release our code

cs.CL↗

arXiv2026

From Proof to Program: Characterizing Tool-Induced Reasoning Hallucinations in Large Language Models

Tool-augmented Language Models (TaLMs) can invoke external tools to solve problems beyond their parametric capacity. However, it remains unclear whether these tool-enabled gains reflect trustworthy reasoning. Focusing on the Code Interpreter tool, we show that even when tools are selected and executed correctly, TaLMs treat tool outputs as substitutes for reasoning, producing solutions that appear correct but lack coherent justification. We term this failure mode Tool-Induced Myopia (TIM), and study it using PYMATH, a benchmark of 1,679 competition-level mathematical problems for which Python code is helpful but not sufficient. We further develop a multi-dimensional evaluation suite to quantify reasoning degradation in TaLMs relative to their non-tool counterparts. Our findings reveal that while TaLMs achieve up to a 19.3 percentage point gain in final-answer accuracy, their reasoning behavior consistently deteriorates (e.g., non-tool LLMs win up to 41.5% more often in pairwise comparisons of the reasoning process). This degradation intensifies with tool use; the more frequently a model invokes tools, the less coherent its reasoning becomes. Moreover, tool use shifts errors from arithmetic mistakes toward global reasoning failures (logic, assumption, creativity); with TIM present in ~55% of high-risk cases. Finally, we propose a preference-optimization-based framework that realigns TaLMs to use tools as assistive evidence, improving both final-answer accuracy and reasoning depth under tool use. Codes and data are available at: https://github.com/megagonlabs/TIM.

cs.CL↗

arXiv2026

Blue Data Intelligence Layer: Streaming Data and Agents for Multi-source Multi-modal Data-Centric Applications

NL2SQL systems aim to address the growing need for natural language interaction with data. However, real-world information rarely maps to a single SQL query because (1) users express queries iteratively (2) questions often span multiple data sources beyond the closed-world assumption of a single database, and (3) queries frequently rely on commonsense or external knowledge. Consequently, satisfying realistic data needs require integrating heterogeneous sources, modalities, and contextual data. In this paper, we present Blue's Data Intelligence Layer (DIL) designed to support multi-source, multi-modal, and data-centric applications. Blue is a compound AI system that orchestrates agents and data for enterprise settings. DIL serves as the data intelligence layer for agentic data processing, to bridge the semantic gap between user intent and available information by unifying structured enterprise data, world knowledge accessible through LLMs, and personal context obtained through interaction. At the core of DIL is a data registry that stores metadata for diverse data sources and modalities to enable both native and natural language queries. DIL treats LLMs, the Web, and the User as source 'databases', each with their own query interface, elevating them to first-class data sources. DIL relies on data planners to transform user queries into executable query plans. These plans are declarative abstractions that unify relational operators with other operators spanning multiple modalities. DIL planners support decomposition of complex requests into subqueries, retrieval from diverse sources, and finally reasoning and integration to produce final results. We demonstrate DIL through two interactive scenarios in which user queries dynamically trigger multi-source retrieval, cross-modal reasoning, and result synthesis, illustrating how compound AI systems can move beyond single database NL2SQL.

cs.AI↗

arXiv2026

Towards Reliable Benchmarking: A Contamination Free, Controllable Evaluation Framework for Multi-step LLM Function Calling

Existing benchmarks for tool-augmented language models (TaLMs) lack fine-grained control over task difficulty and remain vulnerable to data contamination. We present FuncBenchGen, a unified, contamination-free framework that evaluates TaLMs by generating synthetic multi-step tool-use tasks to stress-test TaLMs. The key idea is to cast tool use as traversal over a hidden function-dependency DAG where models must infer the correct sequence of calls to compute a target value. FuncBenchGen allows precise control over task difficulty (e.g., graph size, dependency depth, and distractor functions) while avoiding pretraining/test-time leakage. Our evaluation demonstrates reasoning-optimized models consistently outperform general-purpose models with GPT-5 significantly outperforming other available models. Performance declines sharply as dependency depth increases. Furthermore, connected distractors -- irrelevant functions sharing type-compatible variables with relevant functions -- prove especially difficult to handle. Also, strong models often make syntactically valid function calls but propagate incorrect or stale argument values across steps, revealing brittle state tracking by LLMs in multi-turn tool use. Motivated by this observation, we introduce a simple mitigation strategy that explicitly restates prior variable values to the agent at each step. Surprisingly, this lightweight change yields substantial gains across models. e.g., yielding an improvement in success rate from 62.5% to 81.3% for GPT-5.

cs.CL↗

arXiv2026

From Task Solving to Robust Real-World Adaptation in LLM Agents

Large language models are increasingly deployed as specialized agents that plan, call tools, and take actions over extended horizons. Yet many existing evaluations assume a "clean interface" where dynamics are specified and stable, tools and sensors are reliable, and success is captured by a single explicit objective-often overestimating real-world readiness. In practice, agents face underspecified rules, unreliable signals, shifting environments, and implicit, multi-stakeholder goals. The challenge is therefore not just solving tasks, but adapting while solving: deciding what to trust, what is wanted, when to verify, and when to fall back or escalate. We stress-test deployment-relevant robustness under four operational circumstances: partial observability, dynamic environments, noisy signals, and dynamic agent state. We benchmark agentic LLMs in a grid-based game with a simple goal but long-horizon execution. Episodes violate clean-interface assumptions yet remain solvable, forcing agents to infer rules, pay for information, adapt to environmental and internal shifts, and act cautiously under noise. Across five state-of-the-art LLM agents, we find large gaps between nominal task-solving and deployment-like robustness. Performance generally degrades as grid size and horizon increase, but rankings are unstable: weaker models can beat stronger ones when strategy matches the uncertainty regime. Despite no explicit instruction, agents trade off completion, efficiency, and penalty avoidance, suggesting partial objective inference. Ablations and feature analyses reveal model-specific sensitivities and failure drivers, motivating work on verification, safe action selection, and objective inference under partial observability, noise, and non-stationarity.

cs.CL↗

arXiv2025

From Single to Multi: How LLMs Hallucinate in Multi-Document Summarization

Although many studies have investigated and reduced hallucinations in large language models (LLMs) for single-document tasks, research on hallucination in multi-document summarization (MDS) tasks remains largely unexplored. Specifically, it is unclear how the challenges arising from handling multiple documents (e.g., repetition and diversity of information) affect models outputs. In this work, we investigate how hallucinations manifest in LLMs when summarizing topic-specific information from multiple documents. Since no benchmarks exist for investigating hallucinations in MDS, we use existing news and conversation datasets, annotated with topic-specific insights, to create two novel multi-document benchmarks. When evaluating 5 LLMs on our benchmarks, we observe that on average, up to 75% of the content in LLM-generated summary is hallucinated, with hallucinations more likely to occur towards the end of the summaries. Moreover, when summarizing non-existent topic-related information, gpt-3.5-turbo and GPT-4o still generate summaries about 79.35% and 44% of the time, raising concerns about their tendency to fabricate content. To understand the characteristics of these hallucinations, we manually evaluate 700+ insights and find that most errors stem from either failing to follow instructions or producing overly generic insights. Motivated by these observations, we investigate the efficacy of simple post-hoc baselines in mitigating hallucinations but find them only moderately effective. Our results underscore the need for more effective approaches to systematically mitigate hallucinations in MDS. We release our dataset and code at github.com/megagonlabs/Hallucination_MDS.

cs.CL↗

arXiv2025

Mixed Signals: Decoding VLMs' Reasoning and Underlying Bias in Vision-Language Conflict

Vision-language models (VLMs) have demonstrated impressive performance by effectively integrating visual and textual information to solve complex tasks. However, it is not clear how these models reason over the visual and textual data together, nor how the flow of information between modalities is structured. In this paper, we examine how VLMs reason by analyzing their biases when confronted with scenarios that present conflicting image and text cues, a common occurrence in real-world applications. To uncover the extent and nature of these biases, we build upon existing benchmarks to create five datasets containing mismatched image-text pairs, covering topics in mathematics, science, and visual descriptions. Our analysis shows that VLMs favor text in simpler queries but shift toward images as query complexity increases. This bias correlates with model scale, with the difference between the percentage of image- and text-preferred responses ranging from +56.8% (image favored) to -74.4% (text favored), depending on the task and model. In addition, we explore three mitigation strategies: simple prompt modifications, modifications that explicitly instruct models on how to handle conflicting information (akin to chain-of-thought prompting), and a task decomposition strategy that analyzes each modality separately before combining their results. Our findings indicate that the effectiveness of these strategies in identifying and mitigating bias varies significantly and is closely linked to the model's overall performance on the task and the specific modality in question.

cs.AI↗

arXiv2025

Insight-RAG: Enhancing LLMs with Insight-Driven Augmentation

Retrieval Augmented Generation (RAG) frameworks have shown significant promise in leveraging external knowledge to enhance the performance of large language models (LLMs). However, conventional RAG methods often retrieve documents based solely on surface-level relevance, leading to many issues: they may overlook deeply buried information within individual documents, miss relevant insights spanning multiple sources, and are not well-suited for tasks beyond traditional question answering. In this paper, we propose Insight-RAG, a novel framework designed to address these issues. In the initial stage of Insight-RAG, instead of using traditional retrieval methods, we employ an LLM to analyze the input query and task, extracting the underlying informational requirements. In the subsequent stage, a specialized LLM -- trained on the document database -- is queried to mine content that directly addresses these identified insights. Finally, by integrating the original query with the retrieved insights, similar to conventional RAG approaches, we employ a final LLM to generate a contextually enriched and accurate response. Using two scientific paper datasets, we created evaluation benchmarks targeting each of the mentioned issues and assessed Insight-RAG against traditional RAG pipeline. Our results demonstrate that the Insight-RAG pipeline successfully addresses these challenges, outperforming existing methods by a significant margin in most cases. These findings suggest that integrating insight-driven retrieval within the RAG framework not only enhances performance but also broadens the applicability of RAG to tasks beyond conventional question answering.

cs.CL↗

arXiv2025

LLMs Are Not Intelligent Thinkers: Introducing Mathematical Topic Tree Benchmark for Comprehensive Evaluation of LLMs

Large language models (LLMs) demonstrate impressive capabilities in mathematical reasoning. However, despite these achievements, current evaluations are mostly limited to specific mathematical topics, and it remains unclear whether LLMs are genuinely engaging in reasoning. To address these gaps, we present the Mathematical Topics Tree (MaTT) benchmark, a challenging and structured benchmark that offers 1,958 questions across a wide array of mathematical subjects, each paired with a detailed hierarchical chain of topics. Upon assessing different LLMs using the MaTT benchmark, we find that the most advanced model, GPT-4, achieved a mere 54\% accuracy in a multiple-choice scenario. Interestingly, even when employing Chain-of-Thought prompting, we observe mostly no notable improvement. Moreover, LLMs accuracy dramatically reduced by up to 24.2 percentage point when the questions were presented without providing choices. Further detailed analysis of the LLMs' performance across a range of topics showed significant discrepancy even for closely related subtopics within the same general mathematical area. In an effort to pinpoint the reasons behind LLMs performances, we conducted a manual evaluation of the completeness and correctness of the explanations generated by GPT-4 when choices were available. Surprisingly, we find that in only 53.3\% of the instances where the model provided a correct answer, the accompanying explanations were deemed complete and accurate, i.e., the model engaged in genuine reasoning.

cs.CL↗

arXiv2025

Evaluating Bias in LLMs for Job-Resume Matching: Gender, Race, and Education

Large Language Models (LLMs) offer the potential to automate hiring by matching job descriptions with candidate resumes, streamlining recruitment processes, and reducing operational costs. However, biases inherent in these models may lead to unfair hiring practices, reinforcing societal prejudices and undermining workplace diversity. This study examines the performance and fairness of LLMs in job-resume matching tasks within the English language and U.S. context. It evaluates how factors such as gender, race, and educational background influence model decisions, providing critical insights into the fairness and reliability of LLMs in HR applications. Our findings indicate that while recent models have reduced biases related to explicit attributes like gender and race, implicit biases concerning educational background remain significant. These results highlight the need for ongoing evaluation and the development of advanced bias mitigation strategies to ensure equitable hiring practices when using LLMs in industry settings.

cs.CL↗

arXiv2025

Learning Beyond the Surface: How Far Can Continual Pre-Training with LoRA Enhance LLMs' Domain-Specific Insight Learning?

Large Language Models (LLMs) have demonstrated remarkable performance on various tasks, yet their ability to extract and internalize deeper insights from domain-specific datasets remains underexplored. In this study, we investigate how continual pre-training can enhance LLMs' capacity for insight learning across three distinct forms: declarative, statistical, and probabilistic insights. Focusing on two critical domains: medicine and finance, we employ LoRA to train LLMs on two existing datasets. To evaluate each insight type, we create benchmarks to measure how well continual pre-training helps models go beyond surface-level knowledge. We also assess the impact of document modification on capturing insights. The results show that, while continual pre-training on original documents has a marginal effect, modifying documents to retain only essential information significantly enhances the insight-learning capabilities of LLMs.

cs.CL↗

arXiv2024

Multi-Conditional Ranking with Large Language Models

Utilizing large language models (LLMs) to rank a set of items has become a common approach in recommendation and retrieval systems. Typically, these systems focus on ordering a substantial number of documents in a monotonic order based on a given query. However, real-world scenarios often present a different challenge: ranking a comparatively smaller set of items, but according to a variety of diverse and occasionally conflicting conditions. In this paper, we define and explore the task of multi-conditional ranking by introducing MCRank, a benchmark tailored for assessing multi-conditional ranking across various item types and conditions. Our analysis of LLMs using MCRank indicates a significant decrease in performance as the number and complexity of items and conditions grow. To overcome this limitation, we propose a novel decomposed reasoning method, consisting of EXtracting and Sorting the conditions, and then Iteratively Ranking the items (EXSIR). Our extensive experiments show that this decomposed reasoning method enhances LLMs' performance significantly, achieving up to a 14.4% improvement over existing LLMs. We also provide a detailed analysis of LLMs performance across various condition categories, and examine the effectiveness of decomposition step. Furthermore, we compare our method with existing approaches such as Chain-of-Thought and existing ranking models, demonstrating the superiority of our approach and complexity of MCR task. We released our dataset and code.

cs.CL↗

arXiv2024

A Blueprint Architecture of Compound AI Systems for Enterprise

Large Language Models (LLMs) have showcased remarkable capabilities surpassing conventional NLP challenges, creating opportunities for use in production use cases. Towards this goal, there is a notable shift to building compound AI systems, wherein LLMs are integrated into an expansive software infrastructure with many components like models, retrievers, databases and tools. In this paper, we introduce a blueprint architecture for compound AI systems to operate in enterprise settings cost-effectively and feasibly. Our proposed architecture aims for seamless integration with existing compute and data infrastructure, with ``stream'' serving as the key orchestration concept to coordinate data and instructions among agents and other components. Task and data planners, respectively, break down, map, and optimize tasks and data to available agents and data sources defined in respective registries, given production constraints such as accuracy and latency.

cs.DB↗

arXiv2024

Reasoning Capacity in Multi-Agent Systems: Limitations, Challenges and Human-Centered Solutions

Remarkable performance of large language models (LLMs) in a variety of tasks brings forth many opportunities as well as challenges of utilizing them in production settings. Towards practical adoption of LLMs, multi-agent systems hold great promise to augment, integrate, and orchestrate LLMs in the larger context of enterprise platforms that use existing proprietary data and models to tackle complex real-world tasks. Despite the tremendous success of these systems, current approaches rely on narrow, single-focus objectives for optimization and evaluation, often overlooking potential constraints in real-world scenarios, including restricted budgets, resources and time. Furthermore, interpreting, analyzing, and debugging these systems requires different components to be evaluated in relation to one another. This demand is currently not feasible with existing methodologies. In this postion paper, we introduce the concept of reasoning capacity as a unifying criterion to enable integration of constraints during optimization and establish connections among different components within the system, which also enable a more holistic and comprehensive approach to evaluation. We present a formal definition of reasoning capacity and illustrate its utility in identifying limitations within each component of the system. We then argue how these limitations can be addressed with a self-reflective process wherein human-feedback is used to alleviate shortcomings in reasoning and enhance overall consistency of the system.

cs.CL↗