Searcharxiv⌕ Search

arXiv subjects

Saurabh Jha

Publications and source records attributed to Saurabh Jha.

At least 19 recordsLinked to original sources
arXiv2026

Towards a Belief-Based World Model for LLM Agents

Large language models (LLMs) are being used as policies for autonomous decision-making and planning in many domains. Despite their strong reasoning capabilities, LLMs struggle with long-horizon tasks, especially under partial observability. World models are a promising way to enhance policy performance, both during training and inference. During inference, agents currently use world models to simulate the consequences of candidate actions before committing to an action, which can improve decision-making. However, we argue that simulation alone is an incomplete interface for decision-making under partial observability: simulation doesn't adequately capture uncertainty about the current state, which agents may need for accurate decision-making. We address this limitation with Belief-Based World Models (BB-WMs), which model and maintain a belief that LLMs can query to access information on what is known and uncertain about the current state. Before developing methods to learn accurate BB-WMs, we first ask a more fundamental question: does exposing a world model's belief directly to an LLM policy improve decision-making? Our results show that giving LLM agents access to world model beliefs improves task performance under partial observability, while remaining complementary to existing simulation-based world models. Code is released at https://github.com/skumar-ml/belief-world-models.

cs.AI↗
arXiv2026

The Missing Layer: Specification Infrastructure for AI Oversight

AI safety has a missing layer. Interpretability, formal methods, security engineering, evaluation methodology, and reinforcement-learning safety each produce substantial work, but the resulting artifacts do not compose into deployable oversight: every team fielding an agentic system builds its own audit schema, policy dialect, monitoring stack, and escalation path, mostly reinventions of patterns understood elsewhere. We diagnose this as a coordination gap, not a research gap, and propose a two-axis taxonomy: five technical layers (Legibility, Specification, Mediation, Evaluation, Escalation) crossed with six concerns spanning alignment, robustness, adversarial defense, security, governance, and accountability, populating the resulting 5x6 matrix with existing work. Layer 2 (Specification), where humans translate intent into machine-checkable artifacts, is the connective tissue every layer depends on, yet it lacks four marks of a mature engineering discipline: shared vocabulary, design principles, composability standards, and governance practices. We propose six design principles for Layer 2, from elicitability and composability to adversary-awareness, traceability, and governability, made concrete through worked examples and a reference architecture turning specifications into runtime enforcement, evaluation, and escalation. Existing systems such as Cedar, Constitutional AI, and Open Policy Agent each address a fragment of Layer 2 well and the matrix poorly; treating them as fragments of one shared layer makes composition tractable. As evidence, we introduce CARMA, a Layer 2 prototype for autonomous ETL agents in which one specification drives enforcement, evaluation, and escalation, with every decision traceable to a versioned specification, naming what AI oversight is missing and giving independent teams principles to build the missing pieces so they compose.

cs.CR↗
arXiv2026

CADENCE: Closing the Reasoning Gap via Coverage-Adaptive On-Policy Distillation

On-policy knowledge distillation transfers reasoning from large teachers to compact students, but existing approaches suffer three compounding failure modes: (i) cold-start collapse, where a fresh student assigns near-zero mass to teacher-preferred tokens; (ii) state-agnostic divergence scheduling, where time-only forward/reverse-KL interpolation ignores the student's coverage state; and (iii) binary reward sparsity, where pass/fail signals discard information from partially correct traces. We present CADENCE, a unified framework with a targeted fix for each. Its DRIFT mechanism schedules a per-token convex mixture of forward-KL and reverse-KL surrogate objectives on student-sampled trajectories (per-token surrogates, not sequence-level KL gradient estimators). Six components extend it: (A) COVA, a coverage-adaptive $β$ schedule accelerating the forward-to-reverse transition; (B) FTB, a forking-token boost concentrating gradient at high-entropy positions via a globally-normalized entropy reference; (C) CCD, a dense reward adding numerical-proximity partial credit for incorrect-but-close traces; (D) LAP, brevity-preferential correct-rollout reinforcement; (E) EMR, an entropy-matching calibration regularizer; (F) BSD, a bootstrapped self-distillation phase. On GSM8K and MATH-500 (corrected 512-token protocol, 5 seeds, reported std), CADENCE distills a 0.5B student from a 1.5B teacher to 69.8 $\pm$ 0.5% GSM8K pass@1 (from 48.7% pretrained; 63.2% of the teacher gap closed) and to 72.1 $\pm$ 0.4% with a 3B teacher (76.2% closed), beating the strongest matched-compute label-using baseline (DRIFT+binary reward) by +4.4 $\pm$ 0.7 points. All experiments run on a single Apple Mac Studio (M-series, 64GB unified memory), showing principled distillation reaches strong reasoning quality without datacenter-scale hardware.

cs.LG↗
arXiv2026

PRAXIS: Integrating Program Analysis with Observability for Root-Cause Analysis

Unresolved production cloud incidents cost an average of over $2M per hour. This paper introduces PRAXIS, an orchestrator that manages and deploys an agentic workflow for diagnosing code- and configuration-caused cloud incidents. PRAXIS employs an LLM-driven structured traversal over two types of graph: (1) a service dependency graph (SDG) that captures microservice-level dependencies; and (2) a hammock-block program dependence graph (PDG) that captures code-level dependencies for each microservice. Compared to state-of-the-art ReAct baselines, PRAXIS improves RCA accuracy by up to 6.3x while reducing token consumption by 5.3x. PRAXIS is demonstrated on a set of 30 comprehensive real-world incidents that is being compiled into an RCA benchmark.

cs.DC↗
arXiv2026

Old Universe, Young SNe Ia: A Statistical Analysis of Type Ia Supernova Progenitor Age from 6,983 TITAN Host Galaxies, and Implications for Cosmology

Correlations between standardized Type Ia supernova (SN Ia) luminosities and host-galaxy properties are routinely modeled to avoid bias in cosmological parameter inference. A recent hypothesis attributes these correlations to progenitor-age variations and, combined with a strong ($\sim$5-6 Gyr) age evolution between low- and high-redshift samples, could alter cosmological conclusions. We test this scenario using the SN Ia host galaxies of TITAN DR1, the largest low-redshift sample of its kind to date (6,983 hosts; 0 $\lesssim$ z $\lesssim$ 0.15). Progenitor ages are estimated by combining host-galaxy star-formation histories (SFHs) with empirical delay-time distributions. The SFHs are constrained via spectral energy distribution (SED) fitting of photometry spanning ultraviolet (UV) to mid-infrared (MIR) wavelengths, enabling robust separation of dusty star-forming and quiescent systems. The resulting progenitor-age distribution has a mean of 3.5 Gyr, substantially younger than predicted by strong-evolution models. It is strongly peaked near 2.2 Gyr, predominantly from star-forming hosts (60% of the sample), with a smaller, broader component centered near 6.0 Gyr from quiescent systems. Restricting to high-mass galaxies (in order to isolate progenitor effects from the mass-step), the age difference between host types reduces to 3.3 Gyr which, under the age-dependence hypothesis, would imply a 0.10 mag luminosity offset, inconsistent with observed standardized magnitudes. We infer a modest 1.5 Gyr evolution in mean progenitor age over cosmic time which, combined with observed age-Hubble-residual (HR) relations, yields a maximum redshift-dependent bias of $Δ$HR = $-0.007^{+0.012}_{-0.014}$ mag, consistent with zero. We find no evidence for a large, unmodeled progenitor-age systematic beyond what is already captured, to good approximation, by standard host-mass corrections.

astro-ph.CO↗
arXiv2026

Forge-UGC: FX optimization and register-graph engine for universal graph compiler

We present Forge-UGC (FX Optimization and Register-Graph Engine for Universal Graph Compilation), a four-phase compiler for transformer deployment on heterogeneous accelerator hardware, validated on Intel AI Boost NPU. Existing frameworks such as OpenVINO and ONNX Runtime often use opaque compilation pipelines, limited pass-level visibility, and weak buffer management, which can lead to higher compilation cost and runtime overhead. Forge-UGC addresses this with a hardware-agnostic design that separates graph capture, optimization, intermediate representation lowering, and backend scheduling. Phase 1 captures graphs with torch.export at the ATen operator level, supporting modern transformer components such as rotary position embeddings, grouped-query attention, and SwiGLU without manual decomposition. Phase 2 applies six optimization passes: dead code elimination, common subexpression elimination, constant folding, attention fusion, operator fusion, and layout optimization, reducing graph node count by 14.2 to 21.9%. Phase 3 lowers the optimized graph into a typed intermediate representation with explicit virtual register assignments. Phase 4 performs liveness analysis, linear-scan buffer allocation, reducing peak buffer count by 30 to 48%, and device-affinity scheduling, reducing NPU-CPU transitions by 42 to 65%. Across six model families ranging from 125M to 8B parameters, evaluated on WikiText-103 and GLUE, Forge-UGC delivers 6.9 to 9.2x faster compilation than OpenVINO and ONNX Runtime, 18.2 to 35.7% lower inference latency, and 30.2 to 40.9% lower energy per inference. Fidelity is preserved, with max absolute logit differences below 2.1e-5 and KL divergence below 8.4e-9. We also introduce Fusion Gain Ratio, Compilation Efficiency Index, and per-pass execution profiling for systematic evaluation of NPU compilation pipelines.

cs.AR↗
arXiv2026

QEIL v2: Heterogeneous Computing for Edge Intelligence via Roofline-Derived Pareto-Optimal Energy Modeling and Multi-Objective Orchestration

Deploying large language models (LLMs) on heterogeneous edge devices demands frameworks that jointly optimize energy efficiency, inference quality, and reliability. Our prior QEIL v1 (Kumar & Jha, 2026) achieved 4.82x IPW improvement but relied on static efficiency factors, greedy optimization, and unverified candidate selection. QEIL v2 replaces every static heuristic with physics-grounded, runtime-adaptive models. We introduce three device-workload metrics: DASI (roofline-derived compute utilization), CPQ (memory pressure from allocation theory), and Phi (thermal yield from CMOS leakage physics), forming a unified energy equation with every coefficient traceable to semiconductor physics. For optimization, PGSAM (Pareto-Guided Simulated Annealing with Momentum) simultaneously minimizes energy, latency, and device underutilization. At inference time, the EAC/ARDE selection cascade with CSVET early stopping provides progressive verification among repeated samples. Evaluated on WikiText-103, GSM8K, and ARC-Challenge across seven model families (125M-8B parameters, including one pre-quantized variant), QEIL v2 achieves 75.7% pass@k at 63.8W (IPW=0.9749), a 2.86x improvement over standard inference. When applied to a 4-bit Llama-3.1-8B, QEIL v2's physics-grounded routing achieves IPW=1.024 at 54.8W -- the first edge orchestration system to surpass the IPW=1.0 empirical reference mark, with the gain attributable entirely to QEIL v2's workload-adaptive device allocation on a model with reduced memory bandwidth requirements. Total energy drops 75.6% vs. standard with 38.3% latency reduction, zero thermal throttling, and 100% fault recovery across all benchmarks and model families.

cs.DC↗
arXiv2026

A Context Engineering Framework for Improving Enterprise AI Agents based on Digital-Twin MDP

Despite rapid progress in AI agents for enterprise automation and decision-making, their real-world deployment and further performance gains remain constrained by limited data quality and quantity, complex real-world reasoning demands, difficulties with self-play, and the lack of reliable feedback signals. To address these challenges, we propose a lightweight, model-agnostic framework for improving LLM-based enterprise agents via offline reinforcement learning (RL). The proposed Context Engineering via DT-MDP (DT-MDP-CE) framework comprises three key components: (1) A Digital-Twin Markov Decision Process (DT-MDP), which abstracts the agent's reasoning behavior as a finite MDP; (2) A robust contrastive inverse RL, which, armed with the DT-MDP, to efficiently estimate a well-founded reward function and induces policies from mixed-quality offline trajectories; and (3) RL-guided context engineering, which uses the policy obtained from the integrated process of (1) and (2), to improve the agent's decision-making behavior. As a case study, we apply the framework to a representative task in the enterprise-oriented domain of IT automation. Extensive experimental results demonstrate consistent and significant improvements over baseline agents across a wide range of evaluation settings, suggesting that the framework can generalize to other agents sharing similar characteristics in enterprise environments.

cs.AI↗
arXiv2026

STRATUS: A Multi-agent System for Autonomous Reliability Engineering of Modern Clouds

In cloud-scale systems, failures are the norm. A distributed computing cluster exhibits hundreds of machine failures and thousands of disk failures; software bugs and misconfigurations are reported to be more frequent. The demand for autonomous, AI-driven reliability engineering continues to grow, as existing humanin-the-loop practices can hardly keep up with the scale of modern clouds. This paper presents STRATUS, an LLM-based multi-agent system for realizing autonomous Site Reliability Engineering (SRE) of cloud services. STRATUS consists of multiple specialized agents (e.g., for failure detection, diagnosis, mitigation), organized in a state machine to assist system-level safety reasoning and enforcement. We formalize a key safety specification of agentic SRE systems like STRATUS, termed Transactional No-Regression (TNR), which enables safe exploration and iteration. We show that TNR can effectively improve autonomous failure mitigation. STRATUS significantly outperforms state-of-the-art SRE agents in terms of success rate of failure mitigation problems in AIOpsLab and ITBench (two SRE benchmark suites), by at least 1.5 times across various models. STRATUS shows a promising path toward practical deployment of agentic systems for cloud reliability.

cs.DC↗
arXiv2026

RAMP: Reinforcement Adaptive Mixed Precision Quantization for Efficient On Device LLM Inference

Post training quantization is essential for deploying large language models (LLMs) on resource constrained hardware, yet state of the art methods enforce uniform bit widths across layers, yielding suboptimal accuracy efficiency trade offs. We present RAMP (Reinforcement Adaptive Mixed Precision), an off policy Soft Actor Critic framework that learns per layer bit width assignments to minimize perplexity under a global bit budget. The policy conditions on an 11 dimensional embedding of activation statistics, weight properties, and structural descriptors, enabling zero shot transfer across model families and scales. To enable stable sub 4 bit quantization, we introduce Scale Folding, a preconditioning technique that migrates activation outliers into weights via per channel scaling and normalization layer compensation. A quality prioritized reward with asymmetric penalties and budget cliffs drives rapid convergence. On Llama 2 7B, RAMP achieves 5.54 perplexity at 3.68GB (3.65 effective bits), outperforming uniform 4 bit AWQ (5.60 at 3.90 GB) and GPTQ by 6% in size and 1% to3% in quality. Critically, a policy trained only on Llama 2 7B generalizes zero shot to Llama 2 13B and Mistral 7B, often surpassing target specific training, supporting the hypothesis that quantization sensitivity is primarily architectural. The HALO pipeline exports allocations to GGUF format for kernel free inference on CPUs, GPUs, and edge devices, retaining 99.5% of FP16 commonsense reasoning performance.

cs.LG↗
arXiv2026

The Energy of Falsehood: Detecting Hallucinations via Diffusion Model Likelihoods

Large Language Models (LLMs) frequently hallucinate plausible but incorrect assertions, a vulnerability often missed by uncertainty metrics when models are confidently wrong. We propose DiffuTruth, an unsupervised framework that reconceptualizes fact verification via non equilibrium thermodynamics, positing that factual truths act as stable attractors on a generative manifold while hallucinations are unstable. We introduce the Generative Stress Test, claims are corrupted with noise and reconstructed using a discrete text diffusion model. We define Semantic Energy, a metric measuring the semantic divergence between the original claim and its reconstruction using an NLI critic. Unlike vector space errors, Semantic Energy isolates deep factual contradictions. We further propose a Hybrid Calibration fusing this stability signal with discriminative confidence. Extensive experiments on FEVER demonstrate DiffuTruth achieves a state of the art unsupervised AUROC of 0.725, outperforming baselines by 1.5 percent through the correction of overconfident predictions. Furthermore, we show superior zero shot generalization on the multi hop HOVER dataset, outperforming baselines by over 4 percent, confirming the robustness of thermodynamic truth properties to distribution shifts.

cs.CL↗
arXiv2026

StreamServe: Adaptive Speculative Flows for Low-Latency Disaggregated LLM Serving

Efficient LLM serving must balance throughput and latency across diverse, bursty workloads. We introduce StreamServe, a disaggregated prefill decode serving architecture that combines metric aware routing across compute lanes with adaptive speculative decoding that tunes speculation depth online from runtime signals. StreamServe comprises four components: StreamScheduler for request orchestration, FlowGuard for multi signal routing, PipeServe Engine for disaggregated prefill decode execution on multi GPU, and SpecuStream for runtime adaptive speculation. We evaluate StreamServe on four benchmarks ALPACA, GSM8K, HUMANEVAL, and SUM with 80 queries each and 320 total using 4 A800 40GB GPUs configured as two stream pairs. Across these workloads, StreamServe reduces latency by 11 to 18 times relative to tensor parallel vLLM baselines and reaches throughput up to 2235 tokens per second on summarization tasks. Time per output token remains stable across configurations, indicating that the gains arise from architectural efficiency rather than token quality degradation. Although evaluated on a single node 4 GPU setup, these results suggest that jointly adapting routing and speculation within a disaggregated framework creates a distinct operating regime for LLM inference.

cs.DC↗
arXiv2026

Think Locally, Explain Globally: Graph-Guided LLM Investigations via Local Reasoning and Belief Propagation

LLM agents excel when environments are mostly static and the needed information fits in a model's context window, but they often fail in open-ended investigations where explanations must be constructed by iteratively mining evidence from massive, heterogeneous operational data. These investigations exhibit hidden dependency structure: entities interact, signals co-vary, and the importance of a fact may only become clear after other evidence is discovered. Because the context window is bounded, agents must summarize intermediate findings before their significance is known, increasing the risk of discarding key evidence. ReAct-style agents are especially brittle in this regime. Their retrieve-summarize-reason loop makes conclusions sensitive to exploration order and introduces run-to-run non-determinism, producing a reliability gap where Pass-at-k may be high but Majority-at-k remains low. Simply sampling more rollouts or generating longer reasoning traces does not reliably stabilize results, since hypotheses cannot be autonomously checked as new evidence arrives and there is no explicit mechanism for belief bookkeeping and revision. In addition, ReAct entangles semantic reasoning with controller duties such as tool orchestration and state tracking, so execution errors and plan drift degrade reasoning while consuming scarce context. We address these issues by formulating investigation as abductive reasoning over a dependency graph and proposing EoG (Explanations over Graphs), a disaggregated framework in which an LLM performs bounded local evidence mining and labeling (cause vs symptom) while a deterministic controller manages traversal, state, and belief propagation to compute a minimal explanatory frontier. On a representative ITBench diagnostics task, EoG improves both accuracy and run-to-run consistency over ReAct baselines, including a 7x average gain in Majority-at-k entity F1.

cs.AI↗
arXiv2026

SN 2023zcu: A Type IIP SN with Early Flash Features

We present a detailed photometric and spectroscopic analysis of the Type IIP supernova SN~2023zcu, which exploded in the galaxy NGC~2139 (redshift $z$ = 0.006). SN~2023zcu exhibits a well-sampled light curve covering the rise, plateau, and nebular phases. It has an optically thick phase of $100.6 \pm 0.6$ d with a magnitude drop of $\sim$1.7 mag in the {\em V} band during the transition between the plateau and the nebular phases. Weak emission features in the early-time spectra indicate a low-level interaction between circumstellar material (CSM) and the SN ejecta. The spectral evolution is well sampled and exhibits a prominent P-Cygni profile of H$α$, a defining characteristic of Type IIP SNe. Signatures of metal-line formation (e.g., \ion{Fe}{2}, \ion{Ca}{2} near-infrared triplet) are also evident in the spectra as the SN evolves. Spectral modeling with the radiative-transfer code \texttt{TARDIS} during the early photospheric phase (8.7--35.5 d since explosion) yields photospheric temperatures decreasing from $\sim$9,000 to $\sim$6,000 K and expansion velocities declining from $\sim$10,000 to $\sim$5,400 km s$^{-1}$. A tailored expanding photosphere method (EPM) fit based on the \texttt{TARDIS} models provides a distance estimate of $27.8 \pm 2.0$ Mpc. Nebular-phase spectra and bolometric light-curve modeling suggest a progenitor mass in the range 12--15 M$_\odot$. This thorough analysis helps to constrain progenitor properties and explosion parameters, thereby strengthening our understanding of Type IIP SNe.

astro-ph.HE↗
arXiv2025

The Ultraviolet Type Ia Supernova CubeSat (UVIa): Science Motivation & Mission Concept

The Ultraviolet (UV) Type Ia Supernova Mission (UVIa) is a CubeSat/SmallSat concept that stands to test critical space-borne UV technology for future missions like the Habitable Worlds Observatory (HWO) while elucidating long-standing questions about the explosion mechanisms of Type Ia supernovae (SNe Ia). UVIa will observe whether any SNe Ia emit excess UV light shortly after explosion to test progenitor/explosion models and provide follow-up over many days to characterize their UV and optical flux variations over time, assembling a comprehensive multi-band UV and optical low-redshift anchor sample for upcoming high-redshift SNe Ia surveys (e.g., Euclid, Vera Rubin Observatory, Nancy Roman Space Telescope). UVIa's mission profile requires it to perform rapid and frequent visits to newly discovered SNe Ia, simultaneously observing each SNe Ia in two UV bands (FUV: 1500-1800A and NUV: 1800-2400A) and one optical band (u-band: 3000-4200A). In this study, we describe the UVIa mission concept science motivation and basic mission design. The UVIa mission concept has been submitted to the CubeSats category of the NASA ROSES Astrophysics Research & Analysis (APRA) program (\$10M cost cap) and NASA Astrophysics Pioneers program (\$20M cost cap).

astro-ph.IM↗
arXiv2025

Story of Two GPUs: Characterizing the Resilience of Hopper H100 and Ampere A100 GPUs

This study characterizes GPU resilience in Delta, a large-scale AI system that consists of 1,056 A100 and H100 GPUs, with over 1,300 petaflops of peak throughput. We used 2.5 years of operational data (11.7 million GPU hours) on GPU errors. Our major findings include: (i) H100 GPU memory resilience is worse than A100 GPU memory, with 3.2x lower per-GPU MTBE for memory errors, (ii) The GPU memory error-recovery mechanisms on H100 GPUs are insufficient to handle the increased memory capacity, (iii) H100 GPUs demonstrate significantly improved GPU hardware resilience over A100 GPUs with respect to critical hardware components, (iv) GPU errors on both A100 and H100 GPUs frequently result in job failures due to the lack of robust recovery mechanisms at the application level, and (v) We project the impact of GPU node availability on larger-scales and find that significant overprovisioning of 5% is necessary to handle GPU failures.

cs.DC↗
arXiv2025

Repairing Tool Calls Using Post-tool Execution Reflection and RAG

Agentic systems interact with external systems by calling tools such as Python functions, REST API endpoints, or command line tools such as kubectl in Kubernetes. These tool calls often fail for various syntactic and semantic reasons. Some less obvious semantic errors can only be identified and resolved after analyzing the tool's response. To repair these errors, we develop a post-tool execution reflection component that combines large language model (LLM)-based reflection with domain-specific retrieval-augmented generation (RAG) using documents describing both the specific tool being called and troubleshooting documents related to the tool. For this paper, we focus on the use case of the kubectl command line tool to manage Kubernetes, a platform for orchestrating cluster applications. Through a larger empirical study and a smaller manual evaluation, we find that our RAG-based reflection will repair kubectl commands such that they are both more likely to successfully execute (pass rate) for 55% of our models evaluated and 36% more likely to correctly answer the user query on average. We find that troubleshooting documents improve pass rate compared to official documentation by an average of 10%.

cs.SE↗
arXiv2025

CPU-Limits kill Performance: Time to rethink Resource Control

Research in compute resource management for cloud-native applications is dominated by the problem of setting optimal CPU limits -- a fundamental OS mechanism that strictly restricts a container's CPU usage to its specified CPU-limits . Rightsizing and autoscaling works have innovated on allocation/scaling policies assuming the ubiquity and necessity of CPU-limits . We question this. Practical experiences of cloud users indicate that CPU-limits harms application performance and costs more than it helps. These observations are in contradiction to the conventional wisdom presented in both academic research and industry best practices. We argue that this indiscriminate adoption of CPU-limits is driven by erroneous beliefs that CPU-limits is essential for operational and safety purposes. We provide empirical evidence making a case for eschewing CPU-limits completely from latency-sensitive applications. This prompts a fundamental rethinking of auto-scaling and billing paradigms and opens new research avenues. Finally, we highlight specific scenarios where CPU-limits can be beneficial if used in a well-reasoned way (e.g. background jobs).

cs.DC↗