Waseem AlShikh

Melisa Russak, Umar Jamil, Christopher Bryant et al.

In this paper, we introduce Writing in the Margins (WiM), a new inference pattern for Large Language Models designed to optimize the handling of long input sequences in retrieval-oriented tasks. This approach leverages the chunked prefill of the key-value cache to perform segment-wise inference, which enables efficient processing of extensive contexts along with the generation and classification of intermediate information ("margins") that guide the model towards specific tasks. This method increases computational overhead marginally while significantly enhancing the performance of off-the-shelf models without the need for fine-tuning. Specifically, we observe that WiM provides an average enhancement of 7.5% in accuracy for reasoning skills (HotpotQA, MultiHop-RAG) and more than a 30.0% increase in the F1-score for aggregation tasks (CWE). Additionally, we show how the proposed pattern fits into an interactive retrieval design that provides end-users with ongoing updates about the progress of context processing, and pinpoints the integration of relevant information into the final response. We release our implementation of WiM using Hugging Face Transformers library at https://github.com/writer/writing-in-the-margins.

Waseem AlShikh, Manhal Daaboul, Kirk Goddard et al.

In this paper, we introduce the Instruction Following Score (IFS), a metric that detects language models' ability to follow instructions. The metric has a dual purpose. First, IFS can be used to distinguish between base and instruct models. We benchmark publicly available base and instruct models, and show that the ratio of well formatted responses to partial and full sentences can be an effective measure between those two model classes. Secondly, the metric can be used as an early stopping criteria for instruct tuning. We compute IFS for Supervised Fine-Tuning (SFT) of 7B and 13B LLaMA models, showing that models learn to follow instructions relatively early in the training process, and the further finetuning can result in changes in the underlying base model semantics. As an example of semantics change we show the objectivity of model predictions, as defined by an auxiliary metric ObjecQA. We show that in this particular case, semantic changes are the steepest when the IFS tends to plateau. We hope that decomposing instruct tuning into IFS and semantic factors starts a new trend in better controllable instruct tuning and opens possibilities for designing minimal instruct interfaces querying foundation models.

Kiran Kamble, Melisa Russak, Dmytro Mozolevskyi et al.

We propose a new long-context financial benchmark, FailSafeQA, designed to test the robustness and context-awareness of LLMs against six variations in human-interface interactions in LLM-based query-answer systems within finance. We concentrate on two case studies: Query Failure and Context Failure. In the Query Failure scenario, we perturb the original query to vary in domain expertise, completeness, and linguistic accuracy. In the Context Failure case, we simulate the uploads of degraded, irrelevant, and empty documents. We employ the LLM-as-a-Judge methodology with Qwen2.5-72B-Instruct and use fine-grained rating criteria to define and calculate Robustness, Context Grounding, and Compliance scores for 24 off-the-shelf models. The results suggest that although some models excel at mitigating input perturbations, they must balance robust answering with the ability to refrain from hallucinating. Notably, Palmyra-Fin-128k-Instruct, recognized as the most compliant model, maintained strong baseline performance but encountered challenges in sustaining robust predictions in 17% of test cases. On the other hand, the most robust model, OpenAI o3-mini, fabricated information in 41% of tested cases. The results demonstrate that even high-performing models have significant room for improvement and highlight the role of FailSafeQA as a tool for developing LLMs optimized for dependability in financial applications. The dataset is available at: https://huggingface.co/datasets/Writer/FailSafeQA

Rakshith Vasudev, Melisa Russak, Dan Bikel et al.

Proactive interventions by LLM critic models are often assumed to improve reliability, yet their effects at deployment time are poorly understood. We show that a binary LLM critic with strong offline accuracy (AUROC 0.94) can nevertheless cause severe performance degradation, inducing a 26 percentage point (pp) collapse on one model while affecting another by near zero pp. This variability demonstrates that LLM critic accuracy alone is insufficient to determine whether intervention is safe. We identify a disruption-recovery tradeoff: interventions may recover failing trajectories but also disrupt trajectories that would have succeeded. Based on this insight, we propose a pre-deployment test that uses a small pilot of 50 tasks to estimate whether intervention is likely to help or harm, without requiring full deployment. Across benchmarks, the test correctly anticipates outcomes: intervention degrades performance on high-success tasks (0 to -26 pp), while yielding a modest improvement on the high-failure ALFWorld benchmark (+2.8 pp, p=0.014). The primary value of our framework is therefore identifying when not to intervene, preventing severe regressions before deployment.

Waseem AlShikh, Muayad Sayed Ali, Brian Kennedy et al.

As AI agents proliferate across industries and applications, evaluating their performance based solely on infrastructural metrics such as latency, time-to-first-token, or token throughput is proving insufficient. These metrics fail to capture the quality of an agent's decisions, its operational autonomy, or its ultimate business value. This white paper proposes a novel, comprehensive framework of eleven outcome-based, task-agnostic performance metrics for AI agents that transcend domain boundaries. These metrics are designed to enable organizations to evaluate agents based on the quality of their decisions, their degree of autonomy, their adaptability to new challenges, and the tangible business value they deliver, regardless of the underlying model architecture or specific use case. We introduce metrics such as Goal Completion Rate (GCR), Autonomy Index (AIx), Multi-Step Task Resilience (MTR), and Business Impact Efficiency (BIE). Through a large-scale simulated experiment involving four distinct agent architectures (ReAct, Chain-of-Thought, Tool-Augmented, Hybrid) across five diverse domains (Healthcare, Finance, Marketing, Legal, and Customer Service), we demonstrate the framework's efficacy. Our results reveal significant performance trade-offs between different agent designs, highlighting the Hybrid Agent as the most consistently high-performing model across the majority of our proposed metrics, achieving an average Goal Completion Rate of 88.8\% and the highest Return on Investment (ROI). This work provides a robust, standardized methodology for the holistic evaluation of AI agents, paving the way for more effective development, deployment, and governance.

Raghav Kapoor, Yash Parag Butala, Melisa Russak et al. · cmu

For decades, human-computer interaction has fundamentally been manual. Even today, almost all productive work done on the computer necessitates human input at every step. Autonomous virtual agents represent an exciting step in automating many of these menial tasks. Virtual agents would empower users with limited technical proficiency to harness the full possibilities of computer systems. They could also enable the efficient streamlining of numerous computer tasks, ranging from calendar management to complex travel bookings, with minimal human intervention. In this paper, we introduce OmniACT, the first-of-a-kind dataset and benchmark for assessing an agent's capability to generate executable programs to accomplish computer tasks. Our scope extends beyond traditional web automation, covering a diverse range of desktop applications. The dataset consists of fundamental tasks such as "Play the next song", as well as longer horizon tasks such as "Send an email to John Doe mentioning the time and place to meet". Specifically, given a pair of screen image and a visually-grounded natural language task, the goal is to generate a script capable of fully executing the task. We run several strong baseline language model agents on our benchmark. The strongest baseline, GPT-4, performs the best on our benchmark However, its performance level still reaches only 15% of the human proficiency in generating executable scripts capable of completing the task, demonstrating the challenge of our task for conventional web agents. Our benchmark provides a platform to measure and evaluate the progress of language model agents in automating computer tasks and motivates future work towards building multimodal models that bridge large language models and the visual grounding of computer screens.

Zhenyu Zhao, Sander Land, Dan Bikel et al.

Reasoning in Large Language Models incurs significant inference-time compute, yet the token-level information structure of reasoning traces remains underexplored. We observe that reasoning tokens split into two functional types: low-entropy \textit{structural} tokens (recurring phrases that scaffold the reasoning process) and higher-entropy \textit{organic} tokens (problem-specific content that drives toward a solution). This asymmetry motivates a simple, model-agnostic compression pipeline: apply cross-word BPE merges on a model's own reasoning traces to derive \textit{supertokens} that capture frequent structural patterns, then teach the model to adopt them via supervised fine-tuning. Across three model families and five mathematical reasoning benchmarks, our approach shortens reasoning traces by 8.1\% on average with no statistically significant accuracy loss on any model--benchmark pair. Beyond compression, supertokens act as interpretable reasoning-move annotations (backtracking, verification, strategy shifts), exposing the model's high-level strategy at a glance. Analyzing transitions between structural categories reveals systematic differences between correct and incorrect traces: correct traces show productive recovery (backtracking followed by strategy shifts and verification), while incorrect traces are dominated by confusion cycles (repeated hedging and unresolved contradictions). These diagnostic signals suggest applications in reward shaping and early stopping for RL-based reasoning training.

Zhenyu Zhao, Aparna Balagopalan, Adi Agrawal et al.

Given the increased use of LLMs in financial systems today, it becomes important to evaluate the safety and robustness of such systems. One failure mode that LLMs frequently display in general domain settings is that of sycophancy. That is, models prioritize agreement with expressed user beliefs over correctness, leading to decreased accuracy and trust. In this work, we focus on evaluating sycophancy that LLMs display in agentic financial tasks. Our findings are three-fold: first, we find the models show only low to modest drops in performance in the face of user rebuttals or contradictions to the reference answer, which distinguishes sycophancy that models display in financial agentic settings from findings in prior work. Second, we introduce a suite of tasks to test for sycophancy by user preference information that contradicts the reference answer and find that most models fail in the presence of such inputs. Lastly, we benchmark different modes of recovery such as input filtering with a pretrained LLM.

Shelly Bensal, Umar Jamil, Christopher Bryant et al.

We explore a method for improving the performance of large language models through self-reflection and reinforcement learning. By incentivizing the model to generate better self-reflections when it answers incorrectly, we demonstrate that a model's ability to solve complex, verifiable tasks can be enhanced even when generating synthetic data is infeasible and only binary feedback is available. Our framework operates in two stages: first, upon failing a given task, the model generates a self-reflective commentary analyzing its previous attempt; second, the model is given another attempt at the task with the self-reflection in context. If the subsequent attempt succeeds, the tokens generated during the self-reflection phase are rewarded. Our experimental results show substantial performance gains across a variety of model architectures, as high as 34.7% improvement at math equation writing and 18.1% improvement at function calling. Notably, smaller fine-tuned models (1.5 billion to 7 billion parameters) outperform models in the same family that are 10 times larger. Our novel paradigm is thus an exciting pathway to more useful and reliable language models that can self-improve on challenging tasks with limited external feedback.

Dmytro Mozolevskyi, Waseem AlShikh

This research paper presents a comprehensive analysis of integrating advanced language models with search and retrieval systems in the fields of information retrieval and natural language processing. The objective is to evaluate and compare various state-of-the-art methods based on their performance in terms of accuracy and efficiency. The analysis explores different combinations of technologies, including Azure Cognitive Search Retriever with GPT-4, Pinecone's Canopy framework, Langchain with Pinecone and different language models (OpenAI, Cohere), LlamaIndex with Weaviate Vector Store's hybrid search, Google's RAG implementation on Cloud VertexAI-Search, Amazon SageMaker's RAG, and a novel approach called KG-FID Retrieval. The motivation for this analysis arises from the increasing demand for robust and responsive question-answering systems in various domains. The RobustQA metric is used to evaluate the performance of these systems under diverse paraphrasing of questions. The report aims to provide insights into the strengths and weaknesses of each method, facilitating informed decisions in the deployment and development of AI-driven search and retrieval systems.