arXiv subjects

Sushant Mehta

Publications and source records attributed to Sushant Mehta.

At least 19 recordsLinked to original sources

arXiv2026

Chartography: A Benchmark for Professional Chart Understanding

Professionals across medicine, engineering, finance, manufacturing, and the sciences often make consequential decisions from charts. Existing chart benchmarks do not sufficiently measure this ability: they are dominated by bar, line, and pie formats, rely on shorter reasoning chains, and are nearing saturation, with frontier models already scoring 80-90%. We introduce Chartography, a benchmark of 100 tasks that pair charts drawn from professional practice, in domain-specific formats that standard chart benchmarks rarely include, with questions written by professionals who read these charts for a living and independently verified by three additional experts. In an evaluation of 30 frontier-model configurations (20 scored trials per task), the best configuration reaches only 45.0% mean pass@1; the remainder span 9.0-39.5%. Failures concentrate in visual perception: models can miss nuanced features, misread values along sparsely labeled axes, mishandle projected 3D geometry, and violate domain conventions encoded in the chart. We release all tasks, images, provenance metadata, and evaluation code.

cs.CV↗

arXiv2026

Post-Training on Office Work Improves Software Engineering: A Behavioral Account of Cross-Domain Transfer

Long-horizon tasks require agents to maintain coherent state and goals across nested and branching work. We call this capability goal-directed execution (GDE): the repeated application of four behaviors, namely selecting goals, constructing task-relevant state, maintaining fidelity to higher-level objectives, and verifying completion against the environment. We hypothesize that long-horizon post-training strengthens these behaviors across domains. We test this by post-training Qwen3.5-122B-A10B on 363 Long-Horizon Multi-Tool Agent (LHMTA) tasks drawn from office workflows. The collection contained no software-engineering tasks, yet the model's pass@1 improved by 5.8 points on SWE-Bench Pro. Matched trajectory analysis shows gains in all four GDE behaviors in both office workflows and software repositories. Aggregate SWE-Bench Pro statistics showed related changes in information gathering, implementation, and verification. Together, the results support a behavioral interpretation in which long-horizon post-training changed how the model organized and applied knowledge across tasks, with effects extending beyond the training domain.

cs.AI↗

arXiv2026

Cross-Benchmark Generalization in Long-Horizon Agents

For reinforcement learning (RL) in self-contained environments, a policy can get rewards by exploiting environment-specific regularities (tool schemas, grader parsing, task templates) rather than by acquiring transferable skill, and an in-distribution holdout shares those regularities. We argue that the discriminating question is behavioral, namely how a trained agent acts, and that cross-benchmark transfer is the right place to look for it. We post-train an open-weight mixture-of-experts model (Qwen3.5-122B-A10B) on 363 long-horizon Model Context Protocol (MCP) tasks across 27 categories, using a two-stage SFT-then-RL pipeline. Toolathlon performance informed the initial base-family and SFT-teacher choices, but no external-benchmark task or grader entered training and no external score informed the reward, training hyperparameters, trained-checkpoint selection, or stopping. At greedy pass@1, the trained model improves over the base on five reported external evaluations: Toolathlon (+9.6 pp), $\tau^2$-Bench (+5.3 pp), BFCL-V4 (+3.5 pp), SWE-Bench Pro (+5.8 pp), and Terminal-Bench 2 (+2.8 pp). Both software-engineering benchmarks improve despite the training collection containing no software-engineering tasks. An exploratory paired-trajectory analysis identifies four recurring behavioral differences (more careful local-goal formation, building goal-relevant working state, keeping parent goals stable through local repairs, and verifying completion) that appear in analogous forms across office workflows and code. These results provide descriptive evidence that long-horizon multi-tool post-training can change ways of working that transfer beyond its training domain.

cs.SE↗

arXiv2026

HANDBOOK.md: A Benchmark for Long-Context Agentic Instruction Following

Language-model agents are increasingly deployed under standing instructions: a system prompt, a policy file, or a skills document is placed in context, and the agent is trusted to let that document govern every action that follows. Existing benchmarks rarely test this deployment pattern directly; they measure whether an agent can complete a task, not whether a long, binding policy document constrains its behavior over an extended tool-use horizon. We present HANDBOOK_md, a benchmark of 65 agentic tasks modeled on how employees follow company handbooks. Each task places an agent in a self-contained company environment (a file workspace with mock email, chat, calendar, issue-tracking, and commerce services exposed over the Model Context Protocol) and instructs it to carry out routine professional work governed by an expert-written standard operating procedure of 20-124 pages. Tasks span five domains (finance, medical billing, insurance, logistics, and HR) and 10 fictional companies. To resist memorization, every task modifies one of 10 base handbooks, altering the specific rules and thresholds on which grading depends, so no two tasks share the same set of policies. Grading is fully deterministic: each task carries a rubric of programmatic criteria (824 in total) that check both that required actions occurred and that prohibited actions did not. Under strict grading, where a trial passes only if every criterion is satisfied, the strongest evaluated model passes 36.2% of trials, and most frontier models remain below 25%. Failures follow consistent patterns: agents let a plausible but unauthorized in-environment request override the standing policy, perform a required check and then act against its result, lose rule details over long horizons, and report compliance they did not achieve. We release the tasks, environments, and evaluation harness.

cs.AI↗

arXiv2026

GDP.pdf: Benchmarking Grounded Multimodal Reasoning over Professional PDF Documents

A large share of day-to-day work in professional domains happens inside PDF files: benefits packets, leases, datasheets, clinical guidelines, construction plans. Benchmarks for document AI have generally measured the required capabilities in isolation: OCR, layout analysis, chart reasoning, table QA, document VQA. A high score on any one of them does not necessarily reveal whether a model can answer a realistic question that someone in the field would actually ask about a specific PDF. GDP_pdf is a benchmark built to measure this directly. It consists of question-document pairs authored by working professionals in ten fields, and a candidate question was kept only when at least two frontier multimodal models failed it in a way that mattered: a wrong answer, missed decisive evidence, or a fabricated claim, rather than a superficial difference such as style. Each item comes with a rubric of atomic criteria, so we can report a graded rubric score as well as a strict task-level pass rate, and each item is tagged against a taxonomy of eleven capabilities in three tiers, spanning text extraction and grounding, table and chart comprehension, cross-referencing, spatial reasoning, and abstention on unsupported queries. We report results for seventeen frontier models on the 100-item benchmark: the best model passes only 30.7% of the items and the worst passes 2%. Most errors trace back to a small set of recurring loss patterns: misaligned tables, misread charts, skipped footnotes and exclusions, miscounted floor-plan symbols, scan noise, and amendments that supersede earlier text.

cs.CV↗

arXiv2026

ComplexConstraints and Beyond: Expert Rubrics for RLVR

Evaluation protocols can lag behind LLM capabilities. Programmatically verified benchmarks cover narrow surface constraints, whereas real-world instruction following and agentic workflows require judging semantic, contextual, and policy-dependent behavior. We study expert-curated rubric-based evaluation as a unified mechanism for measurement and reinforcement-learning rewards across two settings: complex instruction following and enterprise agentic tasks. We identify rubric-design choices that affect reward quality, including maximum viable atomicity, intent-aware criterion design, and LLM-judge calibration. We introduce ComplexConstraints, an expert-curated instruction-following suite comprising a public 75-prompt benchmark with 1,559 rubric criteria and a disjoint 1,000-prompt training set, with 10-40 atomic criteria per prompt. Empirically, rubric rewards improve training in both fixed task datasets, such as ComplexConstraints, and stateful RL environments, such as CoreCraft. Training a 4B model on ComplexConstraints improves mean criterion pass rate by +15.5 pp on a held-out split, bringing it within 0.5 pp of the untrained baseline of a roughly 60x larger Qwen3 model, and the gains transfer to external benchmarks the model never saw during training: +8.4 pp on AdvancedIF and +10.1 pp on MultiChallenge. In CoreCraft, rubric-reward RL likewise transfers to out-of-distribution benchmarks (+4.5 pp BFCL, +7.4 pp tau^2-Bench, +6.8 pp Toolathlon). These results show that expert-authored rubrics provide effective evaluation targets and scalable reward signals for improving LLM instruction following and agentic behavior.

cs.AI↗

arXiv2026

Riemann-Bench: A Benchmark for Moonshot Mathematics

Recent AI systems have achieved gold-medal-level performance on the International Mathematical Olympiad, demonstrating remarkable proficiency at competition-style problem solving. However, competition mathematics represents only a narrow slice of mathematical reasoning: problems are drawn from limited domains, require minimal advanced machinery, and can often reward insightful tricks over deep theoretical knowledge. We introduce Riemann-Bench, a private benchmark of expert-curated problems designed to evaluate AI systems on research-level mathematics that goes far beyond the olympiad frontier. Problems are authored by Ivy League mathematics professors, graduate students, and PhD-holding IMO medalists, and routinely took their authors weeks to solve independently. Each problem undergoes double-blind verification by two independent domain experts who must solve the problem from scratch, and yields a unique, closed-form solution assessed by programmatic verifiers. We evaluate frontier models as unconstrained research agents, with full access to coding tools, search, and open-ended reasoning, using an unbiased statistical estimator computed over 100 independent runs per problem. Our results reveal that all frontier models currently score below 10%, exposing a substantial gap between olympiad-level problem solving and genuine research-level mathematical reasoning. By keeping the benchmark fully private, we ensure that measured performance reflects authentic mathematical capability rather than memorization of training data.

cs.AI↗

arXiv2026

EnterpriseBench Corecraft: Training Generalizable Agents on High-Fidelity RL Environments

We show that training AI agents on high-fidelity reinforcement learning environments produces capabilities that generalize beyond the training distribution. We introduce CoreCraft, the first environment in EnterpriseBench, Surge AI's suite of agentic RL environments. CoreCraft is a fully operational enterprise simulation of a customer support organization, comprising over 2,500 entities across 14 entity types with 23 unique tools, designed to measure whether AI agents can perform the multi-step, domain-specific work that real jobs demand. Frontier models such as GPT-5.2 and Claude Opus 4.6 solve fewer than 30% of tasks when all expert-authored rubric criteria must be satisfied. Using this environment, we train GLM 4.6 with Group Relative Policy Optimization (GRPO) and adaptive clipping. After a single epoch of training, the model improves from 25.37% to 36.76% task pass rate on held-out evaluation tasks. More importantly, these gains transfer to out-of-distribution benchmarks: +4.5% on BFCL Parallel, +7.4% on Tau2-Bench Retail, and +6.8% on Tool Decathlon (Pass@1). We believe three environment properties are consistent with the observed transfer: task-centric world building that optimizes for diverse, challenging tasks; expert-authored rubrics enabling reliable reward computation; and enterprise workflows that reflect realistic professional patterns. Our results suggest that environment quality, diversity, and realism are key factors enabling generalizable agent capabilities.

cs.AI↗

arXiv2026

MATRAG: Multi-Agent Transparent Retrieval-Augmented Generation for Explainable Recommendations

Large Language Model (LLM)-based recommendation systems have demonstrated remarkable capabilities in understanding user preferences and generating personalized suggestions. However, existing approaches face critical challenges in transparency, knowledge grounding, and the ability to provide coherent explanations that foster user trust. We introduce MATRAG (Multi-Agent Transparent Retrieval-Augmented Generation), a novel framework that combined multi-agent collaboration with knowledge graph-augmented retrieval to deliver explainable recommendations. MATRAG employs four specialized agents: a User Modeling Agent that constructs dynamic preference profiles, an Item Analysis Agent that extracts semantic features from knowledge graphs, a Reasoning Agent that synthesizes collaborative and content-based signals, and an Explanation Agent that generates natural language justifications grounded in retrieved knowledge. Our framework incorporates a transparency scoring mechanism that quantifies explanation faithfulness and relevance. Extensive experiments on three benchmark datasets (Amazon Reviews, MovieLens-1M, and Yelp) demonstrate that MATRAG achieves state-of-the-art performance, improving recommendation accuracy by 12.7\% (Hit Rate) and 15.3\% (NDCG) over leading baselines, while human evaluation confirms that 87.4\% of generated explanations are rated as helpful and trustworthy by domain experts. Our work establishes new benchmarks for transparent, agentic recommendation systems and provides actionable insights for deploying LLM-based recommenders in production environments.

cs.IR↗

arXiv2026

HELM: A Human-Centered Evaluation Framework for LLM-Powered Recommender Systems

The integration of Large Language Models (LLMs) into recommendation systems has introduced unprecedented capabilities for natural language understanding, explanation generation, and conversational interactions. However, existing evaluation methodologies focus predominantly on traditional accuracy metrics, failing to capture the multifaceted human-centered qualities that determine the real-world user experience. We introduce \framework{} (\textbf{H}uman-centered \textbf{E}valuation for \textbf{L}LM-powered reco\textbf{M}menders), a comprehensive evaluation framework that systematically assesses LLM-powered recommender systems across five human-centered dimensions: \textit{Intent Alignment}, \textit{Explanation Quality}, \textit{Interaction Naturalness}, \textit{Trust \& Transparency}, and \textit{Fairness \& Diversity}. Through extensive experiments involving three state-of-the-art LLM-based recommenders (GPT-4, LLaMA-3.1, and P5) across three domains (movies, books, and restaurants), and rigorous evaluation by 12 domain experts using 847 recommendation scenarios, we demonstrate that \framework{} reveals critical quality dimensions invisible to traditional metrics. Our results show that while GPT-4 achieves superior explanation quality (4.21/5.0) and interaction naturalness (4.35/5.0), it exhibits a significant popularity bias (Gini coefficient 0.73) compared to traditional collaborative filtering (0.58). We release \framework{} as an open-source toolkit to advance human-centered evaluation practices in the recommender systems community.

cs.IR↗

arXiv2026

The Hierarchy of Agentic Capabilities: Evaluating Frontier Models on Realistic RL Environments

The advancement of large language model (LLM) based agents has shifted AI evaluation from single-turn response assessment to multi-step task completion in interactive environments. We present an empirical study evaluating frontier AI models on 150 workplace tasks within a realistic e-commerce RL environment from Surge. Our analysis reveals an empirically-derived \emph{hierarchy of agentic capabilities} that models must master for real-world deployment: (1) tool use, (2) planning and goal formation, (3) adaptability, (4) groundedness, and (5) common-sense reasoning. Even the best-performing models fail approximately 40\% of the tasks, with failures clustering predictably along this hierarchy. Weaker models struggle with fundamental tool use and planning, whereas stronger models primarily fail on tasks requiring contextual inference beyond explicit instructions. We introduce a task-centric design methodology for RL environments that emphasizes diversity and domain expert contributions, provide detailed failure analysis, and discuss implications for agent development. Our findings suggest that while current frontier models can demonstrate coherent multi-step behavior, substantial capability gaps remain before achieving human-level task completion in realistic workplace settings.

cs.AI↗

arXiv2025

Beyond Accuracy: A Multi-Dimensional Framework for Evaluating Enterprise Agentic AI Systems

Current agentic AI benchmarks predominantly evaluate task completion accuracy, while overlooking critical enterprise requirements such as cost-efficiency, reliability, and operational stability. Through systematic analysis of 12 main benchmarks and empirical evaluation of state-of-the-art agents, we identify three fundamental limitations: (1) absence of cost-controlled evaluation leading to 50x cost variations for similar precision, (2) inadequate reliability assessment where agent performance drops from 60\% (single run) to 25\% (8-run consistency), and (3) missing multidimensional metrics for security, latency, and policy compliance. We propose \textbf{CLEAR} (Cost, Latency, Efficacy, Assurance, Reliability), a holistic evaluation framework specifically designed for enterprise deployment. Evaluation of six leading agents on 300 enterprise tasks demonstrates that optimizing for accuracy alone yields agents 4.4-10.8x more expensive than cost-aware alternatives with comparable performance. Expert evaluation (N=15) confirms that CLEAR better predicts production success (correlation $\rho=0.83$) compared to accuracy-only evaluation ($\rho=0.41$).

cs.AI↗

arXiv2025

Beyond Surface-Level Similarity: Hierarchical Contamination Detection for Synthetic Training Data in Foundation Models

Synthetic data has become essential for training foundation models, yet benchmark contamination threatens evaluation integrity. Although existing detection methods identify token-level overlap, they fail to detect semantic-level contamination where synthetic data conceptually resemble benchmarks without lexical overlap. This gap is critical as foundation models increasingly train on synthetic data that may implicitly encode benchmark knowledge. We propose a hierarchical contamination detection framework operating at four levels: token level, semantic level, reasoning pattern, and performance cliff detection. Through controlled experiments on MMLU, GSM8K and HumanEval, we demonstrate that semantic-level contamination evades existing methods (F1=0.17-0.49) but is effectively detected by our hierarchical approach (F1 = 0.76), with an average improvement of 26. 5\% over state-of-the-art baselines. Our framework provides practitioners with practical tools for audit pipelines and enables responsible deployment of synthetic training data.

cs.LG↗

arXiv2025

Scaling Laws and In-Context Learning: A Unified Theoretical Framework

In-context learning (ICL) enables large language models to adapt to new tasks from demonstrations without parameter updates. Despite extensive empirical studies, a principled understanding of ICL emergence at scale remains more elusive. We present a unified theoretical framework connecting scaling laws to ICL emergence in transformers. Our analysis establishes that ICL performance follows power-law relationships with model depth $L$, width $d$, context length $k$, and training data $D$, with exponents determined by task structure. We show that under specific conditions, transformers implement gradient-based metalearning in their forward pass, with an effective learning rate $\eta_{\text{eff}} = \Theta(1/\sqrt{Ld})$. We demonstrate sharp phase transitions at critical scales and derive optimal depth-width allocations favoring $L^* \propto N^{2/3}$, $d^* \propto N^{1/3}$ for the fixed parameter budget $N = Ld$. Systematic experiments on synthetic tasks validate our predictions, with measured scaling exponents closely matching theory. This work provides both necessary and sufficient conditions for the emergence of ICLs and establishes fundamental computational limits on what transformers can learn in-context.

cs.LG↗

arXiv2025

When Are Learning Biases Equivalent? A Unifying Framework for Fairness, Robustness, and Distribution Shift

Machine learning systems exhibit diverse failure modes: unfairness toward protected groups, brittleness to spurious correlations, poor performance on minority sub-populations, which are typically studied in isolation by distinct research communities. We propose a unifying theoretical framework that characterizes when different bias mechanisms produce quantitatively equivalent effects on model performance. By formalizing biases as violations of conditional independence through information-theoretic measures, we prove formal equivalence conditions relating spurious correlations, subpopulation shift, class imbalance, and fairness violations. Our theory predicts that a spurious correlation of strength $\alpha$ produces equivalent worst-group accuracy degradation as a sub-population imbalance ratio $r \approx (1+\alpha)/(1-\alpha)$ under feature overlap assumptions. Empirical validation in six datasets and three architectures confirms that predicted equivalences hold within the accuracy of the worst group 3\%, enabling the principled transfer of debiasing methods across problem domains. This work bridges the literature on fairness, robustness, and distribution shifts under a common perspective.

cs.LG↗

arXiv2025

Muon: Training and Trade-offs with Latent Attention and MoE

We present a comprehensive theoretical and empirical study of the Muon optimizer for training transformers only with a small to medium decoder (30M - 200M parameters), with an emphasis on its mathematical foundations, convergence properties and synergistic interactions with modern architectural optimizations. Building on recent work showing Muon's scalability, we provide rigorous theoretical analysis including: (i)showing the convergence rate under standard assumptions, (ii) spectral regularization properties that prevent gradient explosion, (iii) connection to natural gradient descent on the Stiefel manifold, and (iv) equivalence to steepest gradient descent under the spectral norm. Crucially, we demonstrate that Muon expands the Pareto frontier in the compute-time trade-off by maintaining superior data efficiency at large batch sizes, a key finding of~\cite{essentialai2025muon} that we validate across our model scales. Empirically, Muon reaches the target loss with 48-52\% of the training calculated by AdamW while maintaining or improving the final perplexity, consistent with larger-scale results. When combined with Multi-Head Latent Attention (MLA) and Mixture-of-Experts (MoE), we observe multiplicative efficiency gains: MLA+MoE+Muon achieves 68\% memory reduction and 3.2$\times$ inference speedup, while improving perplexity by 8-12\%. We provide detailed procedures on 15 architectural and optimizer components, stability analyzes across 100+ training runs, and practical implementation guidelines including Newton-Schulz coefficients $(3.4445, -4.7750, 2.0315)$ optimized by~\cite{su2024muonblog}. Our theoretical analysis and comprehensive experiments establish Muon as a principled, robust alternative to AdamW that particularly excels when combined with modern efficiency techniques and large-batch training regimes.

cs.LG↗

arXiv2025

Unifying Mixture of Experts and Multi-Head Latent Attention for Efficient Language Models

We present MoE-MLA-RoPE, a novel architecture combination that combines Mixture of Experts (MoE) with Multi-head Latent Attention (MLA) and Rotary Position Embeddings (RoPE) for efficient language modeling. Our approach addresses the fundamental trade-off between model capacity and computational efficiency through three key innovations: (1) fine-grained expert routing with 64 micro-experts and top-$k$ selection, enabling flexible specialization through 3.6 * 10^7 possible expert combinations; (2) shared expert isolation that dedicates 2 always active experts for common patterns while routing to 6 of 62 specialized experts; and (3) gradient-conflict-free load balancing that maintains expert utilization without interfering with primary loss optimization. Extensive experiments on models ranging from 17M to 202M parameters demonstrate that MoE-MLA-RoPE with compression ratio r=d/2 achieves 68% KV cache memory reduction and 3.2x inference speedup while maintaining competitive perplexity (0.8% degradation). Compared to the parameters with 53.9M parameters, MoE-MLA-RoPE improves the validation loss by 6.9% over the vanilla transformers while using 42% fewer active parameters per forward pass. FLOP-matched experiments reveal even larger gains: 11.1% improvement with 3.2x inference acceleration. Automated evaluation using GPT-4 as a judge confirms quality improvements in generation, with higher scores on coherence (8.1/10), creativity (7.9/10) and grammatical correctness (8.2/10). Our results establish that architectural novelty, not parameter scaling, defines the efficiency frontier for resource-constrained language model deployment.

cs.AI↗

arXiv2025

Latent Multi-Head Attention for Small Language Models

We present the first comprehensive study of latent multi-head attention (MLA) for small language models, revealing interesting efficiency-quality trade-offs. Training 30M-parameter GPT models on 100,000 synthetic stories, we benchmark three architectural variants: standard multi-head attention (MHA), MLA, and MLA with rotary positional embeddings (MLA+RoPE). Our key finding is that MLA+RoPE with half-rank latent dimensions (r = d/2) achieves a 45% KV-cache memory reduction while incurring only a 0.3% increase in validation loss (essentially matching MHA quality)- a Pareto improvement for memory constrained deployment. We further show that RoPE is crucial for MLA in small models: without it, MLA underperforms vanilla attention by 3-5%, but with RoPE, it surpasses vanilla by 2%. Inference benchmarks on NVIDIA A100 GPUs reveal that MLA with r=d/2 achieves a 1.4 times speedup over full-rank MLA while maintaining the memory savings. GPT-4 evaluations corroborate perplexity results, with ours achieving the highest quality scores (7.4/10) across grammar, creativity, and consistency metrics. Code and models will be released upon acceptance.

cs.CL↗