Amir Saeidi

Amir Saeidi, Venkatesh Mishra, Souradeep Mukhopadhyay et al.

Large Language Models are being increasingly deployed as the decision-making core of autonomous agents capable of effecting change in external environments. Yet, in conversational benchmarks, which simulate real-world customer-centric issue resolution scenarios, these agents frequently fail due to the cascading effects of incorrect decision-making. These challenges are particularly pronounced for open-source LLMs with smaller parameter sizes, limited context windows, and constrained inference budgets, which contribute to increased error accumulation in agentic settings. To tackle these challenges, we present the Failure-Aware Meta-Agentic (FAMA) framework. FAMA operates in two stages: first, it analyzes failure trajectories from baseline agents to identify the most prevalent errors; second, it employs an orchestration mechanism that activates a minimal subset of specialized agents tailored to address these failures by injecting a targeted context for the tool-use agent before the decision-making step. Experiments across open-source LLMs demonstrate performance gains up to 27% across evaluation modes over standard baselines. These results highlight that targeted curation of context through specialized agents to address common failures is a valuable design principle for building reliable, multi-turn tool-use LLM agents that simulate real-world conversational scenarios.

Md Nayem Uddin, Amir Saeidi, Divij Handa et al.

This paper introduces UnSeenTimeQA, a novel data contamination-free time-sensitive question-answering (TSQA) benchmark. It differs from existing TSQA benchmarks by avoiding web-searchable queries grounded in the real world. We present a series of time-sensitive event scenarios based on synthetically generated facts. It requires large language models (LLMs) to engage in genuine temporal reasoning without depending on the factual knowledge acquired during the pre-training phase. Our data generation framework enables on-demand generation of new samples, mitigating the risk of data leakage. We designed three types of time-sensitive questions to test LLMs' temporal reasoning abilities over sequential and parallel event occurrences. Our evaluation of five LLMs on synthetic fact-based TSQA reveals mixed results: while they perform well on simpler subsets, their overall performance remains inferior as compared to real world fact-based TSQA. Error analysis indicates that LLMs face difficulties in reasoning over long-range event dependencies and parallel events.

Neeraj Varshney, Satyam Raj, Venkatesh Mishra et al.

Large Language Models (LLMs) have achieved remarkable performance across a wide variety of natural language tasks. However, they have been shown to suffer from a critical limitation pertinent to 'hallucination' in their output. Recent research has focused on investigating and addressing this problem for a variety of tasks such as biography generation, question answering, abstractive summarization, and dialogue generation. However, the crucial aspect pertaining to 'negation' has remained considerably underexplored. Negation is important because it adds depth and nuance to the understanding of language and is also crucial for logical reasoning and inference. In this work, we address the above limitation and particularly focus on studying the impact of negation in LLM hallucinations. Specifically, we study four tasks with negation: 'false premise completion', 'constrained fact generation', 'multiple choice question answering', and 'fact generation'. We show that open-source state-of-the-art LLMs such as LLaMA-2-chat, Vicuna, and Orca-2 hallucinate considerably on all these tasks involving negation which underlines a critical shortcoming of these models. Addressing this problem, we further study numerous strategies to mitigate these hallucinations and demonstrate their impact.

Amir Saeidi, Jurriaan Hage, Ravi Khadka et al.

Many of the existing approaches for program comprehension rely on the linguistic information found in source code, such as identifier names and comments. Semantic clustering is one such technique for modularization of the system that relies on the informal semantics of the program, encoded in the vocabulary used in the source code. Treating the source code as a collection of tokens loses the semantic information embedded within the identifiers. We try to overcome this problem by introducing context models for source code identifiers to obtain a semantic kernel, which can be used for both deriving the topics that run through the system as well as their clustering. In the first model, we abstract an identifier to its type representation and build on this notion of context to construct contextual vector representation of the source code. The second notion of context is defined based on the flow of data between identifiers to represent a module as a dependency graph where the nodes correspond to identifiers and the edges represent the data dependencies between pairs of identifiers. We have applied our approach to 10 medium-sized open source Java projects, and show that by introducing contexts for identifiers, the quality of the modularization of the software systems is improved. Both of the context models give results that are superior to the plain vector representation of documents. In some cases, the authoritativeness of decompositions is improved by 67%. Furthermore, a more detailed evaluation of our approach on JEdit, an open source editor, demonstrates that inferred topics through performing topic analysis on the contextual representations are more meaningful compared to the plain representation of the documents. The proposed approach in introducing a context model for source code identifiers paves the way for building tools that support developers in program comprehension tasks such as application and domain concept location, software modularization and topic analysis.

Amir Saeidi, Shivanshu Verma, Md Nayem Uddin et al.

This study evaluates Direct Preference Optimization (DPO) and its variants for aligning Large Language Models (LLMs) with human preferences, testing three configurations: (1) with Supervised Fine Tuning (SFT), (2) without SFT, and (3) without SFT but using an instruction tuned model. We further investigate how training set size influences model performance. Our evaluation spans 13 benchmarks covering dialogue, reasoning, mathematical problem-solving, question answering, truthfulness, MT-Bench, Big Bench, and the Open LLM Leaderboard. We find that: (1) alignment methods often achieve near optimal performance even with smaller subsets of training data; (2) although they offer limited improvements on complex reasoning tasks, they enhance mathematical problem-solving; and (3) using an instruction tuned model improves truthfulness. These insights highlight the conditions under which alignment methods excel, as well as their limitations.

Divij Handa, Zehua Zhang, Amir Saeidi et al.

Recent advancements in Large Language Model (LLM) safety have primarily focused on mitigating attacks crafted in natural language or common ciphers (e.g. Base64), which are likely integrated into newer models' safety training. However, we reveal a paradoxical vulnerability: as LLMs advance in reasoning, they inadvertently become more susceptible to novel jailbreaking attacks. Enhanced reasoning enables LLMs to interpret complex instructions and decode complex user-defined ciphers, creating an exploitable security gap. To study this vulnerability, we introduce Attacks using Custom Encryptions (ACE), a jailbreaking technique that encodes malicious queries with novel ciphers. Extending ACE, we introduce Layered Attacks using Custom Encryptions (LACE), which applies multi-layer ciphers to amplify attack complexity. Furthermore, we develop CipherBench, a benchmark designed to evaluate LLMs' accuracy in decoding encrypted benign text. Our experiments reveal a critical trade-off: LLMs that are more capable of decoding ciphers are more vulnerable to LACE, with success rates on gpt-oss-20b escalating from 60% under ACE to 72% with LACE. These findings highlight a critical insight: as LLMs become more adept at deciphering complex user ciphers--many of which cannot be preemptively included in safety training--they become increasingly exploitable.

Amir Saeidi, Yiran Luo, Agneet Chatterjee et al.

Recent advancements in human preference optimization, originally developed for Large Language Models (LLMs), have shown significant potential in improving text-to-image diffusion models. These methods aim to learn the distribution of preferred samples while distinguishing them from less preferred ones. However, within the existing preference datasets, the original caption often does not clearly favor the preferred image over the alternative, which weakens the supervision signal available during training. To address this issue, we introduce Dual Caption Preference Optimization (DCPO), a data augmentation and optimization framework that reinforces the learning signal by assigning two distinct captions to each preference pair. This encourages the model to better differentiate between preferred and less-preferred outcomes during training. We also construct Pick-Double Caption, a modified version of Pick-a-Pic v2 with separate captions for each image, and propose three different strategies for generating distinct captions: captioning, perturbation, and hybrid methods. Our experiments show that DCPO significantly improves image quality and relevance to prompts, outperforming Stable Diffusion (SD) 2.1, SFT_Chosen, Diffusion-DPO, and MaPO across multiple metrics, including Pickscore, HPSv2.1, GenEval, CLIPscore, and ImageReward, fine-tuned on SD 2.1 as the backbone.

Venkatesh Mishra, Amir Saeidi, Satyam Raj et al.

Recent advances in reasoning and planning capabilities of large language models (LLMs) have enabled their potential as autonomous agents capable of tool use in dynamic environments. However, in multi-turn conversational environments like $τ$-bench, these agents often struggle with consistent reasoning, adherence to domain-specific policies, and extracting correct information over a long horizon of tool-calls and conversation. To capture and mitigate these failures, we conduct a comprehensive manual analysis of the common errors occurring in the conversation trajectories. We then experiment with reformulations of inputs to the tool-calling agent for improvement in agent decision making. Finally, we propose the Input-Reformulation Multi-Agent (IRMA) framework, which automatically reformulates user queries augmented with relevant domain rules and tool suggestions for the tool-calling agent to focus on. The results show that IRMA significantly outperforms ReAct, Function Calling, and Self-Reflection by 16.1%, 12.7%, and 19.1%, respectively, in overall pass^5 scores. These findings highlight the superior reliability and consistency of IRMA compared to other methods in dynamic environments.

Amir Saeidi, Jurriaan Hage, Ravi Khadka et al.

Program comprehension concerns the ability of an individual to make an understanding of an existing software system to extend or transform it. Software systems comprise of data that are noisy and missing, which makes program understanding even more difficult. A software system consists of various views including the module dependency graph, execution logs, evolutionary information and the vocabulary used in the source code, that collectively defines the software system. Each of these views contain unique and complementary information; together which can more accurately describe the data. In this paper, we investigate various techniques for combining different sources of information to improve the performance of a program comprehension task. We employ state-of-the-art techniques from learning to 1) find a suitable similarity function for each view, and 2) compare different multi-view learning techniques to decompose a software system into high-level units and give component-level recommendations for refactoring of the system, as well as cross-view source code search. The experiments conducted on 10 relatively large Java software systems show that by fusing knowledge from different views, we can guarantee a lower bound on the quality of the modularization and even improve upon it. We proceed by integrating different sources of information to give a set of high-level recommendations as to how to refactor the software system. Furthermore, we demonstrate how learning a joint subspace allows for performing cross-modal retrieval across views, yielding results that are more aligned with what the user intends by the query. The multi-view approaches outlined in this paper can be employed for addressing problems in software engineering that can be encoded in terms of a learning problem, such as software bug prediction and feature location.