arXiv subjects

Sumit Gulwani

Publications and source records attributed to Sumit Gulwani.

At least 19 recordsLinked to original sources

arXiv2026

FlashScript: Compiling Recurring Agent Workflows into Reusable Neuro-Symbolic Programs

Many agent tasks recur with new data or small changes, yet agents regenerate their reasoning and actions on each run, producing different results. Distilling a prior run into a skill helps the agent avoid earlier mistakes, but it still re-runs the full process each time, so runs can vary and incur that cost. We compile a successful agent run into a reusable neuro-symbolic program (NSP) with a set of guidelines we call FlashScript. Actions and data transformations become code, instance-specific values become parameters, and decisions that need model judgment become LLM calls. We identify three layers where these calls belong and three techniques that replace most of them with code. The compilation does not remove nondeterminism; it localizes it, so the deterministic program and its neural calls can be reviewed and tested separately. Across 120 WorkBench analytics tasks with Claude Haiku 4.5, the NSPs were correct and made identical tool calls on every run, using about 70 times fewer input tokens and a third of the runtime of a skill-guided agent whose runs were inconsistent. In a composite version of the tasks, the skill-guided agent's accuracy fell to 23%, and its output inconsistency increased, while the NSPs remained correct and identical across all runs. The NSPs successfully generalized to all 108 unseen instances of the same templates. Our findings suggest that NSPs from agent trajectories can be considered cumulative software assets for end-users to test, maintain, and reuse in their recurring workflows.

cs.HC↗

arXiv2026

RubricCraft: Where Solutions Diverge, Criteria Emerge

Evaluating language agents on open-ended, artifact-producing tasks is difficult: model-based graders can be inconsistent, while exhaustive programmatic checks are often impractical. Textual rubrics make evaluation criteria explicit, but writing them requires domain expertise. Rubrics generated from task descriptions alone can miss requirements and failure modes that become apparent in attempted solutions. We introduce RubricCraft, which treats rubric construction as the generation of natural-language specifications guided by candidate solutions. Given a task description and independently sampled solver trajectories, RubricCraft compares how each attempt satisfies or violates the task requirements and uses these differences to construct rubric criteria. It applies the criteria to the collected solutions and revises checks that overlook errors or penalize valid alternatives. The resulting rubrics capture partial correctness without requiring any single solution to serve as a complete reference. On SpreadsheetBench v2, against a prompt-only rubric written and applied by the same model, RubricCraft rubrics pass fewer wrong outputs (29.5% against 42.6%), fail none of the outputs the benchmark's own checker accepts, and rank outputs by correctness more often, while separating models as often. Repair on a single attempt leaves grading unchanged, but repair driven by attempts from three model families acts on different tasks, and an evidence audit judges most of those repairs improvements, many on tasks whose reference solution is itself wrong.

cs.SE↗

arXiv2026

fxBench: Evaluating and Understanding Formula Suggestions in Spreadsheets

Spreadsheets are among the most widely used tools for computations, and formulas are their core building blocks. Besides agentic spreadsheet tools, proactive formula suggestions enable significant productivity gains and are now available in popular spreadsheet software (Microsoft Excel and Google Sheets). Such suggestions pose interesting challenges: the right context must be collected and represented for a machine learning model, which must then be able to write a correct formula, in a way that is fast and cheap enough to trigger on every = (or subsequent keystrokes for auto-completion) without taking users out of their editing flow. Surprisingly, there is no public benchmark for formula suggestions on real spreadsheets. In this paper, we therefore (1) introduce fxBench as a manually curated benchmark of 503 formula suggestion tasks based on Sheetpedia, (2) break down the formula suggestion problem into four capabilities (gather context, understand context, understand intent, write formula) that a formula suggestion system requires, and (3) evaluate existing and new approaches to understand bottlenecks in these capabilities that pose interesting research directions. Specifically, we annotate each benchmark with the cells that describe intent to determine the theoretical performance limit of a given context selection strategy, we evaluate performance of small and large models on different representations of context, including a dense format with high theoretical coverage at low token counts, and we fine-tune smaller language models to show that we can learn to understand such dense formats. Additionally, we perform quantitative and qualitative analysis on failure modes to understand bottlenecks in the formula writing capability; the best configuration only achieved 58.8% correct suggestions.

cs.LG↗

arXiv2026

Whose turn is it? Context-guided model routing for stateful agents

Model routing can reduce the cost of LLM applications by assigning simpler requests to cheaper models without sacrificing the quality of the response. However, most routers treat requests as independent. Stateful agents do not operate under this assumption: a user turn may require multiple model and tool calls, its difficulty depends on the state produced by earlier turns, and changing models can sacrifice reuse of the conversational prompt cache. We study zero-shot, cost-aware routing at the user-turn boundary, systematically varying the routing rule, access to session history, candidate models that differ in size, reasoning effort, or model provider, and prompt-cache behavior. We evaluate on multi-turn versions of SheetCopilot and SpreadsheetBench 2, where persistent workbooks and cross-turn dependencies make isolated routing decisions especially brittle. At a matched offloading rate, giving the router session history improves task pass rate over the same router without history, with gains for every model pair; history-aware routing remains within one point of always using the expensive model. We also found that cost depends on execution dynamics rather than token prices alone: the cheaper model's per-task cost advantage narrows from 6.8x to 3.2x on longer sessions. A closed form cache analysis further shows that downward switches (expensive to cheap) recover their cache cost within two LLM calls, whereas upward switches do not break even under the evaluated conditions. Together, these results show that agent routing is a sequential decision problem: effective policies must account for what happened earlier in the session and for the asymmetric cost of switching models.

cs.AI↗

arXiv2026

Dude, Where's My State? Execution Information Requirements for Stateful Agents

Long-running agents must preserve information that later steps depend on. We introduce the Execution Information Requirement (EIR), a lower bound on the information that must remain accessible for correct completion under specified task and access conditions. We develop LACUNA, a framework that generates tasks with known dependencies and varies information demand, retention, and recovery separately from the difficulty of individual operations. Across four models, restoring a missing result raises accuracy on affected recall steps to 100%, compared with 0% for equal-length irrelevant information. Sufficient storage alone does not ensure success: retention policies can discard required results, errors can propagate through later computations, and agents can stop before recovery is complete. We also introduce VESTIGE, which uses agent execution traces to construct semantic graphs and measure information demand for real tasks. Across 72,562 software-agent trajectories, VESTIGE reveals a steeper distance-related decline in solution-relevant rereading for failed runs (RR 0.951 per distance doubling), while adjusted peak demand alone is not associated with failure. Together, these contributions support evaluating whether agents preserve and recover the information their tasks require.

cs.AI↗

arXiv2026

On Device Agentic Operation Caches -- Classifier-Centric NL-to-Action Generation

Agentic AI is increasingly being embedded in software applications to provide natural language interfaces to features and functionality. In most cases these agents are powered by enterprise (100+ billion parameter) or frontier class large language models that require substantial computational resources run and depend on cloud hosted inference to handle the task of transforming natural language inputs into actionable software operations. This reliance on cloud-hosted inference introduces substantial network latency on top of LLM inference times, creates data privacy concerns, and, given the costs of running these models, can rapidly escalate expenses associated with supporting agentic features. This paper introduces a novel means of converting the NL-to-Action problem from a generative one into a classification-centric formulation via on-device operation caches. These caches allow an agentic system to handle frequently occurring classes of actions completely on-device -- reducing latency, enhancing privacy, and lowering operational costs. We show that for a classic NL-to-Formula task, generating Excel Formula in response to user requests, this approach reduces total inference cost by 56% when compared to cloud-only model-routing based inference and, on cache hits, reduces the latency to response latency by 5x.

cs.AI↗

arXiv2026

Coding Agents are Strong Prompt Optimizers

Search-based prompt optimizers improve prompts through iterative search: they propose edits, execute fresh rollouts, score the resulting trajectories, and retain only edits that improve a validation metric. We show that this optimization loop is unnecessary. Given only a static corpus of agent trajectories, an off-the-shelf coding agent can directly synthesize an optimized prompt, requiring neither environment access nor validation data. We call this approach \textit{Coding-Agent Skill Distillation} (CASD). The key insight is reflection scope. Rather than reasoning over a small batch of trajectories at each optimization step, the coding agent writes and executes analysis code to compute corpus-wide statistics, identifies systematic failure modes, inspects representative episodes, and distills the resulting insights into behavioral rules. Across four agentic benchmarks (ALFWorld, $τ^2$-bench retail and telecom, and SpreadsheetBench-Verified), under matched data access, a single CASD pass outperforms GEPA, a state-of-the-art reflective prompt optimizer, on three of four benchmarks and outperforms validation-gated reflective search (SkillOpt) on all four, improving the unoptimized baseline by 16.6 percentage points on average versus 10.9 for GEPA and 5.3 for SkillOpt. Because CASD performs a single offline analysis pass rather than iterative search, producing an optimized prompt costs approximately \$1.60---over $22\times$ cheaper than validation-gated search. Even when competing methods are granted additional validation data and unrestricted environment access, CASD remains ahead on two of four benchmarks. These results suggest that corpus-scale statistical reflection is a viable alternative to iterative search for prompt optimization.

cs.AI↗

arXiv2026

Back to the Future: A workbook time machine for spread sheet creation benchmarks

We introduce the workbook time machine, a pipeline that automatically creates benchmarks evaluating the ability of language models to create derived objects in spreadsheets (formulas, charts, pivot tables, and conditional formatting). Applied to public workbook corpora, it produces wtmcorpus--a collection of (input workbook, output workbook, query) triples spanning four artifact types and varying complexity. From this corpus we curate wtmbench, a 150-task evaluation benchmark with queries at three levels of specificity. We evaluate existing spreadsheet manipulation agents and baselines on wtmbench across artifact types, step complexity, and instruction granularity. Our evaluations show that query specificity, agent orchestration, and interface API used to control spreadsheets play a big role in LLM performance on Excel tasks.

cs.AI↗

arXiv2026

Reason Wide, Not Deep: Amortizing the Reasoning Premium into Distilled Skills

Reasoning modes of language models outperform their non-reasoning counterparts on multi-step agentic tasks, but pay a 3-6x premium in output tokens on every episode -- much of it spent re-deriving procedures that are shared across episodes of the same domain. We show this recurring cost can be amortized: a coding agent analyses a small corpus of existing trajectories from a training split and compiles a compact natural-language skill that is injected into the non-reasoning model's system prompt. Across four agentic benchmarks (ALFWorld, tau$^2$-bench telecom and retail, and SpreadsheetBench-Verified), skills recover 55%-100%+ of the reasoning gap for GPT-5.4-mini on held-out tasks -- exceeding the reasoning mode outright on two of four -- while emitting 2.7-6x fewer output tokens and zero reasoning tokens. Notably, reasoning traces are not a prerequisite: skills distilled from non-reasoning trajectories alone remain competitive with skills distilled from paired reasoning/non-reasoning corpora, with domain-dependent differences between the two sources. We interpret these results through a search lens: test-time reasoning is deep search inside a single episode, re-paid at every deployment, while corpus distillation is wide search across episodes, paid once. The two recover overlapping procedural knowledge, and width over cheap trajectories is often the better buy -- with the residual gap on some domains (telecom, SpreadsheetBench) delineating where genuinely per-instance deep search remains necessary.

cs.AI↗

arXiv2026

Plans Work in Mysterious Ways: Evaluating a Plan Mode for Spreadsheet Agents

Plan Modes have become standard features in agentic programming tools, allowing users to gain transparency and control by working with the agent to develop a plan before task execution. However, it remains unclear whether the benefits of this feature translate to end-user programming environments such as spreadsheets. Since spreadsheet programmers tend to work iteratively and care less about technical correctness, upfront planning may not fit into their workflows as easily. In this paper, we build a prototype of a Plan Mode for spreadsheet programming and evaluate it against a non-planning baseline through a within-subjects user study (N=24). We found that despite similar task outcomes with both tools, using Plan Mode led to a reduction in refinement and a better perception of the tool across dimensions of creativity support and human-machine collaboration. We discuss the implications of these results for the future design of Plan Modes, and for the broader role of human-AI planning in end-user programming.

cs.HC↗

arXiv2026

Improving Language Agents through BREW: Bootstrapping expeRientially-learned Environmental knoWledge

Large Language Model (LLM)-based agents are increasingly capable of complex, multi-step tasks such as GUI automation, tool use, and data manipulation, yet they cannot learn from experience: each new session rediscovers solutions from scratch. We introduce BREW (Bootstrapping expeRientially-learned Environmental knoWledge), a framework that distills an agent's past interaction trajectories into a structured, retrievable knowledge base (KB) of natural-language recipes, concept-level procedural documents that capture what to do, when it applies, and what to watch out for. Drawing on the principle of library learning from program synthesis, BREW decomposes agent memory into modular, concept-localized documents and formalizes KB construction as a state-space search problem. To navigate this space, we introduce Expand-and-Gather Monte Carlo Tree Search (EG-MCTS), a reward-guided algorithm that jointly optimizes recipe accuracy and retrievability across parallel, per-concept search trees. We further adapt hindsight relabeling to convert near-miss trajectories into positive demonstrations, surfacing latent agent competencies as reusable knowledge. On three domain-grounded benchmarks, OSWorld, tau^2-Bench, and SpreadSheetBench, BREW achieves 10-20% gains in task success and 10-15% fewer execution steps over base agents, while consistently outperforming existing memory-augmented baselines that can degrade below memoryless performance. The resulting KB is inspectable, modular, and extensible, providing a transparent and controllable substrate for agent optimization.

cs.AI↗

arXiv2026

A Benchmark and Framework for Evaluating Next Action Predictions in Spreadsheets

Predictive code completion greatly accelerates how quickly developers work. In spreadsheets, despite being much more common, such auto-completion features are virtually non-existent. To address this gap, we introduce a benchmark for systems that observe a sequence of user actions in a spreadsheet and predict future actions. Two challenges are (1) the absence of edit histories in public spreadsheet corpora and (2) the complex space of spreadsheet actions (spatial, temporal, composite). To address (1), we manually curate 52 sequences of 12K actions that recreate spreadsheets from public corpora, seeded by parametrized heuristics and LLM refinement. To address (2), we propose an online evaluation that expects a prediction after each user action, accepts or rejects that prediction, updates the future actions upon acceptance, and repeats this until the target spreadsheet is obtained. We use multiple baseline predictors (including zero-shot LLMs, fine-tuned SLMs, and classical models) and analyze different properties that our benchmark teaches us, including but not limited to: properties of saved actions and false positives, efficiency, effect of user profiles, effect of triggers, and effect of context.

cs.SE↗

arXiv2026

SkillAxe: Sharpening LLM-Authored Agent Skills Through Evaluation-Guided Self-Refinement

Skill documents, structured natural-language instructions that guide Large Language Model (LLM) agents, are critical to modern agent frameworks, yet LLMs struggle to write skills that actually work. On SkillsBench, human-authored skills improve pass rates by 16.2 percentage points, while LLM-authored skills provide no measurable gain. We introduce SkillAxe, a fully unsupervised framework that enables LLMs to iteratively diagnose and refine their own skills. SkillAxe decomposes skill quality into four interpretable dimensions (quality impact, trigger precision, instruction compliance with fault attribution, and solution-path coverage), producing structured improvement briefs that require no ground-truth labels, test suites, or environment rewards. On SkillsBench, SkillAxe improves pass rates by 28\% relative over unimproved LLM skills and closes 47--67\% of the gap to human-authored skills. We validate the approach as a continuous improvement engine in the wild on SpreadsheetBench, where a SkillAxe-built skill library learns from past agent trajectories and raises pass rate from 16.0\% to 52.0\% using only 22 skills.

cs.MA↗

arXiv2026

Auditing and Controlling AI Agent Actions in Spreadsheets

Advances in AI agent capabilities have outpaced users' ability to meaningfully oversee their execution. AI agents can perform sophisticated, multi-step knowledge work autonomously from start to finish, yet this process remains effectively inaccessible during execution, often buried within large volumes of intermediate reasoning and outputs: by the time users receive the output, all underlying decisions have already been made without their involvement. This lack of transparency leaves users unable to examine the agent's assumptions, identify errors before they propagate, or redirect execution when it deviates from their intent. The stakes are particularly high in spreadsheet environments, where process and artifact are inseparable. Each decision the agent makes is recorded directly in cells that belong to and reflect on the user. We introduce Pista, a spreadsheet AI agent that decomposes execution into auditable, controllable actions, providing users with visibility into the agent's decision-making process and the capacity to intervene at each step. A formative study (N = 8) and a within-subjects summative evaluation (N = 16) comparing Pista to a baseline agent demonstrated that active participation in execution influenced not only task outcomes but also users' comprehension of the task, their perception of the agent, and their sense of role within the workflow. Users identified their own intent reflected in the agent's actions, detected errors that post-hoc review would have failed to surface, and reported a sense of co-ownership over the resulting output. These findings indicate that meaningful human oversight of AI agents in knowledge work requires not improved post-hoc review mechanisms, but active participation in decisions as they are made.

cs.HC↗

arXiv2026

Exploration vs. Fixation: Scaffolding Divergent and Convergent Thinking for Human-AI Co-Creation with Generative Models

Generative AI has democratized content creation, but popular chatbot-based interfaces often prioritize execution, generating fully rendered artifacts right away. This issue can lead to premature convergence and design fixation, where users are being anchored to initial outputs. Recent works have proposed new interfaces to address this issue by supporting exploration, though typically constrained to be semantically close to a user's initial task framing, potentially limiting the creativity of the outcomes. We examine an approach grounded in the Geneplore model of creative cognition and instantiate it in a human-AI co-creation system, HAICo, for creative image generation. HAICo explicitly structures the creative process into two switchable modes: DIVERGENT mode scaffolds the broad exploration of remote conceptual ideas; CONVERGENT mode supports a targeted refinement of selected ideas. Through a within-subjects study (N=24) on a poster image creation task, we demonstrate that HAICo outperforms ChatGPT across multiple dimensions of creativity and usability. Our results highlight the critical need to shift from pure execution-focused chatbots to scaffolded co-creation systems that actively guide exploration and foster the creative process.

cs.HC↗

arXiv2026

IndiMathBench: Autoformalizing Mathematical Reasoning Problems with a Human Touch

Reliable autoformalization remains challenging even in the era of large language models (LLMs). The scarcity of high-quality training data is a major bottleneck. Expert annotation requires substantial time and deep expertise in both mathematics and theorem proving. We introduce IndiMathBench, a human-verified benchmark designed to evaluate mathematical theorem proving, curated using an AI-powered human-assisted pipeline for formalizing natural language problems in Lean. IndiMathBench is composed of 312 formal Lean 4 theorems paired with their corresponding informal problem statements, sourced from Indian Mathematics Olympiads. Through category-based retrieval, iterative compiler feedback, and multi-model ensembles, our pipeline generates candidate formalizations that experts efficiently validate via an interactive dashboard with automated quality summaries. Evaluation across multiple frontier models demonstrates that autoformalization remains challenging, with substantial gaps between syntactic validity and semantic correctness, while theorem proving success rates remain low even with iterative refinement, demonstrating that \benchmark~presents a challenging testbed for mathematical reasoning. IndiMathBench is available at https://github.com/prmbiy/IndiMathBench.

cs.AI↗

arXiv2026

Closing the Loop: An Instructor-in-the-Loop AI Assistance System for Supporting Student Help-Seeking in Programming Education

Timely and high-quality feedback is essential for effective learning in programming courses; yet, providing such support at scale remains a challenge. While AI-based systems offer scalable and immediate help, their responses can occasionally be inaccurate or insufficient. Human instructors, in contrast, may bring more valuable expertise but are limited in time and availability. To address these limitations, we present a hybrid help framework that integrates AI-generated hints with an escalation mechanism, allowing students to request feedback from instructors when AI support falls short. This design leverages the strengths of AI for scale and responsiveness while reserving instructor effort for moments of greatest need. We deployed this tool in a data science programming course with 82 students. We observe that out of the total 673 AI-generated hints, students rated 146 (22%) as unhelpful. Among those, only 16 (11%) of the cases were escalated to the instructors. A qualitative investigation of instructor responses showed that those feedback instances were incorrect or insufficient roughly half of the time. This finding suggests that when AI support fails, even instructors with expertise may need to pay greater attention to avoid making mistakes. We will publicly release the tool for broader adoption and enable further studies in other classrooms. Our work contributes a practical approach to scaling high-quality support and informs future efforts to effectively integrate AI and humans in education.

cs.CY↗

arXiv2026

An Empirical Investigation of Robustness in Large Language Models under Tabular Distortions

We investigate how large language models (LLMs) fail when tabular data in an otherwise canonical representation is subjected to semantic and structural distortions. Our findings reveal that LLMs lack an inherent ability to detect and correct subtle distortions in table representations. Only when provided with an explicit prior, via a system prompt, do models partially adjust their reasoning strategies and correct some distortions, though not consistently or completely. To study this phenomenon, we introduce a small, expert-curated dataset that explicitly evaluates LLMs on table question answering (TQA) tasks requiring an additional error-correction step prior to analysis. Our results reveal systematic differences in how LLMs ingest and interpret tabular information under distortion, with even SoTA models such as GPT-5.2 model exhibiting a drop of minimum 22% accuracy under distortion. These findings raise important questions for future research, particularly regarding when and how models should autonomously decide to realign tabular inputs, analogous to human behavior, without relying on explicit prompts or tabular data pre-processing.

cs.AI↗