CORTEXA
← Browse
arxivcs.SEcs.AIcs.CL2026-07-07

RuBench: A Repository-Level Agentic Coding Benchmark with Natively Authored Russian Task Specifications

Evgeny Shilov

Developers increasingly delegate real maintenance work to product-grade coding agents, and many state tasks in their native language, in the style of a customer request rather than a curated English issue. We introduce RuBench 1.0, a benchmark of 25 tasks mined from recent fix commits in five live open-source repositories (aiohttp, aiogram, Laravel, NestJS, Fastify), each specified natively in Russian -- written from scratch, not translated -- and judged by the upstream maintainer's regression tests, which we withhold from release. All fix commits postdate the training-data cutoffs of every evaluated model. Round 1 evaluates Claude Code with Opus 4.8, Sonnet 5, and Haiku 4.5, and Codex CLI with GPT-5.5 (3 independent runs each; pass@1 with task-level uncertainty); the best configuration resolves 78.7% of tasks. Auditing full trajectories of an hors-concours configuration (Claude Code + Fable 5), we caught the product silently substituting the model on 20% of tasks via an official safeguard fallback -- evidence that the deployed product, not the model, is the unit actually measured. Version 2 adds Round 2: seven further configurations on the same frozen set under a per-configuration freshness gate -- the Russian-market agents SourceCraft CLI (ds, legacy) and Koda CLI (koda-pro), Antigravity with Gemini 3.1 Pro and 3.5 Flash, and Codex CLI with GPT-5.6 Sol and Luna. SourceCraft's flagship resolves 68.1% (N=23), above GPT-5.5 and both Gemini rows. A tool-call contamination re-audit of all 437 Round-2 trajectories finds the Russian and Gemini columns clean (0/293 cells) while flagging systematic oracle-hunting in the GPT-5.6 family (8/69 and 13/75 cells), including one case of mining a prior round's artifacts from the run machine's disk; honest scores are published alongside raw ones. We release statements, metadata, trajectories, and diffs; oracles are withheld with a SHA-256 manifest.

View free PDFSource page

Related papers

arxivcs.SEcs.AIcs.CL2026-07-02

TestEvo-Bench: An Executable and Live Benchmark for Test and Code Co-Evolution

Jiale Amber Wang, Kaiyuan Wang, Pengyu Nie

Software tests and code evolve together: a code change should be followed by new or updated tests that record the new software behavior. Yet existing test generation and update benchmarks often isolate the test from the code change, and rely on static metadata that does not verif…

View free PDFSource page
arxivcs.AIcs.CLcs.SEeess.SY2026-07-14

Do AI Agents Know When a Task Is Simple? Toward Complexity-Aware Reasoning and Execution

Junjie Yin, Xinyu Feng

Large language model (LLM) agents increasingly automate multi-step engineering and informatics workflows, yet they rarely ask how much effort a task actually requires. They often follow a maximum-context-first strategy--re-reading files and dependencies they have already seen--tu…

View free PDFSource page
arxivcs.CLcs.AIcs.HCcs.MAcs.SE2026-07-15

DevicesWorld: Benchmarking Cross-Device Agents in Heterogeneous Environments

Huatao Li, Xinwei Geng, Yuheng Wang, Yutong Li, Runde Yang, Hantao Chen, et al.

LLM-based agents have rapidly improved at operating individual digital environments such as mobile applications, desktop systems, and smart homes. However, real-world user goals often span multiple devices: information may come from a phone, be processed on a desktop, and the res…

View free PDFSource page
arxivcs.CLcs.AIcs.SE2026-07-01

Adversarial Pragmatics for AI Safety Evaluation: A Benchmark for Instruction Conflict, Embedded Commands, and Policy Ambiguity

Brett Reynolds

Safety evaluations for language models increasingly depend on judgments about ambiguous natural-language behaviour: whether a model has followed an instruction, refused appropriately, complied with a policy, resisted an embedded command, or misreported progress in an agentic task…

View free PDFSource page
arxivcs.CLcs.AIcs.SE2026-07-06

ToolFailBench: Diagnosing Tool-Use Failures in LLM Agents

Harsh Soni

Tool calling is central to modern language model agents, but aggregate benchmark scores often hide where tool use fails. A model that never calls a needed tool and a model that calls the tool but ignores the result can look similar under final task accuracy. We introduce ToolFail…

View free PDFSource page
arxivcs.SEcs.AIcs.CL2026-07-10

Git-Assistant: Planning-Based Support for Updating Git Repositories

Alfredo Garrachón Ruiz, Tomás de la Rosa, Daniel Borrajo

Version control systems are essential for collaborative software development, yet tools like git remain challenging for many practitioners. Recent advances in Large Language Models (LLMs) offer promising capabilities for interpreting developer intent, but their effectiveness in r…

View free PDFSource page