arXiv subjects

Shreshth Rajan

Publications and source records attributed to Shreshth Rajan.

6 recordsLinked to original sources

arXiv2026

Auditing Reward Hackability in Code RL Training Environments

We measure the rate at which code RL environments accept incorrect solutions as correct. On a 49-task sample of SWE-bench Verified, 28.5% of tasks have test suites weak enough that a Docker-verified incorrect patch passes them. On 20 R2E-Gym tasks across 6 repositories, the same pipeline at single-shot exploit generation yields 25.0%. A random-effects meta-analysis over 134 frontier model submissions to SWE-bench Verified finds, within the same human-rated difficulty stratum, model Pass@1 is +14.14 percentage points higher on flagged-hackable tasks than on robust ones (95% CI [+11.80, +16.48]; one-sided p < 10^-6; I^2 = 0%; 123 of 134 models positive). We then describe a procedure for hardening the broken tasks. An inline LLM judge with a Docker gold-sanity gate runs each generated test against the gold solution before the judge is consulted. On the 11 broken tasks in the audit, the gate flags 65 of 105 decisive LLM-generated tests as failing on the gold patch itself, a 61.9% per-augmentation defect rate the LLM judge alone misses. With diversity-biased retry, the loop converges 9 of 11 tasks to a gated upgrade.

cs.AI↗

arXiv2026

MultiVer: Zero-Shot Multi-Agent Vulnerability Detection

We present MultiVer, a zero-shot multi-agent system for vulnerability detection that achieves state-of-the-art recall without fine-tuning. A four-agent ensemble (security, correctness, performance, style) with union voting achieves 82.7% recall on PyVul, exceeding fine-tuned GPT-3.5 (81.3%) by 1.4 percentage points -- the first zeroshot system to surpass fine-tuned performance on this benchmark. On SecurityEval, the same architecture achieves 91.7% detection rate, matching specialized systems. The recall improvement comes at a precision cost: 48.8% precision versus 63.9% for fine-tuned baselines, yielding 61.4% F1. Ablation experiments isolate component contributions: the multi-agent ensemble adds 17 percentage points recall over single-agent security analysis. These results demonstrate that for security applications where false negatives are costlier than false positives, zero-shot multi-agent ensembles can match and exceed fine-tuned models on the metric that matters most.

cs.MA↗

arXiv2026

Social Contagion and Bank Runs: An Agent-Based Model with LLM Depositors

Digital banking and online communication have made modern bank runs faster and more networked than the canonical queue-at-the-branch setting. While equilibrium models explain why strategic complementarities generate run risk, they offer limited guidance on how beliefs synchronize and propagate in real time. We develop a process-based agent-based model that makes the information and coordination layer explicit. Banks follow cash-first withdrawal processing with discounted fire-sale liquidation and an endogenous stress index. Depositors are heterogeneous in risk tolerance and in the weight placed on fundamentals versus social information, communicating on a heavy-tailed network calibrated to Twitter activity during March 2023. Depositor behavior is generated by a constrained large language model that maps each agent's information set into a discrete action and an optional post; we validate this policy against laboratory coordination evidence and theoretical benchmarks. Across 4,900 configurations and full LLM simulations, three findings emerge. Within-bank connectivity raises the likelihood and speed of withdrawal cascades holding fundamentals fixed. Cross-bank contagion exhibits a sharp phase transition near spillover rates of 0.10. Depositor overlap and network amplification interact nonlinearly, so channels weak in isolation become powerful in combination. In an SVB, First Republic, and regional bank scenario disciplined by crisis-era data, the model reproduces the observed ordering of failures and predicts substantially higher withdrawal rates among uninsured depositors. The results frame social correlation as a measurable amplifier of run risk alongside balance-sheet fundamentals.

physics.soc-ph↗

arXiv2025

Uncertainty-Gated Region-Level Retrieval for Robust Semantic Segmentation

Semantic segmentation of outdoor street scenes plays a key role in applications such as autonomous driving, mobile robotics, and assistive technology for visually-impaired pedestrians. For these applications, accurately distinguishing between key surfaces and objects such as roads, sidewalks, vehicles, and pedestrians is essential for maintaining safety and minimizing risks. Semantic segmentation must be robust to different environments, lighting and weather conditions, and sensor noise, while being performed in real-time. We propose a region-level, uncertainty-gated retrieval mechanism that improves segmentation accuracy and calibration under domain shift. Our best method achieves an 11.3% increase in mean intersection-over-union while reducing retrieval cost by 87.5%, retrieving for only 12.5% of regions compared to 100% for always-on baseline.

cs.CV↗

arXiv2025

Arithmetic-Intensity-Aware Quantization

As modern neural networks become increasingly memory-bound, inference throughput is limited by DRAM bandwidth rather than compute. We present Arithmetic-Intensity-Aware Quantization (AIQ), a mixed precision quantization framework that chooses per-layer bit-widths to maximize arithmetic intensity (AI) while minimizing accuracy loss. AIQ is a post-training quantization method that uses search algorithms over per-layer quantization schemes to minimize a weighted loss over AI and accuracy. On ResNet-20/CIFAR-10, AIQ increases AI by ~50% over an FP32 baseline while keeping test accuracy within ~1 percentage point, and outperforming global uniform quantization schemes. On a memory-bound MobileNetV2 architecture, AIQ configurations give a 1.66x higher throughput than the FP32 baseline while keeping test accuracy within 1 percentage point. We also find that AIQ naturally quantizes larger layers more aggressively.

cs.LG↗

arXiv2025

Multi-Agent Code Verification via Information Theory

LLMs generate buggy code: 29.6% of SWE-bench solved patches fail, 62% of BaxBench solutions have vulnerabilities, and existing tools only catch 65% of bugs with 35% false positives. We built CodeX-Verify, a multi-agent system that uses four specialized agents to detect different types of bugs. We prove mathematically that combining agents with different detection patterns finds more bugs than any single agent when the agents look for different problems, using submodularity of mutual information under conditional independence. Measuring agent correlation of rho = 0.05 to 0.25 confirms they detect different bugs. Testing on 99 code samples with verified labels shows our system catches 76.1% of bugs, matching the best existing method (Meta Prompt Testing: 75%) while running faster and without test execution. We tested all 15 agent combinations and found that using multiple agents improves accuracy by 39.7 percentage points (from 32.8% to 72.4%) compared to single agents, with diminishing returns of +14.9pp, +13.5pp, and +11.2pp for agents 2, 3, and 4, validating our theoretical model. The best two-agent combination (Correctness + Performance) reaches 79.3% accuracy. Testing on 300 real patches from Claude Sonnet 4.5 runs in under 200ms per sample, making this practical for production use.

cs.SE↗