paper_id stringlengths 12 12 | title stringlengths 32 143 | published_date stringdate 2025-06-04 03:17:19 2026-08-19 05:44:29 | arxiv_categories listlengths 1 5 | legal_license stringclasses 1
value | license_url stringclasses 1
value | commercial_use_allowed bool 1
class | commercial_ip_safety_score int64 60 80 | commercial_ip_verdict stringclasses 2
values | commercial_ip_justification stringclasses 3
values | academic_citations_count int64 0 0 | influential_citations_count int64 0 0 | field_weighted_citation_impact_fwci float64 1 1 | tldr_neural_summary stringlengths 191 235 | primary_author stringlengths 7 18 | total_authors_count int64 1 172 | authors listlengths 1 172 | affiliations listlengths 1 1 | arxiv_url stringlengths 33 33 | code_audit_status stringclasses 2
values | primary_repo_url stringlengths 14 65 | repo_urls listlengths 0 4 | github_stars int64 0 131k | github_forks int64 0 4.29k | github_last_commit stringclasses 5
values | github_license stringclasses 4
values | importance_score float64 72.2 141 | abstract stringlengths 710 1.92k | title_vector_384d list | abstract_vector_384d list | target_industry stringclasses 1
value | code_agent_execution_mode stringclasses 8
values | programming_languages_supported listlengths 1 3 | coding_foundation_backbone stringclasses 5
values | tested_benchmarks listlengths 1 3 | extracted_benchmark_scores listlengths 1 3 | benchmark_leaderboard_json listlengths 1 3 | semantic_nearest_neighbors_top3 listlengths 3 3 | reproduction_recipe stringlengths 90 176 | core_problem_addressed stringlengths 62 454 | key_technical_innovation stringlengths 82 615 | key_quantitative_result stringlengths 39 423 | cluster_id int64 1 8 | cluster_topic_name stringclasses 8
values | trend_velocity_tier stringclasses 1
value | papers_last_30_days int64 232 232 | audited_at stringdate 2026-08-21 14:35:24 2026-08-21 14:41:50 |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
2605.18401v2 | SkillsVote: Lifecycle Governance of Agent Skills from Collection, Recommendation to Evolution | 2026-05-18T13:44:19Z | [
"cs.CL",
"cs.AI"
] | ArXiv Standard | http://arxiv.org/licenses/nonexclusive-distrib/1.0/ | true | 60 | Internal R&D Only (Copyleft or Academic Terms) | ArXiv Standard Distribution; Repository License Unspecified | 0 | 0 | 1 | Proposes Agent Skills offer a structured artifact for combining procedural guidance, executable res... to enhance autonomous code synthesis, achieving Experiments on Terminal-Bench 2.0 and SWE-Bench Pro show that SkillsVote improve.... | Hongyi Liu | 7 | [
"Hongyi Liu",
"Haoyan Yang",
"Tao Jiang",
"Bo Tang",
"Feiyu Xiong",
"Yuyu Luo",
"Zhiyu Li"
] | [
"Autonomous Software Engineering AI Lab"
] | http://arxiv.org/abs/2605.18401v2 | NOT_DETECTED | Not Applicable | [] | 0 | 0 | Not Applicable | Not Applicable | 140.86 | Long-horizon LLM agents generate traces that could become reusable experience, but raw trajectories are noisy, local, and hard to govern. Agent Skills offer a structured artifact for combining procedural guidance, executable resources, and applicability boundaries. Yet open skill ecosystems contain redundant, uneven, e... | [
-0.05816999822854996,
-0.06106400117278099,
-0.08495599776506424,
0.004093999974429607,
0.0069519998505711555,
0.04622599855065346,
-0.02043999917805195,
-0.03943600133061409,
-0.045768000185489655,
0.00534099992364645,
-0.008522000163793564,
0.04272200167179108,
-0.005563999991863966,
0.1... | [
-0.020181000232696533,
-0.08380500227212906,
-0.040803998708724976,
0.021090000867843628,
0.12466099858283997,
-0.03423300012946129,
-0.038763001561164856,
-0.03721199929714203,
-0.01526699960231781,
-0.024966999888420105,
-0.05469899997115135,
-0.04503199830651283,
0.025644000619649887,
0... | AI Code Generation, Autonomous SWE Agents & Program Synthesis | Autonomous Repo-Level SWE Agent (SWE-bench) | [
"Python"
] | GPT-4o / Codex Specialized Coding LLM | [
"SWE-bench (Autonomous Software Engineering)"
] | [
"Quantitative Program Synthesis & Pass@k Metric"
] | [
{
"benchmark_name": "SWE-bench (Autonomous Software Engineering)",
"metric": "pass@1 / resolved",
"score": "Quantitative Program Synthesis & Pass@k Metric"
}
] | [
{
"paper_id": "2605.25430v1",
"title": "CODESKILL: Learning Self-Evolving Skills for Coding Agents",
"cosine_sim": 0.685
},
{
"paper_id": "2506.11425v2",
"title": "Agent-RLVR: Training Software Engineering Agents via Guidance and Environment Re",
"cosine_sim": 0.6537
},
{
"paper_... | No public code repository attached. Paper provides formal program synthesis specification. | Long-horizon LLM agents generate traces that could become reusable experience, but raw trajectories are noisy, local, and hard to govern. | Agent Skills offer a structured artifact for combining procedural guidance, executable resources, and applicability boundaries. | Experiments on Terminal-Bench 2.0 and SWE-Bench Pro show that SkillsVote improves agent performance on challenging agentic coding benchmarks. | 1 | Autonomous Repo-Level Software Engineering Agents (SWE-bench) | Explosive (>50/mo) | 232 | 2026-08-21T14:36:55.066395 |
2604.05278v1 | Spec Kit Agents: Context-Grounded Agentic Workflows | 2026-04-07T00:26:49Z | [
"cs.SE",
"cs.AI",
"cs.MA"
] | ArXiv Standard | http://arxiv.org/licenses/nonexclusive-distrib/1.0/ | true | 60 | Internal R&D Only (Copyleft or Academic Terms) | ArXiv Standard Distribution; Repository License Unspecified | 0 | 0 | 1 | Proposes Spec Kit Agents, a multi-agent SDD pipeline (with PM and developer roles) that adds phase-... to enhance autonomous code synthesis, achieving We further evaluate the framework on SWE-bench Lite, where augmentation hooks im.... | Pardis Taghavi | 2 | [
"Pardis Taghavi",
"Santosh Bhavani"
] | [
"Autonomous Software Engineering AI Lab"
] | http://arxiv.org/abs/2604.05278v1 | VERIFIED_LIVE | https://github.com/github/spec-kit | [
"https://github.com/github/spec-kit"
] | 131,000 | 0 | 2026-08-21 | Unspecified | 136.91 | Spec-driven development (SDD) with AI coding agents provides a structured workflow, but agents often remain "context blind" in large, evolving repositories, leading to hallucinated APIs and architectural violations. We present Spec Kit Agents, a multi-agent SDD pipeline (with PM and developer roles) that adds phase-lev... | [
-0.017906000837683678,
-0.0004659999976865947,
-0.03569300100207329,
-0.062470000237226486,
0.0623060017824173,
-0.09127400070428848,
0.07774300128221512,
0.019391000270843506,
-0.06134599819779396,
-0.023907000198960304,
-0.034935999661684036,
-0.10392600297927856,
-0.0004290000069886446,
... | [
-0.040824998170137405,
-0.03677999973297119,
-0.015138999558985233,
-0.002188999904319644,
0.037714000791311264,
-0.09035100042819977,
-0.03119100071489811,
0.004739999771118164,
-0.03911599889397621,
-0.02028699964284897,
-0.1005370020866394,
-0.1059970036149025,
0.01862500049173832,
0.01... | AI Code Generation, Autonomous SWE Agents & Program Synthesis | Autonomous Repo-Level SWE Agent (SWE-bench) | [
"Multi-Language / Polyglot"
] | Specialized Autoregressive Code Transformer | [
"SWE-bench (Autonomous Software Engineering)"
] | [
"achieving 58.2"
] | [
{
"benchmark_name": "SWE-bench (Autonomous Software Engineering)",
"metric": "pass@1 / resolved",
"score": "achieving 58.2"
}
] | [
{
"paper_id": "2605.15226v1",
"title": "Is Agentic AI Ready for Real-World Hardware Engineering? A Deep Dive with Phoeni",
"cosine_sim": 0.7629
},
{
"paper_id": "2510.23761v2",
"title": "TDFlow: Agentic Workflows for Test Driven Development",
"cosine_sim": 0.751
},
{
"paper_id": ... | git clone https://github.com/github/spec-kit && cd spec-kit && (pip install -e . || pip install -r requirements.txt) | Spec-driven development (SDD) with AI coding agents provides a structured workflow, but agents often remain "context blind" in large, evolving repositories, leading to hallucinated APIs and architectural violations. | We present Spec Kit Agents, a multi-agent SDD pipeline (with PM and developer roles) that adds phase-level, context-grounding hooks. | We further evaluate the framework on SWE-bench Lite, where augmentation hooks improve baseline by 1.7 percent, achieving 58.2 percent Pass@1. | 8 | Code Foundation Models & Specialized Instruction Distillation | Explosive (>50/mo) | 232 | 2026-08-21T14:37:51.079701 |
2607.01929v1 | Beyond Textual Repository Exploration: Dual-Modal Structural Reasoning for Agentic Issue Resolution | 2026-07-02T09:23:48Z | [
"cs.SE"
] | ArXiv Standard | http://arxiv.org/licenses/nonexclusive-distrib/1.0/ | true | 80 | Enterprise Safe (Commercial Training & Deployment Allowed) | ArXiv Standard Distribution; Permissive Open-Source Software (MIT) | 0 | 0 | 1 | Proposes As issue-resolution agents traverse repositories through fragmented textual observations,... to enhance autonomous code synthesis, achieving Further ablation studies demonstrate that the gains arise not only from textual.... | Jiayi Zhang | 4 | [
"Jiayi Zhang",
"Kai Huang",
"Yang Liu",
"Chunyang Chen"
] | [
"Autonomous Software Engineering AI Lab"
] | http://arxiv.org/abs/2607.01929v1 | VERIFIED_LIVE | https://github.com/colbymchenry/codegraph | [
"https://github.com/colbymchenry/codegraph"
] | 67,528 | 4,285 | 2026-08-20 | MIT | 134.59 | Recent advances in agentic program repair have significantly improved issue resolution by enabling iterative repository exploration. However, existing approaches predominantly rely on sequential, text-based code navigation, which fundamentally limits their ability to reason over large-scale long-horizon repositories wi... | [
0.012072999961674213,
-0.0015310000162571669,
0.010522999800741673,
0.036600999534130096,
0.0860380008816719,
0.00418699998408556,
0.030119000002741814,
0.017093999311327934,
0.027442000806331635,
0.05183300003409386,
-0.017961999401450157,
-0.03453800082206726,
0.039413001388311386,
0.104... | [
-0.048778001219034195,
-0.06391800194978714,
-0.01829499937593937,
0.028967000544071198,
0.04979899898171425,
-0.05796299874782562,
-0.023615999147295952,
0.00711899995803833,
-0.015812000259757042,
0.012107999995350838,
-0.043776001781225204,
-0.005764000117778778,
0.035597000271081924,
0... | AI Code Generation, Autonomous SWE Agents & Program Synthesis | Autonomous Repo-Level SWE Agent (SWE-bench) | [
"Multi-Language / Polyglot"
] | Claude-3.5-Sonnet Agentic SWE Backbone | [
"SWE-bench (Autonomous Software Engineering)"
] | [
"Quantitative Program Synthesis & Pass@k Metric"
] | [
{
"benchmark_name": "SWE-bench (Autonomous Software Engineering)",
"metric": "pass@1 / resolved",
"score": "Quantitative Program Synthesis & Pass@k Metric"
}
] | [
{
"paper_id": "2605.03117v2",
"title": "ARISE: A Repository-level Graph Representation and Toolset for Agentic Program R",
"cosine_sim": 0.7671
},
{
"paper_id": "2605.16352v1",
"title": "LARGER: Lexically Anchored Repository Graph Exploration and Retrieval",
"cosine_sim": 0.7536
},
{... | git clone https://github.com/colbymchenry/codegraph && cd codegraph && (pip install -e . || pip install -r requirements.txt) | Recent advances in agentic program repair have significantly improved issue resolution by enabling iterative repository exploration. | As issue-resolution agents traverse repositories through fragmented textual observations, structural information such as module organization, call relationships, and dependency chains must be repeatedly reconstructed across interaction steps, often leading to exploration drift and incomplete localization. | Further ablation studies demonstrate that the gains arise not only from textual structural information but also from visual externalization of repository dependencies, which better supports long-horizon repository exploration. | 8 | Code Foundation Models & Specialized Instruction Distillation | Explosive (>50/mo) | 232 | 2026-08-21T14:36:02.079528 |
2602.22764v1 | Evaluating and Improving Automated Repository-Level Rust Issue Resolution with LLM-based Agents | 2026-02-26T08:54:09Z | [
"cs.SE"
] | ArXiv Standard | http://arxiv.org/licenses/nonexclusive-distrib/1.0/ | true | 60 | Internal R&D Only (Copyleft or Academic Terms) | ArXiv Standard Distribution; Repository License Unspecified | 0 | 0 | 1 | Proposes Recently, LLM-powered code agents have shown remarkable success in resolving complex softw... to enhance autonomous code synthesis, achieving The evaluation shows that RUSTFORGER using Claude-Sonnet-3.7 significantly outpe.... | Jiahong Xiang | 5 | [
"Jiahong Xiang",
"Wenxiao He",
"Xihua Wang",
"Hongliang Tian",
"Yuqun Zhang"
] | [
"Autonomous Software Engineering AI Lab"
] | http://arxiv.org/abs/2602.22764v1 | NOT_DETECTED | Not Applicable | [] | 0 | 0 | Not Applicable | Not Applicable | 134.25 | The Rust programming language presents a steep learning curve and significant coding challenges, making the automation of issue resolution essential for its broader adoption. Recently, LLM-powered code agents have shown remarkable success in resolving complex software engineering tasks, yet their application to Rust ha... | [
-0.05432000011205673,
-0.019053999334573746,
0.03193800151348114,
0.03148199990391731,
0.05669400095939636,
-0.056724000722169876,
-0.05136200040578842,
-0.0025319999549537897,
-0.044700998812913895,
0.0015030000358819962,
-0.04782399907708168,
-0.010869000107049942,
0.09177800267934799,
-... | [
-0.08384200185537338,
-0.029462000355124474,
0.007918000221252441,
0.0595409981906414,
0.05516500025987625,
-0.08093400299549103,
-0.044084999710321426,
0.026537999510765076,
-0.028132999315857887,
-0.02605300024151802,
-0.10503000020980835,
-0.07940799742937088,
0.0939669981598854,
-0.000... | AI Code Generation, Autonomous SWE Agents & Program Synthesis | Autonomous Repo-Level SWE Agent (SWE-bench) | [
"Rust"
] | Claude-3.5-Sonnet Agentic SWE Backbone | [
"SWE-bench (Autonomous Software Engineering)"
] | [
"Quantitative Program Synthesis & Pass@k Metric"
] | [
{
"benchmark_name": "SWE-bench (Autonomous Software Engineering)",
"metric": "pass@1 / resolved",
"score": "Quantitative Program Synthesis & Pass@k Metric"
}
] | [
{
"paper_id": "2606.07297v1",
"title": "SWE-Explore: Benchmarking How Coding Agents Explore Repositories",
"cosine_sim": 0.755
},
{
"paper_id": "2604.11518v1",
"title": "From Translation to Superset: Benchmark-Driven Evolution of a Production AI Agen",
"cosine_sim": 0.755
},
{
"p... | No public code repository attached. Paper provides formal program synthesis specification. | The Rust programming language presents a steep learning curve and significant coding challenges, making the automation of issue resolution essential for its broader adoption. | Recently, LLM-powered code agents have shown remarkable success in resolving complex software engineering tasks, yet their application to Rust has been limited by the absence of a large-scale, repository-level benchmark. | The evaluation shows that RUSTFORGER using Claude-Sonnet-3.7 significantly outperforms all baselines, resolving 28.6% of tasks on Rust-SWE-bench, i.e., a 34.9% improvement over the strongest baseline, and, in aggregate, uniquely solves 46 tasks that no other agent could solve across all adopted advanced LLMs. | 8 | Code Foundation Models & Specialized Instruction Distillation | Explosive (>50/mo) | 232 | 2026-08-21T14:38:31.236278 |
2605.09051v1 | ParityFuzz: Finding Inconsistencies across Solidity Compilers via Fine-Grained Mutation and Differential Analysis | 2026-05-09T16:53:45Z | [
"cs.SE"
] | ArXiv Standard | http://arxiv.org/licenses/nonexclusive-distrib/1.0/ | true | 60 | Internal R&D Only (Copyleft or Academic Terms) | ArXiv Standard Distribution; Repository License Unspecified | 0 | 0 | 1 | Proposes These inconsistencies hinder contract migration, mislead developers during debugging, and... to enhance autonomous code synthesis, achieving It achieves up to 18x higher compilation success rate and 1.8x higher code cover.... | Bowei Su | 5 | [
"Bowei Su",
"Mingxi Ye",
"Yuhong Na",
"Peilin Zheng",
"Zibin Zheng"
] | [
"Autonomous Software Engineering AI Lab"
] | http://arxiv.org/abs/2605.09051v1 | NOT_DETECTED | Not Applicable | [] | 0 | 0 | Not Applicable | Not Applicable | 124.04 | The Solidity smart contract ecosystem has rapidly grown, leading to multiple compilers targeting different blockchain platforms or improving compilation efficiency. Although many compilers aim to be compatible with the primary Solidity compiler (Solc), significant inconsistencies in compilation and execution remain. Th... | [
-0.11066000163555145,
-0.0005329999839887023,
0.029309000819921494,
0.005584000144153833,
-0.022265000268816948,
-0.11751700192689896,
-0.04494199901819229,
-0.018334999680519104,
-0.07810500264167786,
-0.05143500119447708,
-0.02896299958229065,
-0.07300999760627747,
-0.008156999945640564,
... | [
-0.08967100083827972,
0.012349000200629234,
0.0017399999778717756,
0.009242000058293343,
-0.005783999804407358,
-0.09761399775743484,
-0.05167999863624573,
-0.00254100002348423,
-0.08862199634313583,
-0.040525998920202255,
-0.048374999314546585,
-0.11144399642944336,
0.0034680000972002745,
... | AI Code Generation, Autonomous SWE Agents & Program Synthesis | Code Translation & Transpilation Agent | [
"Solidity / Smart Contracts"
] | CodeQwen / Qwen2.5-Coder Foundation | [
"HumanEval & SWE-bench Functional Synthesis Benchmark"
] | [
"Quantitative Program Synthesis & Pass@k Metric"
] | [
{
"benchmark_name": "HumanEval & SWE-bench Functional Synthesis Benchmark",
"metric": "pass@1 / resolved",
"score": "Quantitative Program Synthesis & Pass@k Metric"
}
] | [
{
"paper_id": "2607.07217v1",
"title": "Finding and Understanding Miscompilation Bugs in the Solidity Compiler",
"cosine_sim": 0.842
},
{
"paper_id": "2607.15762v1",
"title": "GapForge: Directed Compiler Fuzzing via Coverage-Gap Analysis",
"cosine_sim": 0.6979
},
{
"paper_id": "2... | No public code repository attached. Paper provides formal program synthesis specification. | The Solidity smart contract ecosystem has rapidly grown, leading to multiple compilers targeting different blockchain platforms or improving compilation efficiency. | These inconsistencies hinder contract migration, mislead developers during debugging, and may introduce exploitable vulnerabilities, causing financial losses. | It achieves up to 18x higher compilation success rate and 1.8x higher code coverage than state-of-the-art fuzzers. | 7 | Syntax-Guided AST Modeling & Semantic Code Search | Explosive (>50/mo) | 232 | 2026-08-21T14:37:09.664025 |
2511.13761v1 | What happens when nanochat meets DiLoCo? | 2025-11-14T10:30:04Z | [
"cs.DC",
"cs.AI",
"cs.LG"
] | ArXiv Standard | http://arxiv.org/licenses/nonexclusive-distrib/1.0/ | true | 60 | Internal R&D Only (Copyleft or Academic Terms) | ArXiv Standard Distribution; Repository License Unspecified | 0 | 0 | 1 | Proposes The model trade-offs introduced by this shift remain underexplored, and our goal is to stu... to enhance autonomous code synthesis, achieving DiLoCo achieves stable convergence and competitive loss in pretraining but yield.... | Alexander Acker | 6 | [
"Alexander Acker",
"Soeren Becker",
"Sasho Nedelkoski",
"Dominik Scheinert",
"Odej Kao",
"Philipp Wiesner"
] | [
"Autonomous Software Engineering AI Lab"
] | http://arxiv.org/abs/2511.13761v1 | VERIFIED_LIVE | https://github.com/karpathy/nanochat | [
"https://github.com/karpathy/nanochat"
] | 57,400 | 0 | 2026-08-21 | Unspecified | 123.98 | Although LLM training is typically centralized with high-bandwidth interconnects and large compute budgets, emerging methods target communication-constrained training in distributed environments. The model trade-offs introduced by this shift remain underexplored, and our goal is to study them. We use the open-source na... | [
0.0245789997279644,
-0.09005799889564514,
0.00871300045400858,
0.03439300134778023,
0.01057399995625019,
-0.11997199803590775,
-0.0170499999076128,
0.07791300117969513,
0.0061599998734891415,
0.02568100020289421,
0.12115699797868729,
-0.024991000071167946,
-0.07395900040864944,
0.019495999... | [
-0.014174000360071659,
-0.05802199989557266,
0.05764700099825859,
0.047540999948978424,
-0.017190000042319298,
-0.12149500101804733,
-0.07197300344705582,
0.04043800011277199,
-0.05829000100493431,
0.0030129998922348022,
0.02053700014948845,
-0.05826300010085106,
-0.0502999983727932,
-0.04... | AI Code Generation, Autonomous SWE Agents & Program Synthesis | Neural Program Synthesis & Generation Architecture | [
"Multi-Language / Polyglot"
] | Specialized Autoregressive Code Transformer | [
"HumanEval (Functional Python Correctness)"
] | [
"Quantitative Program Synthesis & Pass@k Metric"
] | [
{
"benchmark_name": "HumanEval (Functional Python Correctness)",
"metric": "pass@1 / resolved",
"score": "Quantitative Program Synthesis & Pass@k Metric"
}
] | [
{
"paper_id": "2511.11608v1",
"title": "Why Should the Server Do It All?: A Scalable, Versatile, and Model-Agnostic Fram",
"cosine_sim": 0.5515
},
{
"paper_id": "2602.08676v3",
"title": "LLaDA2.1: Speeding Up Text Diffusion via Token Editing",
"cosine_sim": 0.5234
},
{
"paper_id"... | git clone https://github.com/karpathy/nanochat && cd nanochat && (pip install -e . || pip install -r requirements.txt) | Although LLM training is typically centralized with high-bandwidth interconnects and large compute budgets, emerging methods target communication-constrained training in distributed environments. | The model trade-offs introduced by this shift remain underexplored, and our goal is to study them. | DiLoCo achieves stable convergence and competitive loss in pretraining but yields worse MMLU, GSM8K, and HumanEval scores after mid-training and SFT. | 4 | Formal Verification, Theorem Proving & Symbolic Logic | Explosive (>50/mo) | 232 | 2026-08-21T14:39:48.981685 |
2606.02963v1 | KForge: LLM-Driven Cross-Platform Kernel Generation for AI Accelerators | 2026-06-01T23:48:55Z | [
"cs.LG"
] | ArXiv Standard | http://arxiv.org/licenses/nonexclusive-distrib/1.0/ | true | 60 | Internal R&D Only (Copyleft or Academic Terms) | ArXiv Standard Distribution; Repository License Unspecified | 0 | 0 | 1 | Proposes Agentic pipelines interleave reasoning, tool calls, and multi-agent coordination, each wit... to enhance autonomous code synthesis, achieving On NVIDIA B200, KForge achieves a 2.12$\%$ improvement in end-to-end throughput.... | Taras Sereda | 6 | [
"Taras Sereda",
"Burak Bartan",
"Ankita Nayak",
"Tom St. John",
"Natalie Serrino",
"Zain Asgar"
] | [
"Autonomous Software Engineering AI Lab"
] | http://arxiv.org/abs/2606.02963v1 | NOT_DETECTED | Not Applicable | [] | 0 | 0 | Not Applicable | Not Applicable | 119.93 | Production inference increasingly targets a heterogeneous mix of accelerators. Agentic pipelines interleave reasoning, tool calls, and multi-agent coordination, each with distinct compute and memory profiles. For optimal efficiency, each stage should run on the accelerator best suited to it. This creates a systems chal... | [
-0.03654899820685387,
-0.06515999883413315,
-0.028748000040650368,
0.03246799856424332,
0.06185000017285347,
-0.09104599803686142,
-0.0603020004928112,
-0.04893599823117256,
-0.04920800030231476,
-0.0842209979891777,
-0.04712500050663948,
-0.13733400404453278,
-0.009084999561309814,
-0.083... | [
-0.04515400156378746,
-0.07560399919748306,
-0.06313599646091461,
0.04154599830508232,
0.027215000241994858,
-0.0899370014667511,
-0.03132899850606918,
0.017210999503731728,
-0.029625000432133675,
-0.07179299741983414,
-0.09405100345611572,
-0.12672999501228333,
0.01032199990004301,
0.0030... | AI Code Generation, Autonomous SWE Agents & Program Synthesis | Neural Program Synthesis & Generation Architecture | [
"Multi-Language / Polyglot"
] | Specialized Autoregressive Code Transformer | [
"HumanEval & SWE-bench Functional Synthesis Benchmark"
] | [
"achieves a 2.12",
"achieving a 5.13"
] | [
{
"benchmark_name": "HumanEval & SWE-bench Functional Synthesis Benchmark",
"metric": "pass@1 / resolved",
"score": "achieves a 2.12"
}
] | [
{
"paper_id": "2511.13274v1",
"title": "KForge: Program Synthesis for Diverse AI Hardware Accelerators",
"cosine_sim": 0.8613
},
{
"paper_id": "2510.16996v1",
"title": "STARK: Strategic Team of Agents for Refining Kernels",
"cosine_sim": 0.818
},
{
"paper_id": "2603.08721v2",
... | No public code repository attached. Paper provides formal program synthesis specification. | Production inference increasingly targets a heterogeneous mix of accelerators. | Agentic pipelines interleave reasoning, tool calls, and multi-agent coordination, each with distinct compute and memory profiles. | On NVIDIA B200, KForge achieves a 2.12$\%$ improvement in end-to-end throughput compared to TensorRT-LLM on the gpt-oss-20b inference speed benchmark. | 4 | Formal Verification, Theorem Proving & Symbolic Logic | Explosive (>50/mo) | 232 | 2026-08-21T14:36:33.131487 |
2603.00729v1 | Qwen3-Coder-Next Technical Report | 2026-02-28T16:25:04Z | [
"cs.CL"
] | ArXiv Standard | http://arxiv.org/licenses/nonexclusive-distrib/1.0/ | true | 60 | Internal R&D Only (Copyleft or Academic Terms) | ArXiv Standard Distribution; Repository License Unspecified | 0 | 0 | 1 | Proposes To achieve this, we perform agentic training through large-scale synthesis of verifiable c... to enhance autonomous code synthesis, achieving Across agent-centric benchmarks including SWE-Bench and Terminal-Bench, Qwen3-Co.... | Ruisheng Cao | 20 | [
"Ruisheng Cao",
"Mouxiang Chen",
"Jiawei Chen",
"Zeyu Cui",
"Yunlong Feng",
"Binyuan Hui",
"Yuheng Jing",
"Kaixin Li",
"Mingze Li",
"Junyang Lin",
"Zeyao Ma",
"Kashun Shum",
"Xuwu Wang",
"Jinxi Wei",
"Jiaxi Yang",
"Jiajun Zhang",
"Lei Zhang",
"Zongmeng Zhang",
"Wenting Zhao",
"... | [
"Autonomous Software Engineering AI Lab"
] | http://arxiv.org/abs/2603.00729v1 | VERIFIED_LIVE | https://github.com/QwenLM/Qwen3-Coder | [
"https://github.com/QwenLM/Qwen3-Coder"
] | 16,800 | 0 | 2026-08-21 | Unspecified | 117.55 | We present Qwen3-Coder-Next, an open-weight language model specialized for coding agents. Qwen3-Coder-Next is an 80-billion-parameter model that activates only 3 billion parameters during inference, enabling strong coding capability with efficient inference. In this work, we explore how far strong training recipes can ... | [
-0.1630139946937561,
-0.02379699982702732,
-0.03865800052881241,
-0.042472999542951584,
-0.003106999909505248,
0.025072000920772552,
-0.027929000556468964,
0.00444599986076355,
-0.011842000298202038,
0.05302400141954422,
0.07548099756240845,
-0.04769299924373627,
0.03791100159287453,
0.016... | [
-0.08927900344133377,
-0.0901539996266365,
-0.04215899854898453,
0.01765899918973446,
0.01207599975168705,
-0.011757000349462032,
-0.04595300182700157,
-0.04198500141501427,
-0.00380500010214746,
-0.0008329999982379377,
-0.03073900006711483,
-0.11849399656057358,
0.05341099947690964,
0.030... | AI Code Generation, Autonomous SWE Agents & Program Synthesis | Autonomous Repo-Level SWE Agent (SWE-bench) | [
"Multi-Language / Polyglot"
] | Claude-3.5-Sonnet Agentic SWE Backbone | [
"SWE-bench (Autonomous Software Engineering)"
] | [
"Quantitative Program Synthesis & Pass@k Metric"
] | [
{
"benchmark_name": "SWE-bench (Autonomous Software Engineering)",
"metric": "pass@1 / resolved",
"score": "Quantitative Program Synthesis & Pass@k Metric"
}
] | [
{
"paper_id": "2506.03136v2",
"title": "Co-Evolving LLM Coder and Unit Tester via Reinforcement Learning",
"cosine_sim": 0.7544
},
{
"paper_id": "2607.05471v1",
"title": "KAT-Coder-V2.5 Technical Report",
"cosine_sim": 0.7481
},
{
"paper_id": "2608.17393v1",
"title": "LEGO-RL... | git clone https://github.com/QwenLM/Qwen3-Coder && cd Qwen3-Coder && (pip install -e . || pip install -r requirements.txt) | We present Qwen3-Coder-Next, an open-weight language model specialized for coding agents. | To achieve this, we perform agentic training through large-scale synthesis of verifiable coding tasks paired with executable environments, allowing learning directly from environment feedback via mid-training and reinforcement learning. | Across agent-centric benchmarks including SWE-Bench and Terminal-Bench, Qwen3-Coder-Next achieves competitive performance relative to its active parameter count. | 5 | Interactive Coding Assistants & Human-AI Pair Programming | Explosive (>50/mo) | 232 | 2026-08-21T14:38:28.350988 |
2603.06107v1 | Real-World Fault Detection for C-Extended Python Projects with Automated Unit Test Generation | 2026-03-06T10:05:29Z | [
"cs.SE"
] | ArXiv Standard | http://arxiv.org/licenses/nonexclusive-distrib/1.0/ | true | 60 | Internal R&D Only (Copyleft or Academic Terms) | ArXiv Standard Distribution; Repository License Unspecified | 0 | 0 | 1 | Proposes To overcome this problem, we propose separating the generation and execution stages of the... to enhance autonomous code synthesis, achieving Subprocess-execution allowed automated testing of up to 56.5% more modules and d.... | Lucas Berg | 7 | [
"Lucas Berg",
"Lukas Krodinger",
"Stephan Lukasczyk",
"Annibale Panichella",
"Gordon Fraser",
"Wim Vanhoof",
"Xavier Devroey"
] | [
"Autonomous Software Engineering AI Lab"
] | http://arxiv.org/abs/2603.06107v1 | NOT_DETECTED | Not Applicable | [] | 0 | 0 | Not Applicable | Not Applicable | 116.92 | Many popular Python libraries use C-extensions for performance-critical operations allowing users to combine the best of the two worlds: The simplicity and versatility of Python and the performance of C. A drawback of this approach is that exceptions raised in C can bypass Python's exception handling and cause the enti... | [
-0.09296099841594696,
-0.01591699942946434,
-0.021114999428391457,
0.07172200083732605,
0.035085998475551605,
-0.13524700701236725,
-0.07286900281906128,
0.02332100085914135,
-0.06964100152254105,
-0.0204050000756979,
-0.004652000032365322,
-0.08622299879789352,
0.03316900134086609,
-0.011... | [
-0.10700400173664093,
0.029191000387072563,
-0.039521001279354095,
0.05558300018310547,
0.008333999663591385,
-0.1200999990105629,
-0.07753700017929077,
0.03271099925041199,
-0.04572699964046478,
-0.023290999233722687,
-0.03943299874663353,
-0.08316399902105331,
0.04874400049448013,
0.0040... | AI Code Generation, Autonomous SWE Agents & Program Synthesis | Neural Program Synthesis & Generation Architecture | [
"Python",
"Java / Kotlin"
] | Specialized Autoregressive Code Transformer | [
"HumanEval & SWE-bench Functional Synthesis Benchmark"
] | [
"Quantitative Program Synthesis & Pass@k Metric"
] | [
{
"benchmark_name": "HumanEval & SWE-bench Functional Synthesis Benchmark",
"metric": "pass@1 / resolved",
"score": "Quantitative Program Synthesis & Pass@k Metric"
}
] | [
{
"paper_id": "2507.01477v2",
"title": "Combining Type Inference and Automated Unit Test Generation for Python",
"cosine_sim": 0.7048
},
{
"paper_id": "2606.08588v1",
"title": "LLM vs. Human Unit Tests: Fault Detection on Real Python Bugs",
"cosine_sim": 0.6869
},
{
"paper_id": "... | No public code repository attached. Paper provides formal program synthesis specification. | Many popular Python libraries use C-extensions for performance-critical operations allowing users to combine the best of the two worlds: The simplicity and versatility of Python and the performance of C. | To overcome this problem, we propose separating the generation and execution stages of the test-generation process. | Subprocess-execution allowed automated testing of up to 56.5% more modules and discovered 213 unique crash causes, revealing 32 previously unknown faults. | 7 | Syntax-Guided AST Modeling & Semantic Code Search | Explosive (>50/mo) | 232 | 2026-08-21T14:38:22.082864 |
2605.03956v1 | Generating Proof-of-Vulnerability Tests to Help Enhance the Security of Complex Software | 2026-05-05T16:39:29Z | [
"cs.CR",
"cs.SE"
] | ArXiv Standard | http://arxiv.org/licenses/nonexclusive-distrib/1.0/ | true | 60 | Internal R&D Only (Copyleft or Academic Terms) | ArXiv Standard Distribution; Repository License Unspecified | 0 | 0 | 1 | Proposes Prior work shows that developers often require concrete and executable evidence, i.e., pro... to enhance autonomous code synthesis, achieving PoVSmith substantially outperforms the state-of-the-art LLM-based approach, as i.... | Shravya Kanchi | 5 | [
"Shravya Kanchi",
"Xiaoyan Zang",
"Ying Zhang",
"Danfeng Yao",
"Na Meng"
] | [
"Autonomous Software Engineering AI Lab"
] | http://arxiv.org/abs/2605.03956v1 | NOT_DETECTED | Not Applicable | [] | 0 | 0 | Not Applicable | Not Applicable | 115.33 | Developers create modern software applications (Apps) on top of third-party libraries (Libs). When library vulnerabilities are reachable through application code, the applications can be vulnerable to software supply chain attacks. Prior work shows that developers often require concrete and executable evidence, i.e., p... | [
-0.08836500346660614,
0.06098699942231178,
-0.03901999816298485,
0.01435100007802248,
0.03838599845767021,
-0.016289999708533287,
-0.028596000745892525,
0.053300999104976654,
-0.010355999693274498,
-0.012702999636530876,
-0.054976001381874084,
-0.10575299710035324,
0.12424000352621078,
-0.... | [
-0.11676599830389023,
-0.0031500000040978193,
-0.05471700057387352,
0.00483600003644824,
0.029249999672174454,
-0.03622199967503548,
-0.0037750001065433025,
0.06828100234270096,
-0.0012019999558106065,
-0.04670200124382973,
-0.026089999824762344,
-0.10912100225687027,
0.05251999944448471,
... | AI Code Generation, Autonomous SWE Agents & Program Synthesis | Static Analysis & Vulnerability Remediation | [
"Java / Kotlin"
] | GPT-4o / Codex Specialized Coding LLM | [
"HumanEval & SWE-bench Functional Synthesis Benchmark"
] | [
"Quantitative Program Synthesis & Pass@k Metric"
] | [
{
"benchmark_name": "HumanEval & SWE-bench Functional Synthesis Benchmark",
"metric": "pass@1 / resolved",
"score": "Quantitative Program Synthesis & Pass@k Metric"
}
] | [
{
"paper_id": "2507.15241v1",
"title": "FaultLine: Automated Proof-of-Vulnerability Generation Using LLM Agents",
"cosine_sim": 0.7854
},
{
"paper_id": "2506.11559v1",
"title": "Leveraging GPT-4 for Vulnerability-Witnessing Unit Test Generation",
"cosine_sim": 0.7405
},
{
"paper_... | No public code repository attached. Paper provides formal program synthesis specification. | Developers create modern software applications (Apps) on top of third-party libraries (Libs). | Prior work shows that developers often require concrete and executable evidence, i.e., proof-of-vulnerability (PoV) tests, to decide whether a reported dependency vulnerability poses a practical security risk to their application. | PoVSmith substantially outperforms the state-of-the-art LLM-based approach, as it reduces human involvement while dramatically improving test quality. | 7 | Syntax-Guided AST Modeling & Semantic Code Search | Explosive (>50/mo) | 232 | 2026-08-21T14:37:12.539413 |
2510.14455v1 | Coder as Editor: Code-driven Interpretable Molecular Optimization | 2025-10-16T08:55:06Z | [
"cs.LG",
"q-bio.BM"
] | ArXiv Standard | http://arxiv.org/licenses/nonexclusive-distrib/1.0/ | true | 60 | Internal R&D Only (Copyleft or Academic Terms) | ArXiv Standard Distribution; Repository License Unspecified | 0 | 0 | 1 | Proposes While large language models (LLMs) have shown promise in generating high-level editing int... to enhance autonomous code synthesis, achieving On downstream optimization benchmarks spanning physicochemical properties and ta.... | Wenyu Zhu | 10 | [
"Wenyu Zhu",
"Chengzhu Li",
"Xiaohe Tian",
"Yifan Wang",
"Yinjun Jia",
"Jianhui Wang",
"Bowen Gao",
"Ya-Qin Zhang",
"Wei-Ying Ma",
"Yanyan Lan"
] | [
"Autonomous Software Engineering AI Lab"
] | http://arxiv.org/abs/2510.14455v1 | VERIFIED_LIVE | https://github.com/QwenLM/Qwen2.5-Coder | [
"https://github.com/QwenLM/Qwen2.5-Coder"
] | 16,800 | 0 | 2026-08-21 | Unspecified | 112.15 | Molecular optimization is a central task in drug discovery that requires precise structural reasoning and domain knowledge. While large language models (LLMs) have shown promise in generating high-level editing intentions in natural language, they often struggle to faithfully execute these modifications-particularly wh... | [
-0.08746399730443954,
-0.02586900070309639,
-0.0534450002014637,
0.010858999565243721,
-0.002583000110462308,
-0.06214499846100807,
-0.05366099998354912,
0.0214649997651577,
-0.03148899972438812,
0.0024790000170469284,
-0.03720799833536148,
-0.04629199951887131,
-0.054016999900341034,
0.02... | [
-0.05363300070166588,
-0.04652300104498863,
0.018188999965786934,
0.02793700061738491,
0.012728000059723854,
-0.08590400218963623,
-0.0569939985871315,
0.04344499856233597,
0.03202800080180168,
0.012821000069379807,
-0.07873900234699249,
-0.07241500169038773,
0.008887999691069126,
0.044146... | AI Code Generation, Autonomous SWE Agents & Program Synthesis | Neural Program Synthesis & Generation Architecture | [
"Multi-Language / Polyglot"
] | CodeQwen / Qwen2.5-Coder Foundation | [
"HumanEval & SWE-bench Functional Synthesis Benchmark"
] | [
"98% accuracy"
] | [
{
"benchmark_name": "HumanEval & SWE-bench Functional Synthesis Benchmark",
"metric": "pass@1 / resolved",
"score": "98% accuracy"
}
] | [
{
"paper_id": "2605.02351v2",
"title": "MolViBench: Evaluating LLMs on Molecular Vibe Coding",
"cosine_sim": 0.7663
},
{
"paper_id": "2601.02075v4",
"title": "MDAgent2: Large Language Model for Code Generation and Knowledge Q&A in Molecula",
"cosine_sim": 0.659
},
{
"paper_id": "... | git clone https://github.com/QwenLM/Qwen2.5-Coder && cd Qwen2.5-Coder && (pip install -e . || pip install -r requirements.txt) | Molecular optimization is a central task in drug discovery that requires precise structural reasoning and domain knowledge. | While large language models (LLMs) have shown promise in generating high-level editing intentions in natural language, they often struggle to faithfully execute these modifications-particularly when operating on non-intuitive representations like SMILES. | On downstream optimization benchmarks spanning physicochemical properties and target activities, MECo substantially improves consistency by 38-86 percentage points to 90%+ and achieves higher success rates over SMILES-based baselines while preserving structural similarity. | 4 | Formal Verification, Theorem Proving & Symbolic Logic | Explosive (>50/mo) | 232 | 2026-08-21T14:40:11.751486 |
2604.01483v1 | Type-Checked Compliance: Deterministic Guardrails for Agentic Financial Systems Using Lean 4 Theorem Proving | 2026-04-01T23:39:43Z | [
"cs.LO",
"cs.AI",
"cs.CR"
] | ArXiv Standard | http://arxiv.org/licenses/nonexclusive-distrib/1.0/ | true | 60 | Internal R&D Only (Copyleft or Academic Terms) | ArXiv Standard Distribution; Repository License Unspecified | 0 | 0 | 1 | Proposes This paper presents the Lean-Agent Protocol, a formal-verification-based AI guardrail plat... to enhance autonomous code synthesis, achieving A three-phase implementation roadmap from shadow verification through enterprise.... | Devakh Rashie | 2 | [
"Devakh Rashie",
"Veda Rashi"
] | [
"Autonomous Software Engineering AI Lab"
] | http://arxiv.org/abs/2604.01483v1 | NOT_DETECTED | Not Applicable | [] | 0 | 0 | Not Applicable | Not Applicable | 111.22 | The rapid evolution of autonomous, agentic artificial intelligence within financial services has introduced an existential architectural crisis: large language models (LLMs) are probabilistic, non-deterministic systems operating in domains that demand absolute, mathematically verifiable compliance guarantees. Existing ... | [
-0.0012829999905079603,
0.027456000447273254,
-0.0697299987077713,
-0.03126399964094162,
-0.003708000062033534,
0.0357000008225441,
0.09757100045681,
-0.024515999481081963,
-0.03214000165462494,
0.0728290006518364,
0.013379000127315521,
0.007734000217169523,
0.034651998430490494,
0.0320100... | [
-0.05377199873328209,
-0.006027999799698591,
-0.05203599855303764,
-0.06374599784612656,
-0.04005499929189682,
0.0011790000135079026,
0.0049149999395012856,
-0.017462000250816345,
-0.0053820000030100346,
0.05393100157380104,
-0.01919100061058998,
-0.03683000057935715,
0.03530300036072731,
... | AI Code Generation, Autonomous SWE Agents & Program Synthesis | Formal Logic & Theorem Proving Synthesis | [
"Multi-Language / Polyglot"
] | Specialized Autoregressive Code Transformer | [
"HumanEval & SWE-bench Functional Synthesis Benchmark"
] | [
"Quantitative Program Synthesis & Pass@k Metric"
] | [
{
"benchmark_name": "HumanEval & SWE-bench Functional Synthesis Benchmark",
"metric": "pass@1 / resolved",
"score": "Quantitative Program Synthesis & Pass@k Metric"
}
] | [
{
"paper_id": "2602.11136v2",
"title": "FormalJudge: A Neuro-Symbolic Paradigm for Agentic Oversight",
"cosine_sim": 0.6778
},
{
"paper_id": "2606.06523v2",
"title": "Lean4Agent: Formal Modeling and Verification for Agent Workflow and Trajectory",
"cosine_sim": 0.6578
},
{
"paper... | No public code repository attached. Paper provides formal program synthesis specification. | The rapid evolution of autonomous, agentic artificial intelligence within financial services has introduced an existential architectural crisis: large language models (LLMs) are probabilistic, non-deterministic systems operating in domains that demand absolute, mathematically verifiable compliance guarantees. | This paper presents the Lean-Agent Protocol, a formal-verification-based AI guardrail platform that leverages the Aristotle neural-symbolic model developed by Harmonic AI to auto-formalize institutional policies into Lean 4 code. | A three-phase implementation roadmap from shadow verification through enterprise-scale deployment is provided. | 6 | Static Analysis, Vulnerability Detection & Automated Security Patching | Explosive (>50/mo) | 232 | 2026-08-21T14:38:00.325487 |
2605.26851v1 | LLM-based Mockless Unit Test Generation for Java | 2026-05-26T11:08:04Z | [
"cs.SE"
] | ArXiv Standard | http://arxiv.org/licenses/nonexclusive-distrib/1.0/ | true | 60 | Internal R&D Only (Copyleft or Academic Terms) | ArXiv Standard Distribution; Repository License Unspecified | 0 | 0 | 1 | Proposes MocklessTester, a mockless unit test generation approach built around two strategies: cont... to enhance autonomous code synthesis, achieving Ablation results confirm that all major components contribute positively to the.... | Qinghua Xu | 5 | [
"Qinghua Xu",
"Guancheng Wang",
"Lionel Briand",
"Zhaoqiang Guo",
"Kui Liu"
] | [
"Autonomous Software Engineering AI Lab"
] | http://arxiv.org/abs/2605.26851v1 | VERIFIED_LIVE | https://github.com/agentscope-ai/agentscope-java | [
"https://github.com/agentscope-ai/agentscope-java",
"https://github.com/google/adk-java",
"https://github.com/a2aproject/a2a-java"
] | 5,200 | 0 | 2026-08-21 | Unspecified | 110.84 | Large language models (LLMs) have shown strong potential for automated test generation, yet most approaches to generating Java unit tests still rely on mocking frameworks to handle dependencies. Mockless test generation could exercise more real low-level code, but it faces challenges such as invalid test code generatio... | [
-0.07035800069570541,
0.015937000513076782,
-0.02673099935054779,
-0.014090999960899353,
-0.04995200037956238,
-0.11564499884843826,
-0.06239200010895729,
0.00023799999326001853,
-0.0885540023446083,
-0.023778999224305153,
-0.0021840000990778208,
-0.06192699819803238,
0.0897739976644516,
-... | [
-0.048854999244213104,
0.02302500046789646,
0.01244099996984005,
0.0016749999485909939,
-0.012369999662041664,
-0.06058099865913391,
-0.07020100206136703,
-0.05055699869990349,
-0.09445399791002274,
-0.01027700025588274,
-0.018984999507665634,
-0.09879700094461441,
0.08446399867534637,
0.0... | AI Code Generation, Autonomous SWE Agents & Program Synthesis | Neural Program Synthesis & Generation Architecture | [
"Java / Kotlin"
] | Specialized Autoregressive Code Transformer | [
"Defects4J (Automated Program Repair Benchmark)"
] | [
"Quantitative Program Synthesis & Pass@k Metric"
] | [
{
"benchmark_name": "Defects4J (Automated Program Repair Benchmark)",
"metric": "pass@1 / resolved",
"score": "Quantitative Program Synthesis & Pass@k Metric"
}
] | [
{
"paper_id": "2604.19315v1",
"title": "Improving LLM-Driven Test Generation by Learning from Mocking Information",
"cosine_sim": 0.7725
},
{
"paper_id": "2509.23812v2",
"title": "Navigating the Labyrinth: Path-Sensitive Unit Test Generation with Large Languag",
"cosine_sim": 0.7691
},... | git clone https://github.com/agentscope-ai/agentscope-java && cd agentscope-java && (pip install -e . || pip install -r requirements.txt) | Large language models (LLMs) have shown strong potential for automated test generation, yet most approaches to generating Java unit tests still rely on mocking frameworks to handle dependencies. | We present MocklessTester, a mockless unit test generation approach built around two strategies: context-enriched generation and constraint-enforced fixing. | Ablation results confirm that all major components contribute positively to the final performance. | 7 | Syntax-Guided AST Modeling & Semantic Code Search | Explosive (>50/mo) | 232 | 2026-08-21T14:36:45.814628 |
2510.22210v2 | LSPRAG: LSP-Guided RAG for Language-Agnostic Real-Time Unit Test Generation | 2025-10-25T08:19:21Z | [
"cs.SE",
"cs.AI"
] | ArXiv Standard | http://arxiv.org/licenses/nonexclusive-distrib/1.0/ | true | 60 | Internal R&D Only (Copyleft or Academic Terms) | ArXiv Standard Distribution; Repository License Unspecified | 0 | 0 | 1 | Proposes To address this gap, we present LSPRAG, a framework for concise-context retrieval tailored... to enhance autonomous code synthesis, achieving Compared to the best performance of baselines, LSPRAG increased line coverage by.... | Gwihwan Go | 5 | [
"Gwihwan Go",
"Quan Zhang",
"Chijin Zhou",
"Zhao Wei",
"Yu Jiang"
] | [
"Autonomous Software Engineering AI Lab"
] | http://arxiv.org/abs/2510.22210v2 | NOT_DETECTED | Not Applicable | [] | 0 | 0 | Not Applicable | Not Applicable | 110.28 | Automated unit test generation is essential for robust software development, yet existing approaches struggle to generalize across multiple programming languages and operate within real-time development. While Large Language Models (LLMs) offer a promising solution, their ability to generate high coverage test code dep... | [
-0.09231100231409073,
-0.009336000308394432,
-0.00900499988347292,
0.03705799952149391,
-0.01613900065422058,
-0.0418270006775856,
-0.051920998841524124,
-0.016517000272870064,
-0.020176000893115997,
-0.03928099945187569,
0.03243900090456009,
-0.08161299675703049,
0.01166400033980608,
-0.0... | [
-0.09567400068044662,
-0.005266000051051378,
-0.024519000202417374,
0.02940800040960312,
0.04032700136303902,
-0.05492899939417839,
-0.030155999585986137,
0.035691000521183014,
-0.015776999294757843,
-0.07719700038433075,
0.03231099992990494,
-0.10382299870252609,
0.04523700103163719,
0.00... | AI Code Generation, Autonomous SWE Agents & Program Synthesis | Multi-Turn Interactive Pair Programmer | [
"Python",
"Go",
"Java / Kotlin"
] | Specialized Autoregressive Code Transformer | [
"HumanEval & SWE-bench Functional Synthesis Benchmark"
] | [
"Quantitative Program Synthesis & Pass@k Metric"
] | [
{
"benchmark_name": "HumanEval & SWE-bench Functional Synthesis Benchmark",
"metric": "pass@1 / resolved",
"score": "Quantitative Program Synthesis & Pass@k Metric"
}
] | [
{
"paper_id": "2602.21997v1",
"title": "Enhancing LLM-Based Test Generation by Eliminating Covered Code",
"cosine_sim": 0.7882
},
{
"paper_id": "2507.17271v1",
"title": "Seed&Steer: Guiding Large Language Models with Compilable Prefix and Branch Sign",
"cosine_sim": 0.7478
},
{
"... | No public code repository attached. Paper provides formal program synthesis specification. | Automated unit test generation is essential for robust software development, yet existing approaches struggle to generalize across multiple programming languages and operate within real-time development. | To address this gap, we present LSPRAG, a framework for concise-context retrieval tailored for real-time, language-agnostic unit test generation. | Compared to the best performance of baselines, LSPRAG increased line coverage by up to 174.55% for Golang, 213.31% for Java, and 31.57% for Python. | 7 | Syntax-Guided AST Modeling & Semantic Code Search | Explosive (>50/mo) | 232 | 2026-08-21T14:40:03.472172 |
2506.11781v1 | GeoPandas-AI: A Smart Class Bringing LLM as Stateful AI Code Assistant | 2025-06-13T13:42:17Z | [
"cs.HC",
"cs.SE"
] | ArXiv Standard | http://arxiv.org/licenses/nonexclusive-distrib/1.0/ | true | 60 | Internal R&D Only (Copyleft or Academic Terms) | ArXiv Standard Distribution; Repository License Unspecified | 0 | 0 | 1 | Proposes GeoPandas-AI addresses this gap by integrating LLMs directly into the GeoPandas workflow,... to enhance autonomous code synthesis, achieving Through its innovative combination of conversational interfaces and stateful exp.... | Gaspard Merten | 3 | [
"Gaspard Merten",
"Gilles Dejaegere",
"Mahmoud Sakr"
] | [
"Autonomous Software Engineering AI Lab"
] | http://arxiv.org/abs/2506.11781v1 | VERIFIED_LIVE | https://github.com/sinaptik-ai/pandas-ai | [
"https://github.com/sinaptik-ai/pandas-ai",
"https://github.com/GaspardMerten/geopandas-ai"
] | 23,800 | 0 | 2026-08-21 | Unspecified | 110.17 | Geospatial data analysis plays a crucial role in tackling intricate societal challenges such as urban planning and climate modeling. However, employing tools like GeoPandas, a prominent Python library for geospatial data manipulation, necessitates expertise in complex domain-specific syntax and workflows. GeoPandas-AI ... | [
-0.07660199701786041,
-0.05083199962973595,
-0.003613000037148595,
-0.013778000138700008,
0.02586599998176098,
-0.08126900345087051,
0.021052999421954155,
-0.04361499845981598,
-0.13087299466133118,
0.02115900069475174,
-0.03795799985527992,
-0.033601000905036926,
0.007476999890059233,
-0.... | [
-0.044881001114845276,
-0.028558999300003052,
0.007143000140786171,
0.0005389999714680016,
0.00927400030195713,
-0.15655100345611572,
0.004523999989032745,
-0.04337900131940842,
-0.09047599881887436,
0.019889000803232193,
-0.018001999706029892,
-0.03835900127887726,
0.008434999734163284,
0... | AI Code Generation, Autonomous SWE Agents & Program Synthesis | Multi-Turn Interactive Pair Programmer | [
"Python"
] | Specialized Autoregressive Code Transformer | [
"HumanEval & SWE-bench Functional Synthesis Benchmark"
] | [
"Quantitative Program Synthesis & Pass@k Metric"
] | [
{
"benchmark_name": "HumanEval & SWE-bench Functional Synthesis Benchmark",
"metric": "pass@1 / resolved",
"score": "Quantitative Program Synthesis & Pass@k Metric"
}
] | [
{
"paper_id": "2509.08863v3",
"title": "GeoJSON Agents:A Multi-Agent LLM Architecture for Geospatial Analysis-Function C",
"cosine_sim": 0.6755
},
{
"paper_id": "2511.20656v1",
"title": "Context-Aware Visual Prompting: Automating Geospatial Web Dashboards with Large ",
"cosine_sim": 0.66... | git clone https://github.com/sinaptik-ai/pandas-ai && cd pandas-ai && (pip install -e . || pip install -r requirements.txt) | Geospatial data analysis plays a crucial role in tackling intricate societal challenges such as urban planning and climate modeling. | GeoPandas-AI addresses this gap by integrating LLMs directly into the GeoPandas workflow, transforming the GeoDataFrame class into an intelligent, stateful class for both data analysis and geospatial code development. | Through its innovative combination of conversational interfaces and stateful exploitation of LLMs for code generation and data analysis, GeoPandas-AI introduces a new paradigm for code-copilots and instantiates it for geospatial development. | 3 | Multi-Language Code Translation & Cross-Platform Migration | Explosive (>50/mo) | 232 | 2026-08-21T14:41:42.761068 |
2603.07927v1 | SWE-Fuse: Empowering Software Agents via Issue-free Trajectory Learning and Entropy-aware RLVR Training | 2026-03-09T03:47:10Z | [
"cs.SE",
"cs.AI"
] | ArXiv Standard | http://arxiv.org/licenses/nonexclusive-distrib/1.0/ | true | 60 | Internal R&D Only (Copyleft or Academic Terms) | ArXiv Standard Distribution; Repository License Unspecified | 0 | 0 | 1 | Proposes Recently, numerous LLM-based agents have been developed to address real-world software iss... to enhance autonomous code synthesis, achieving Specifically, SWE-Fuse outperforms the best 8B and 32B baselines by 43.0\% and 6.... | Xin-Cheng Wen | 6 | [
"Xin-Cheng Wen",
"Binbin Chen",
"Haoxuan Lan",
"Hang Yu",
"Peng Di",
"Cuiyun Gao"
] | [
"Autonomous Software Engineering AI Lab"
] | http://arxiv.org/abs/2603.07927v1 | VERIFIED_LIVE | https://github.com/SWE-agent/Mini-SWE-Agent | [
"https://github.com/SWE-agent/Mini-SWE-Agent"
] | 6,700 | 0 | 2026-08-21 | Unspecified | 109.92 | Large language models (LLMs) have transformed the software engineering landscape. Recently, numerous LLM-based agents have been developed to address real-world software issue fixing tasks. Despite their state-of-the-art performance, Despite achieving state-of-the-art performance, these agents face a significant challen... | [
0.008169000037014484,
-0.08773399889469147,
-0.004641000181436539,
0.0073979999870061874,
0.10463900119066238,
0.009387999773025513,
-0.005611000116914511,
-0.013344000093638897,
0.053109001368284225,
-0.003539999946951866,
-0.027256999164819717,
-0.022777000442147255,
0.041179001331329346,
... | [
-0.017217999324202538,
-0.026840999722480774,
0.011500000022351742,
0.05406700074672699,
0.08587999641895294,
-0.03497299924492836,
0.02593100070953369,
-0.005632000043988228,
-0.00962700042873621,
-0.014876999892294407,
-0.04098299890756607,
-0.04198100045323372,
0.04476400092244148,
0.04... | AI Code Generation, Autonomous SWE Agents & Program Synthesis | Autonomous Repo-Level SWE Agent (SWE-bench) | [
"Multi-Language / Polyglot"
] | Claude-3.5-Sonnet Agentic SWE Backbone | [
"SWE-bench (Autonomous Software Engineering)"
] | [
"Quantitative Program Synthesis & Pass@k Metric"
] | [
{
"benchmark_name": "SWE-bench (Autonomous Software Engineering)",
"metric": "pass@1 / resolved",
"score": "Quantitative Program Synthesis & Pass@k Metric"
}
] | [
{
"paper_id": "2601.11655v1",
"title": "Advances and Frontiers of LLM-based Issue Resolution in Software Engineering: A ",
"cosine_sim": 0.7987
},
{
"paper_id": "2507.23361v2",
"title": "SWE-Exp: Experience-Driven Software Issue Resolution",
"cosine_sim": 0.7892
},
{
"paper_id": ... | git clone https://github.com/SWE-agent/Mini-SWE-Agent && cd Mini-SWE-Agent && (pip install -e . || pip install -r requirements.txt) | Large language models (LLMs) have transformed the software engineering landscape. | Recently, numerous LLM-based agents have been developed to address real-world software issue fixing tasks. | Specifically, SWE-Fuse outperforms the best 8B and 32B baselines by 43.0\% and 60.2\% in solve rate, respectively. | 8 | Code Foundation Models & Specialized Instruction Distillation | Explosive (>50/mo) | 232 | 2026-08-21T14:38:19.710898 |
2602.03419v1 | SWE-World: Building Software Engineering Agents in Docker-Free Environments | 2026-02-03T11:44:39Z | [
"cs.SE",
"cs.CL"
] | ArXiv Standard | http://arxiv.org/licenses/nonexclusive-distrib/1.0/ | true | 60 | Internal R&D Only (Copyleft or Academic Terms) | ArXiv Standard Distribution; Repository License Unspecified | 0 | 0 | 1 | Proposes While effective, this paradigm is resource-intensive and difficult to maintain, substantia... to enhance autonomous code synthesis, achieving Experiments on SWE-bench Verified demonstrate that SWE-World raises Qwen2.5-Code.... | Shuang Sun | 14 | [
"Shuang Sun",
"Huatong Song",
"Lisheng Huang",
"Jinhao Jiang",
"Ran Le",
"Zhihao Lv",
"Zongchao Chen",
"Yiwen Hu",
"Wenyang Luo",
"Wayne Xin Zhao",
"Yang Song",
"Hongteng Xu",
"Tao Zhang",
"Ji-Rong Wen"
] | [
"Autonomous Software Engineering AI Lab"
] | http://arxiv.org/abs/2602.03419v1 | VERIFIED_LIVE | https://github.com/SWE-agent/Mini-SWE-Agent | [
"https://github.com/SWE-agent/Mini-SWE-Agent",
"https://github.com/zhenyuhe00/SWE-Swiss",
"https://github.com/RUCAIBox/SWE-World"
] | 6,700 | 0 | 2026-08-21 | Unspecified | 108.56 | Recent advances in large language models (LLMs) have enabled software engineering agents to tackle complex code modification tasks. Most existing approaches rely on execution feedback from containerized environments, which require dependency-complete setup and physical execution of programs and tests. While effective, ... | [
-0.02568499930202961,
0.019207000732421875,
0.014932000078260899,
-0.005078000016510487,
0.058260999619960785,
-0.11757200211286545,
-0.04345100000500679,
-0.0012410000199452043,
0.04235300049185753,
0.021112000569701195,
-0.0832580029964447,
-0.11021099984645844,
0.09520699828863144,
-0.0... | [
-0.0002500000118743628,
-0.025490999221801758,
-0.0062790000811219215,
0.04890099912881851,
0.07395000010728836,
-0.07618299871683121,
-0.04188000038266182,
-0.03573799878358841,
0.004821999929845333,
-0.017407000064849854,
-0.06322299689054489,
-0.09302700310945511,
0.021765999495983124,
... | AI Code Generation, Autonomous SWE Agents & Program Synthesis | Autonomous Repo-Level SWE Agent (SWE-bench) | [
"Multi-Language / Polyglot"
] | CodeQwen / Qwen2.5-Coder Foundation | [
"SWE-bench (Autonomous Software Engineering)"
] | [
"Quantitative Program Synthesis & Pass@k Metric"
] | [
{
"benchmark_name": "SWE-bench (Autonomous Software Engineering)",
"metric": "pass@1 / resolved",
"score": "Quantitative Program Synthesis & Pass@k Metric"
}
] | [
{
"paper_id": "2602.00592v2",
"title": "DockSmith: Scaling Reliable Coding Environments via an Agentic Docker Builder",
"cosine_sim": 0.821
},
{
"paper_id": "2606.28436v1",
"title": "Dockerless: Environment-Free Program Verifier for Coding Agents",
"cosine_sim": 0.7892
},
{
"pape... | git clone https://github.com/SWE-agent/Mini-SWE-Agent && cd Mini-SWE-Agent && (pip install -e . || pip install -r requirements.txt) | Recent advances in large language models (LLMs) have enabled software engineering agents to tackle complex code modification tasks. | While effective, this paradigm is resource-intensive and difficult to maintain, substantially complicating agent training and limiting scalability. | Experiments on SWE-bench Verified demonstrate that SWE-World raises Qwen2.5-Coder-32B from 6.2\% to 52.0\% via Docker-free SFT, 55.0\% with Docker-free RL, and 68.2\% with further TTS. | 8 | Code Foundation Models & Specialized Instruction Distillation | Explosive (>50/mo) | 232 | 2026-08-21T14:38:53.002471 |
2602.02361v1 | SWE-Universe: Scale Real-World Verifiable Environments to Millions | 2026-02-02T17:20:30Z | [
"cs.SE",
"cs.AI"
] | ArXiv Standard | http://arxiv.org/licenses/nonexclusive-distrib/1.0/ | true | 60 | Internal R&D Only (Copyleft or Academic Terms) | ArXiv Standard Distribution; Repository License Unspecified | 0 | 0 | 1 | Proposes To overcome the prevalent challenges of automatic building, such as low production yield,... to enhance autonomous code synthesis, achieving Finally, we applied this technique to Qwen3-Max-Thinking and achieved a score of.... | Mouxiang Chen | 18 | [
"Mouxiang Chen",
"Lei Zhang",
"Yunlong Feng",
"Xuwu Wang",
"Wenting Zhao",
"Ruisheng Cao",
"Jiaxi Yang",
"Jiawei Chen",
"Mingze Li",
"Zeyao Ma",
"Hao Ge",
"Zongmeng Zhang",
"Zeyu Cui",
"Dayiheng Liu",
"Jingren Zhou",
"Jianling Sun",
"Junyang Lin",
"Binyuan Hui"
] | [
"Autonomous Software Engineering AI Lab"
] | http://arxiv.org/abs/2602.02361v1 | VERIFIED_LIVE | https://github.com/SWE-agent/Mini-SWE-Agent | [
"https://github.com/SWE-agent/Mini-SWE-Agent"
] | 6,700 | 0 | 2026-08-21 | Unspecified | 108.52 | We propose SWE-Universe, a scalable and efficient framework for automatically constructing real-world software engineering (SWE) verifiable environments from GitHub pull requests (PRs). To overcome the prevalent challenges of automatic building, such as low production yield, weak verifiers, and prohibitive cost, our fr... | [
-0.027387000620365143,
-0.05132000148296356,
0.015410000458359718,
-0.010448000393807888,
0.10125099867582321,
-0.0887570008635521,
-0.06268899887800217,
0.0009950000094249845,
0.003596999915316701,
-0.0017859999788925052,
-0.020712999626994133,
-0.11566799879074097,
0.0822950005531311,
0.... | [
-0.07904399931430817,
-0.04182799905538559,
-0.03633499890565872,
0.002773999935016036,
0.04835699871182442,
-0.05911700055003166,
-0.04251600056886673,
0.03101399913430214,
0.05398700013756752,
0.018408000469207764,
-0.06064600124955177,
-0.13739599287509918,
0.11214400082826614,
0.054891... | AI Code Generation, Autonomous SWE Agents & Program Synthesis | Autonomous Repo-Level SWE Agent (SWE-bench) | [
"Multi-Language / Polyglot"
] | Specialized Autoregressive Code Transformer | [
"SWE-bench (Autonomous Software Engineering)"
] | [
"75.3% on SWE-Bench"
] | [
{
"benchmark_name": "SWE-bench (Autonomous Software Engineering)",
"metric": "pass@1 / resolved",
"score": "75.3% on SWE-Bench"
}
] | [
{
"paper_id": "2506.07636v2",
"title": "SWE-Dev: Building Software Engineering Agents with Training and Inference Scalin",
"cosine_sim": 0.7504
},
{
"paper_id": "2601.04171v1",
"title": "Agentic Rubrics as Contextual Verifiers for SWE Agents",
"cosine_sim": 0.7459
},
{
"paper_id"... | git clone https://github.com/SWE-agent/Mini-SWE-Agent && cd Mini-SWE-Agent && (pip install -e . || pip install -r requirements.txt) | We propose SWE-Universe, a scalable and efficient framework for automatically constructing real-world software engineering (SWE) verifiable environments from GitHub pull requests (PRs). | To overcome the prevalent challenges of automatic building, such as low production yield, weak verifiers, and prohibitive cost, our framework utilizes a building agent powered by an efficient custom-trained model. | Finally, we applied this technique to Qwen3-Max-Thinking and achieved a score of 75.3% on SWE-Bench Verified. | 8 | Code Foundation Models & Specialized Instruction Distillation | Explosive (>50/mo) | 232 | 2026-08-21T14:38:51.135776 |
2601.04171v1 | Agentic Rubrics as Contextual Verifiers for SWE Agents | 2026-01-07T18:38:23Z | [
"cs.LG"
] | ArXiv Standard | http://arxiv.org/licenses/nonexclusive-distrib/1.0/ | true | 60 | Internal R&D Only (Copyleft or Academic Terms) | ArXiv Standard Distribution; Repository License Unspecified | 0 | 0 | 1 | Proposes Despite its importance, verification in software engineering (SWE) agent settings often re... to enhance autonomous code synthesis, achieving Together, these results suggest that Agentic Rubrics provide an efficient, scala.... | Mohit Raghavendra | 4 | [
"Mohit Raghavendra",
"Anisha Gunjal",
"Bing Liu",
"Yunzhong He"
] | [
"Autonomous Software Engineering AI Lab"
] | http://arxiv.org/abs/2601.04171v1 | VERIFIED_LIVE | https://github.com/SWE-agent/Mini-SWE-Agent | [
"https://github.com/SWE-agent/Mini-SWE-Agent"
] | 6,700 | 0 | 2026-08-21 | Unspecified | 107.48 | Verification is critical for improving agents: it provides the reward signal for Reinforcement Learning and enables inference-time gains through Test-Time Scaling (TTS). Despite its importance, verification in software engineering (SWE) agent settings often relies on code execution, which can be difficult to scale due ... | [
-0.03576099872589111,
-0.058956000953912735,
-0.05778999999165535,
-0.07355300337076187,
0.04129200056195259,
0.03474700078368187,
0.06290300190448761,
0.021508999168872833,
0.006000999826937914,
-0.02111399918794632,
0.0002209999947808683,
-0.02542700059711933,
0.06456100195646286,
0.0442... | [
-0.0833360031247139,
-0.04462999850511551,
-0.07054799795150757,
-0.003536999924108386,
0.0606129989027977,
-0.08120899647474289,
0.01737000048160553,
0.02544100023806095,
-0.023200999945402145,
0.021734999492764473,
-0.052976999431848526,
-0.08531700074672699,
0.08095300197601318,
0.02247... | AI Code Generation, Autonomous SWE Agents & Program Synthesis | Autonomous Repo-Level SWE Agent (SWE-bench) | [
"Python"
] | Specialized Autoregressive Code Transformer | [
"SWE-bench (Autonomous Software Engineering)"
] | [
"achieve a score of 54.2%"
] | [
{
"benchmark_name": "SWE-bench (Autonomous Software Engineering)",
"metric": "pass@1 / resolved",
"score": "achieve a score of 54.2%"
}
] | [
{
"paper_id": "2606.20512v2",
"title": "Probe-and-Refine Tuning of Repository Guidance for Coding Agents",
"cosine_sim": 0.7675
},
{
"paper_id": "2510.22775v1",
"title": "Scalable Supervising Software Agents with Patch Reasoner",
"cosine_sim": 0.7572
},
{
"paper_id": "2602.02361v... | git clone https://github.com/SWE-agent/Mini-SWE-Agent && cd Mini-SWE-Agent && (pip install -e . || pip install -r requirements.txt) | Verification is critical for improving agents: it provides the reward signal for Reinforcement Learning and enables inference-time gains through Test-Time Scaling (TTS). | Despite its importance, verification in software engineering (SWE) agent settings often relies on code execution, which can be difficult to scale due to environment setup overhead. | Together, these results suggest that Agentic Rubrics provide an efficient, scalable, and granular verification signal for SWE agents. | 8 | Code Foundation Models & Specialized Instruction Distillation | Explosive (>50/mo) | 232 | 2026-08-21T14:39:13.871666 |
2507.23370v1 | Trae Agent: An LLM-based Agent for Software Engineering with Test-time Scaling | 2025-07-31T09:37:22Z | [
"cs.SE",
"cs.AI"
] | ArXiv Standard | http://arxiv.org/licenses/nonexclusive-distrib/1.0/ | true | 60 | Internal R&D Only (Copyleft or Academic Terms) | ArXiv Standard Distribution; Repository License Unspecified | 0 | 0 | 1 | Proposes Recent studies have introduced ensemble reasoning techniques to enhance the performance of... to enhance autonomous code synthesis, achieving Trae Agent has achieved first place on the SWE-bench Verified leaderboard, with.... | Trae Research Team | 15 | [
"Trae Research Team",
"Pengfei Gao",
"Zhao Tian",
"Xiangxin Meng",
"Xinchen Wang",
"Ruida Hu",
"Yuanan Xiao",
"Yizhou Liu",
"Zhao Zhang",
"Junjie Chen",
"Cuiyun Gao",
"Yun Lin",
"Yingfei Xiong",
"Chao Peng",
"Xia Liu"
] | [
"Autonomous Software Engineering AI Lab"
] | http://arxiv.org/abs/2507.23370v1 | VERIFIED_LIVE | https://github.com/bytedance/trae-agent | [
"https://github.com/bytedance/trae-agent"
] | 12,000 | 0 | 2026-08-21 | Unspecified | 106.14 | Software issue resolution is a critical challenge in software engineering and has garnered increasing attention in recent years. With the rapid advancement of large language models (LLMs), substantial progress has been made in addressing real-world software engineering tasks. Recent studies have introduced ensemble rea... | [
-0.030900999903678894,
-0.012891000136733055,
-0.06716399639844894,
-0.0030610000248998404,
0.020945999771356583,
-0.14647699892520905,
-0.005013999994844198,
-0.0049430001527071,
-0.027890000492334366,
-0.023142000660300255,
0.00761799979954958,
-0.04166100174188614,
0.015418999828398228,
... | [
-0.06581500172615051,
-0.026924999430775642,
-0.03735800087451935,
0.06126299872994423,
0.04902200028300285,
-0.08236999809741974,
0.01510000042617321,
0.04983700066804886,
-0.001749000046402216,
0.026991000398993492,
-0.0912420004606247,
-0.0812940001487732,
0.06446400284767151,
0.0181799... | AI Code Generation, Autonomous SWE Agents & Program Synthesis | Autonomous Repo-Level SWE Agent (SWE-bench) | [
"Python"
] | Claude-3.5-Sonnet Agentic SWE Backbone | [
"SWE-bench (Autonomous Software Engineering)"
] | [
"Quantitative Program Synthesis & Pass@k Metric"
] | [
{
"benchmark_name": "SWE-bench (Autonomous Software Engineering)",
"metric": "pass@1 / resolved",
"score": "Quantitative Program Synthesis & Pass@k Metric"
}
] | [
{
"paper_id": "2507.23348v1",
"title": "SWE-Debate: Competitive Multi-Agent Debate for Software Issue Resolution",
"cosine_sim": 0.7696
},
{
"paper_id": "2512.23631v2",
"title": "BOAD: Discovering Hierarchical Software Engineering Agents via Bandit Optimizati",
"cosine_sim": 0.7528
},
... | git clone https://github.com/bytedance/trae-agent && cd trae-agent && (pip install -e . || pip install -r requirements.txt) | Software issue resolution is a critical challenge in software engineering and has garnered increasing attention in recent years. | Recent studies have introduced ensemble reasoning techniques to enhance the performance of LLM-based issue resolution. | Trae Agent has achieved first place on the SWE-bench Verified leaderboard, with a notable Pass@1 score of 75.20%. | 8 | Code Foundation Models & Specialized Instruction Distillation | Explosive (>50/mo) | 232 | 2026-08-21T14:41:07.634759 |
2512.18552v3 | Toward Training Superintelligent Software Agents through Self-Play SWE-RL | 2025-12-21T00:49:40Z | [
"cs.SE",
"cs.AI",
"cs.CL",
"cs.LG"
] | ArXiv Standard | http://arxiv.org/licenses/nonexclusive-distrib/1.0/ | true | 60 | Internal R&D Only (Copyleft or Academic Terms) | ArXiv Standard Distribution; Repository License Unspecified | 0 | 0 | 1 | Proposes we present Self-play SWE-RL (SSR), a first step toward training paradigms for superintelli... to enhance autonomous code synthesis, achieving Our results, albeit early, suggest a path where agents autonomously gather exten.... | Yuxiang Wei | 9 | [
"Yuxiang Wei",
"Zhiqing Sun",
"Emily McMilin",
"Jonas Gehring",
"David Zhang",
"Gabriel Synnaeve",
"Daniel Fried",
"Lingming Zhang",
"Sida Wang"
] | [
"Autonomous Software Engineering AI Lab"
] | http://arxiv.org/abs/2512.18552v3 | VERIFIED_LIVE | https://github.com/SWE-bench/SWE-bench | [
"https://github.com/SWE-bench/SWE-bench"
] | 5,700 | 0 | 2026-08-21 | Unspecified | 105.4 | While current software agents powered by large language models (LLMs) and agentic reinforcement learning (RL) can boost programmer productivity, their training data (e.g., GitHub issues and pull requests) and environments (e.g., pass-to-pass and fail-to-pass tests) heavily depend on human knowledge or curation, posing ... | [
-0.03880700096487999,
-0.0783419981598854,
-0.01342300046235323,
-0.0009089999948628247,
-0.0033809999004006386,
-0.035006001591682434,
0.036997001618146896,
0.05174199864268303,
-0.027276000007987022,
-0.019547000527381897,
-0.00017699999443721026,
-0.0006370000191964209,
0.0826100036501884... | [
-0.09534499794244766,
-0.07610700279474258,
-0.02765600010752678,
0.016124999150633812,
0.011274999938905239,
-0.018178999423980713,
0.005313999950885773,
0.03640199825167656,
-0.038972001522779465,
0.02109299972653389,
-0.025999000295996666,
-0.026034999638795853,
0.07060600072145462,
0.0... | AI Code Generation, Autonomous SWE Agents & Program Synthesis | Autonomous Repo-Level SWE Agent (SWE-bench) | [
"Multi-Language / Polyglot"
] | GPT-4o / Codex Specialized Coding LLM | [
"SWE-bench (Autonomous Software Engineering)"
] | [
"Quantitative Program Synthesis & Pass@k Metric"
] | [
{
"benchmark_name": "SWE-bench (Autonomous Software Engineering)",
"metric": "pass@1 / resolved",
"score": "Quantitative Program Synthesis & Pass@k Metric"
}
] | [
{
"paper_id": "2606.07412v1",
"title": "Socratic-SWE: Self-Evolving Coding Agents via Trace-Derived Agent Skills",
"cosine_sim": 0.811
},
{
"paper_id": "2604.10493v1",
"title": "SWE-Shepherd: Advancing PRMs for Reinforcing Code Agents",
"cosine_sim": 0.7704
},
{
"paper_id": "2601... | git clone https://github.com/SWE-bench/SWE-bench && cd SWE-bench && (pip install -e . || pip install -r requirements.txt) | While current software agents powered by large language models (LLMs) and agentic reinforcement learning (RL) can boost programmer productivity, their training data (e.g., GitHub issues and pull requests) and environments (e.g., pass-to-pass and fail-to-pass tests) heavily depend on human knowledge or curation, posing ... | In this paper, we present Self-play SWE-RL (SSR), a first step toward training paradigms for superintelligent software agents. | Our results, albeit early, suggest a path where agents autonomously gather extensive learning experiences from real-world software repositories, ultimately enabling superintelligent systems that exceed human capabilities in understanding how systems are constructed, solving novel challenges, and autonomously creating n... | 8 | Code Foundation Models & Specialized Instruction Distillation | Explosive (>50/mo) | 232 | 2026-08-21T14:39:24.543764 |
2601.10904v1 | ARC Prize 2025: Technical Report | 2026-01-15T23:23:56Z | [
"cs.AI"
] | ArXiv Standard | http://arxiv.org/licenses/nonexclusive-distrib/1.0/ | true | 60 | Internal R&D Only (Copyleft or Academic Terms) | ArXiv Standard Distribution; Repository License Unspecified | 0 | 0 | 1 | Proposes The ARC Prize 2025 global competition targeted the newly released ARC-AGI-2 dataset, which... to enhance autonomous code synthesis, achieving In this paper, we survey the top-performing methods, examine the role of refinem.... | François Chollet | 4 | [
"François Chollet",
"Mike Knoop",
"Gregory Kamradt",
"Bryan Landers"
] | [
"Autonomous Software Engineering AI Lab"
] | http://arxiv.org/abs/2601.10904v1 | VERIFIED_LIVE | https://github.com/fchollet/ARC-AGI | [
"https://github.com/fchollet/ARC-AGI",
"https://github.com/jerber/arc-lang-public",
"https://github.com/ijoffe/ARC-VSA-2025"
] | 4,800 | 0 | 2026-08-21 | Unspecified | 104.91 | The ARC-AGI benchmark series serves as a critical measure of few-shot generalization on novel tasks, a core aspect of intelligence. The ARC Prize 2025 global competition targeted the newly released ARC-AGI-2 dataset, which features greater task complexity compared to its predecessor. The Kaggle competition attracted 1,... | [
-0.08333099633455276,
0.007778000086545944,
0.026125000789761543,
0.02778399921953678,
-0.031062999740242958,
-0.05575399845838547,
0.02115900069475174,
0.0659639984369278,
-0.0742959976196289,
0.001978999935090542,
-0.06441300362348557,
-0.009809000417590141,
0.04946000128984451,
0.023978... | [
-0.093640998005867,
-0.06769999861717224,
0.027115000411868095,
0.03092700056731701,
0.04746900126338005,
-0.0033370000310242176,
-0.014464999549090862,
0.02025900036096573,
-0.031170999631285667,
-0.06516599655151367,
-0.09480399638414383,
-0.06501799821853638,
-0.019971000030636787,
-0.0... | AI Code Generation, Autonomous SWE Agents & Program Synthesis | Neural Program Synthesis & Generation Architecture | [
"Multi-Language / Polyglot"
] | Specialized Autoregressive Code Transformer | [
"HumanEval & SWE-bench Functional Synthesis Benchmark"
] | [
"Quantitative Program Synthesis & Pass@k Metric"
] | [
{
"benchmark_name": "HumanEval & SWE-bench Functional Synthesis Benchmark",
"metric": "pass@1 / resolved",
"score": "Quantitative Program Synthesis & Pass@k Metric"
}
] | [
{
"paper_id": "2603.13372v1",
"title": "The ARC of Progress towards AGI: A Living Survey of Abstraction and Reasoning",
"cosine_sim": 0.7852
},
{
"paper_id": "2607.06764v1",
"title": "Cost-Effective Agent Harnesses for Abstract Reasoning and Generalization on ARC-",
"cosine_sim": 0.6725
... | git clone https://github.com/fchollet/ARC-AGI && cd ARC-AGI && (pip install -e . || pip install -r requirements.txt) | The ARC-AGI benchmark series serves as a critical measure of few-shot generalization on novel tasks, a core aspect of intelligence. | The ARC Prize 2025 global competition targeted the newly released ARC-AGI-2 dataset, which features greater task complexity compared to its predecessor. | In this paper, we survey the top-performing methods, examine the role of refinement loops in AGI progress, discuss knowledge-dependent overfitting, and preview ARC-AGI-3, which introduces interactive reasoning challenges that require exploration, planning, memory, goal acquisition, and alignment capabilities. | 4 | Formal Verification, Theorem Proving & Symbolic Logic | Explosive (>50/mo) | 232 | 2026-08-21T14:39:12.301234 |
2510.19438v1 | AutoMT: A Multi-Agent LLM Framework for Automated Metamorphic Testing of Autonomous Driving Systems | 2025-10-22T10:11:05Z | [
"cs.SE"
] | ArXiv Standard | http://arxiv.org/licenses/nonexclusive-distrib/1.0/ | true | 60 | Internal R&D Only (Copyleft or Academic Terms) | ArXiv Standard Distribution; Repository License Unspecified | 0 | 0 | 1 | Proposes AutoMT, a multi-agent MT framework powered by Large Language Models (LLMs) that automates... to enhance autonomous code synthesis, achieving that AutoMT achieves up to 5 x higher test diversity in follow-up case generatio.... | Linfeng Liang | 6 | [
"Linfeng Liang",
"Chenkai Tan",
"Yao Deng",
"Yingfeng Cai",
"T. Y Chen",
"Xi Zheng"
] | [
"Autonomous Software Engineering AI Lab"
] | http://arxiv.org/abs/2510.19438v1 | VERIFIED_LIVE | https://github.com/udacity/self-driving-car | [
"https://github.com/udacity/self-driving-car"
] | 6,300 | 0 | 2026-08-21 | Unspecified | 103.87 | Autonomous Driving Systems (ADS) are safety-critical, where failures can be severe. While Metamorphic Testing (MT) is effective for fault detection in ADS, existing methods rely heavily on manual effort and lack automation. We present AutoMT, a multi-agent MT framework powered by Large Language Models (LLMs) that autom... | [
0.03116999939084053,
-0.009576000273227692,
0.0016580000519752502,
-0.056814998388290405,
0.061622001230716705,
-0.08245900273323059,
-0.015678999945521355,
-0.022481000050902367,
-0.06532499939203262,
-0.039684001356363297,
-0.02130099944770336,
-0.08171200007200241,
-0.012635000050067902,
... | [
0.010460999794304371,
-0.050172001123428345,
0.012590999715030193,
-0.029966000467538834,
0.06523100286722183,
-0.04216200113296509,
-0.003003000048920512,
-0.010498999617993832,
-0.05811600014567375,
-0.06695999950170517,
-0.057363998144865036,
-0.0959630012512207,
0.08214999735355377,
0.... | AI Code Generation, Autonomous SWE Agents & Program Synthesis | Neural Program Synthesis & Generation Architecture | [
"Multi-Language / Polyglot"
] | Claude-3.5-Sonnet Agentic SWE Backbone | [
"HumanEval & SWE-bench Functional Synthesis Benchmark"
] | [
"Quantitative Program Synthesis & Pass@k Metric"
] | [
{
"benchmark_name": "HumanEval & SWE-bench Functional Synthesis Benchmark",
"metric": "pass@1 / resolved",
"score": "Quantitative Program Synthesis & Pass@k Metric"
}
] | [
{
"paper_id": "2605.13898v1",
"title": "Bidirectional Empowerment of Metamorphic Testing and Large Language Models: A Sy",
"cosine_sim": 0.6868
},
{
"paper_id": "2510.16701v2",
"title": "An Agentic Framework with LLMs for Solving Complex Vehicle Routing Problems",
"cosine_sim": 0.6841
... | git clone https://github.com/udacity/self-driving-car && cd self-driving-car && (pip install -e . || pip install -r requirements.txt) | Autonomous Driving Systems (ADS) are safety-critical, where failures can be severe. | We present AutoMT, a multi-agent MT framework powered by Large Language Models (LLMs) that automates the extraction of Metamorphic Relations (MRs) from local traffic rules and the generation of valid follow-up test cases. | Experiments show that AutoMT achieves up to 5 x higher test diversity in follow-up case generation compared to the best baseline (manual expert-defined MRs) in terms of validation rate, and detects up to 20.55% more behavioral violations. | 1 | Autonomous Repo-Level Software Engineering Agents (SWE-bench) | Explosive (>50/mo) | 232 | 2026-08-21T14:40:05.855618 |
2603.25810v2 | ExVerus: Verus Proof Repair via Counterexample Reasoning | 2026-03-26T18:14:34Z | [
"cs.PL",
"cs.LG"
] | ArXiv Standard | http://arxiv.org/licenses/nonexclusive-distrib/1.0/ | true | 60 | Internal R&D Only (Copyleft or Academic Terms) | ArXiv Standard Distribution; Repository License Unspecified | 0 | 0 | 1 | Proposes EXVERUS, a counterexample-guided framework that enables LLMs to reason about proofs using... to enhance autonomous code synthesis, achieving Large Language Models (LLMs) have shown promising results in automating formal v.... | Jun Yang | 8 | [
"Jun Yang",
"Yuechun Sun",
"Yi Wu",
"Rodrigo Caridad",
"Yongwei Yuan",
"Jianan Yao",
"Shan Lu",
"Kexin Pei"
] | [
"Autonomous Software Engineering AI Lab"
] | http://arxiv.org/abs/2603.25810v2 | VERIFIED_LIVE | https://github.com/verus-lang/verus | [
"https://github.com/verus-lang/verus",
"https://github.com/secure-foundations/human-eval-verus",
"https://github.com/microsoft/verus-copilot-vscode",
"https://github.com/WeituoDAI/verus-study-cases-leetcode"
] | 2,900 | 0 | 2026-08-21 | Unspecified | 103.33 | Large Language Models (LLMs) have shown promising results in automating formal verification. However, existing approaches treat proof generation as a static, end-to-end prediction over source code, relying on limited verifier feedback and lacking access to concrete program behaviors. We present EXVERUS, a counterexampl... | [
-0.11394300311803818,
0.06021200120449066,
0.03724399954080582,
-0.026309000328183174,
0.10491299629211426,
-0.036052998155355453,
0.015951000154018402,
0.01661599986255169,
-0.04504700005054474,
0.07250899821519852,
-0.023685000836849213,
0.0138330003246665,
0.04230500012636185,
0.0277880... | [
-0.10030999779701233,
0.0408330000936985,
0.018539000302553177,
0.003191000083461404,
0.11386200040578842,
-0.05884300172328949,
-0.007996000349521637,
0.0614360012114048,
-0.0419900007545948,
0.04654400050640106,
-0.07553000003099442,
-0.10994400084018707,
0.06732799857854843,
0.038561999... | AI Code Generation, Autonomous SWE Agents & Program Synthesis | Multi-Turn Interactive Pair Programmer | [
"Multi-Language / Polyglot"
] | Specialized Autoregressive Code Transformer | [
"HumanEval & SWE-bench Functional Synthesis Benchmark"
] | [
"Quantitative Program Synthesis & Pass@k Metric"
] | [
{
"benchmark_name": "HumanEval & SWE-bench Functional Synthesis Benchmark",
"metric": "pass@1 / resolved",
"score": "Quantitative Program Synthesis & Pass@k Metric"
}
] | [
{
"paper_id": "2604.22601v1",
"title": "From Natural Language to Verified Code: Toward AI Assisted Problem-to-Code Gener",
"cosine_sim": 0.7596
},
{
"paper_id": "2605.08694v1",
"title": "A Learning Method for Symbolic Systems Using Large Language Models",
"cosine_sim": 0.7522
},
{
... | git clone https://github.com/verus-lang/verus && cd verus && (pip install -e . || pip install -r requirements.txt) | Large Language Models (LLMs) have shown promising results in automating formal verification. | We present EXVERUS, a counterexample-guided framework that enables LLMs to reason about proofs using behavioral feedback via counterexamples. | Large Language Models (LLMs) have shown promising results in automating formal verification. | 6 | Static Analysis, Vulnerability Detection & Automated Security Patching | Explosive (>50/mo) | 232 | 2026-08-21T14:38:09.518585 |
2509.09853v2 | SWE-Effi: Re-Evaluating Software AI Agent System Effectiveness Under Resource Constraints | 2025-09-11T21:04:10Z | [
"cs.SE",
"cs.AI"
] | ArXiv Standard | http://arxiv.org/licenses/nonexclusive-distrib/1.0/ | true | 60 | Internal R&D Only (Copyleft or Academic Terms) | ArXiv Standard Distribution; Repository License Unspecified | 0 | 0 | 1 | Proposes To address this gap, we introduce SWE-Effi, a set of new metrics to re-evaluate AI systems... to enhance autonomous code synthesis, achieving In this paper, we specifically focus on the software engineering scenario by re-.... | Zhiyu Fan | 9 | [
"Zhiyu Fan",
"Kirill Vasilevski",
"Dayi Lin",
"Boyuan Chen",
"Yihao Chen",
"Zhiqing Zhong",
"Jie M. Zhang",
"Pinjia He",
"Ahmed E. Hassan"
] | [
"Autonomous Software Engineering AI Lab"
] | http://arxiv.org/abs/2509.09853v2 | VERIFIED_LIVE | https://github.com/SWE-agent/Mini-SWE-Agent | [
"https://github.com/SWE-agent/Mini-SWE-Agent"
] | 6,700 | 0 | 2026-08-21 | Unspecified | 102.76 | The advancement of large language models (LLMs) and code agents has demonstrated significant potential to assist software engineering (SWE) tasks, such as autonomous issue resolution and feature addition. Existing AI for software engineering leaderboards (e.g., SWE-bench) focus solely on solution accuracy, ignoring the... | [
-0.03121899999678135,
-0.02585900016129017,
-0.05924899876117706,
0.02651200070977211,
0.11124099791049957,
-0.046473998576402664,
-0.004910000134259462,
-0.012737000361084938,
0.008938999846577644,
-0.05149000138044357,
-0.11545199900865555,
-0.03973099961876869,
0.08880600333213806,
0.00... | [
-0.03088200092315674,
-0.04176599904894829,
-0.002609000075608492,
0.04918299987912178,
0.08298099786043167,
-0.018028000369668007,
0.012675000354647636,
0.08822199702262878,
0.017278000712394714,
-0.02609799988567829,
-0.08466699719429016,
-0.0733570009469986,
0.0752130001783371,
0.044218... | AI Code Generation, Autonomous SWE Agents & Program Synthesis | Autonomous Repo-Level SWE Agent (SWE-bench) | [
"Multi-Language / Polyglot"
] | GPT-4o / Codex Specialized Coding LLM | [
"SWE-bench (Autonomous Software Engineering)"
] | [
"Quantitative Program Synthesis & Pass@k Metric"
] | [
{
"benchmark_name": "SWE-bench (Autonomous Software Engineering)",
"metric": "pass@1 / resolved",
"score": "Quantitative Program Synthesis & Pass@k Metric"
}
] | [
{
"paper_id": "2507.11059v3",
"title": "SWE-MERA: A Dynamic Benchmark for Agenticly Evaluating Large Language Models on ",
"cosine_sim": 0.7301
},
{
"paper_id": "2512.09543v2",
"title": "SWEnergy: An Empirical Study on Energy Efficiency in Agentic Issue Resolution Fr",
"cosine_sim": 0.72... | git clone https://github.com/SWE-agent/Mini-SWE-Agent && cd Mini-SWE-Agent && (pip install -e . || pip install -r requirements.txt) | The advancement of large language models (LLMs) and code agents has demonstrated significant potential to assist software engineering (SWE) tasks, such as autonomous issue resolution and feature addition. | To address this gap, we introduce SWE-Effi, a set of new metrics to re-evaluate AI systems in terms of holistic effectiveness scores. | In this paper, we specifically focus on the software engineering scenario by re-ranking popular AI systems for issue resolution on a subset of the SWE-bench benchmark using our new multi-dimensional metrics. | 8 | Code Foundation Models & Specialized Instruction Distillation | Explosive (>50/mo) | 232 | 2026-08-21T14:40:41.297495 |
2512.24873v3 | Let It Flow: Agentic Crafting on Rock and Roll, Building the ROME Model within an Open Agentic Learning Ecosystem | 2025-12-31T14:03:39Z | [
"cs.AI",
"cs.CL"
] | ArXiv Standard | http://arxiv.org/licenses/nonexclusive-distrib/1.0/ | true | 60 | Internal R&D Only (Copyleft or Academic Terms) | ArXiv Standard Distribution; Repository License Unspecified | 0 | 0 | 1 | Proposes Despite its importance, the open-source community lacks a principled, end-to-end ecosystem... to enhance autonomous code synthesis, achieving ROME demonstrates strong performance across benchmarks like SWE-bench Verified a.... | Weixun Wang | 90 | [
"Weixun Wang",
"XiaoXiao Xu",
"Wanhe An",
"Fangwen Dai",
"Wei Gao",
"Yancheng He",
"Ju Huang",
"Qiang Ji",
"Hanqi Jin",
"Xiaoyang Li",
"Yang Li",
"Zhongwen Li",
"Shirong Lin",
"Jiashun Liu",
"Zenan Liu",
"Tao Luo",
"Dilxat Muhtar",
"Yuanbin Qu",
"Jiaqiang Shi",
"Qinghui Sun",
... | [
"Autonomous Software Engineering AI Lab"
] | http://arxiv.org/abs/2512.24873v3 | VERIFIED_LIVE | https://github.com/alibaba/ROLL | [
"https://github.com/alibaba/ROLL",
"https://github.com/alibaba/ROCK"
] | 3,400 | 0 | 2026-08-21 | Unspecified | 101.31 | Agentic crafting requires LLMs to operate in real-world environments over multiple turns by taking actions, observing outcomes, and iteratively refining artifacts. Despite its importance, the open-source community lacks a principled, end-to-end ecosystem to streamline agent development. We introduce the Agentic Learnin... | [
0.053610000759363174,
-0.015247000381350517,
0.003734000027179718,
-0.0012369999894872308,
0.008558000437915325,
-0.010583999566733837,
-0.02453400008380413,
-0.043090999126434326,
-0.05849599838256836,
0.058823999017477036,
-0.007104000076651573,
-0.0276040006428957,
-0.03382999822497368,
... | [
-0.006066999863833189,
-0.042718999087810516,
0.0021490000654011965,
-0.006393000017851591,
0.0400100015103817,
-0.06395000219345093,
-0.032173000276088715,
-0.050220001488924026,
-0.03172300010919571,
0.007315000053495169,
-0.051277998834848404,
-0.09357800334692001,
0.002369999885559082,
... | AI Code Generation, Autonomous SWE Agents & Program Synthesis | Autonomous Repo-Level SWE Agent (SWE-bench) | [
"Multi-Language / Polyglot"
] | Specialized Autoregressive Code Transformer | [
"SWE-bench (Autonomous Software Engineering)"
] | [
"Quantitative Program Synthesis & Pass@k Metric"
] | [
{
"benchmark_name": "SWE-bench (Autonomous Software Engineering)",
"metric": "pass@1 / resolved",
"score": "Quantitative Program Synthesis & Pass@k Metric"
}
] | [
{
"paper_id": "2605.15040v3",
"title": "Orchard: An Open-Source Agentic Modeling Framework",
"cosine_sim": 0.7285
},
{
"paper_id": "2607.05471v1",
"title": "KAT-Coder-V2.5 Technical Report",
"cosine_sim": 0.721
},
{
"paper_id": "2512.04987v1",
"title": "Nex-N1: Agentic Models... | git clone https://github.com/alibaba/ROLL && cd ROLL && (pip install -e . || pip install -r requirements.txt) | Agentic crafting requires LLMs to operate in real-world environments over multiple turns by taking actions, observing outcomes, and iteratively refining artifacts. | Despite its importance, the open-source community lacks a principled, end-to-end ecosystem to streamline agent development. | ROME demonstrates strong performance across benchmarks like SWE-bench Verified and Terminal Bench, proving the effectiveness of ALE. | 1 | Autonomous Repo-Level Software Engineering Agents (SWE-bench) | Explosive (>50/mo) | 232 | 2026-08-21T14:39:21.923602 |
2604.16571v1 | EquivFusion: Unifying Hardware Equivalence Checking from Algorithms to Netlists via MLIR | 2026-04-17T12:09:54Z | [
"cs.AR",
"cs.SE"
] | ArXiv Standard | http://arxiv.org/licenses/nonexclusive-distrib/1.0/ | true | 60 | Internal R&D Only (Copyleft or Academic Terms) | ArXiv Standard Distribution; Repository License Unspecified | 0 | 0 | 1 | Proposes we present EquivFusion, an end-to-end equivalence checking tool tailored for multi-modal c... to enhance autonomous code synthesis, achieving We demonstrate EquivFusion's feasibility to bridge the semantic gap between soft.... | Jiaying Zhu | 7 | [
"Jiaying Zhu",
"Baoqi Zhang",
"Mengxia Tao",
"Kezhi Li",
"Hao Yan",
"Qiang Xu",
"Min Li"
] | [
"Autonomous Software Engineering AI Lab"
] | http://arxiv.org/abs/2604.16571v1 | NOT_DETECTED | Not Applicable | [] | 0 | 0 | Not Applicable | Not Applicable | 100.54 | Ensuring functional consistency between high-level algorithmic models and low-level hardware implementations is a critical challenge, particularly as modern design flows increasingly span heterogeneous abstractions--from deep learning frameworks to hardware netlists. In this paper, we present EquivFusion, an end-to-end... | [
-0.07582899928092957,
-0.03767700120806694,
-0.01825599931180477,
-0.04907599836587906,
0.09989699721336365,
-0.09240599721670151,
0.03428199887275696,
-0.03753900155425072,
-0.05449000000953674,
-0.06383699923753738,
-0.0016299999551847577,
-0.08119399845600128,
-0.030056999996304512,
-0.... | [
-0.07383199781179428,
-0.005485000088810921,
0.013621999882161617,
-0.029170000925660133,
0.08320999890565872,
-0.07936099916696548,
-0.05164499953389168,
-0.022954000160098076,
-0.03421200066804886,
-0.05539799854159355,
-0.020459000021219254,
-0.13150599598884583,
-0.05652200058102608,
-... | AI Code Generation, Autonomous SWE Agents & Program Synthesis | Formal Logic & Theorem Proving Synthesis | [
"Multi-Language / Polyglot"
] | Specialized Autoregressive Code Transformer | [
"HumanEval & SWE-bench Functional Synthesis Benchmark"
] | [
"Quantitative Program Synthesis & Pass@k Metric"
] | [
{
"benchmark_name": "HumanEval & SWE-bench Functional Synthesis Benchmark",
"metric": "pass@1 / resolved",
"score": "Quantitative Program Synthesis & Pass@k Metric"
}
] | [
{
"paper_id": "2603.09161v2",
"title": "Wrong Code, Right Structure: Learning Netlist Representations from Imperfect LLM",
"cosine_sim": 0.6957
},
{
"paper_id": "2603.25768v1",
"title": "UCAgent: An End-to-End Agent for Block-Level Functional Verification",
"cosine_sim": 0.6801
},
{
... | No public code repository attached. Paper provides formal program synthesis specification. | Ensuring functional consistency between high-level algorithmic models and low-level hardware implementations is a critical challenge, particularly as modern design flows increasingly span heterogeneous abstractions--from deep learning frameworks to hardware netlists. | In this paper, we present EquivFusion, an end-to-end equivalence checking tool tailored for multi-modal circuit designs. | We demonstrate EquivFusion's feasibility to bridge the semantic gap between software specifications and hardware realizations, showcasing its effectiveness in facilitating "shift-left" formal verification for datapath-intensive hardware designs. | 6 | Static Analysis, Vulnerability Detection & Automated Security Patching | Explosive (>50/mo) | 232 | 2026-08-21T14:37:38.378332 |
2606.02091v2 | DFlare: Scaling Up Draft Capacity for Block Diffusion Speculative Decoding | 2026-06-01T11:18:30Z | [
"cs.CL"
] | ArXiv Standard | http://arxiv.org/licenses/nonexclusive-distrib/1.0/ | true | 60 | Internal R&D Only (Copyleft or Academic Terms) | ArXiv Standard Distribution; Repository License Unspecified | 0 | 0 | 1 | Proposes However, the state-of-the-art method DFlash constrains all draft layers to share a single... to enhance autonomous code synthesis, achieving Our code is available at https://github.com/Tencent/AngelSlim.. | Jiebin Zhang | 12 | [
"Jiebin Zhang",
"Zhenghan Yu",
"Song Liu",
"Eugene J. Yu",
"Zheng Li",
"Dawei Zhu",
"Jiangshan Duo",
"Weimin Xiong",
"Yifan Song",
"Guanghua Yu",
"Jianchen Zhu",
"Sujian Li"
] | [
"Autonomous Software Engineering AI Lab"
] | http://arxiv.org/abs/2606.02091v2 | VERIFIED_LIVE | https://github.com/Tencent/AngelSlim | [
"https://github.com/Tencent/AngelSlim"
] | 1,500 | 0 | 2026-08-21 | Unspecified | 100.29 | Block diffusion speculative decoding accelerates LLM inference by predicting all tokens within a block simultaneously for the target model to verify in parallel. Predicting an entire block at once requires a sufficiently capable draft model and effective utilization of the target model's internal knowledge. However, th... | [
-0.048652999103069305,
-0.07485699653625488,
-0.04909000173211098,
0.027946999296545982,
0.049063000828027725,
-0.07000499963760376,
-0.029752999544143677,
-0.04559199884533882,
0.04258999973535538,
-0.003585000056773424,
-0.024235999211668968,
0.04196099936962128,
-0.05737299844622612,
-0... | [
-0.047775998711586,
-0.09775400161743164,
0.0073449998162686825,
0.09732499718666077,
0.11798399686813354,
-0.04136500135064125,
-0.04715999960899353,
-0.028706999495625496,
0.051086001098155975,
-0.034233998507261276,
-0.06694000214338303,
-0.06716100126504898,
0.013465999625623226,
0.019... | AI Code Generation, Autonomous SWE Agents & Program Synthesis | Neural Program Synthesis & Generation Architecture | [
"Multi-Language / Polyglot"
] | Specialized Autoregressive Code Transformer | [
"HumanEval & SWE-bench Functional Synthesis Benchmark"
] | [
"Quantitative Program Synthesis & Pass@k Metric"
] | [
{
"benchmark_name": "HumanEval & SWE-bench Functional Synthesis Benchmark",
"metric": "pass@1 / resolved",
"score": "Quantitative Program Synthesis & Pass@k Metric"
}
] | [
{
"paper_id": "2605.15609v1",
"title": "PSD: Pushing the Pareto Frontier of Diffusion LLMs via Parallel Speculative Deco",
"cosine_sim": 0.7105
},
{
"paper_id": "2509.20416v1",
"title": "FastEagle: Cascaded Drafting for Accelerating Speculative Decoding",
"cosine_sim": 0.6877
},
{
... | git clone https://github.com/Tencent/AngelSlim && cd AngelSlim && (pip install -e . || pip install -r requirements.txt) | Block diffusion speculative decoding accelerates LLM inference by predicting all tokens within a block simultaneously for the target model to verify in parallel. | However, the state-of-the-art method DFlash constrains all draft layers to share a single fused representation derived from only a few target layers, limiting per-layer expressiveness and hindering further scaling of draft capacity. | Our code is available at https://github.com/Tencent/AngelSlim. | 4 | Formal Verification, Theorem Proving & Symbolic Logic | Explosive (>50/mo) | 232 | 2026-08-21T14:36:36.059985 |
2604.22046v1 | Call-Chain-Aware LLM-Based Test Generation for Java Projects | 2026-04-23T20:03:18Z | [
"cs.SE",
"cs.AI"
] | ArXiv Standard | http://arxiv.org/licenses/nonexclusive-distrib/1.0/ | true | 60 | Internal R&D Only (Copyleft or Academic Terms) | ArXiv Standard Distribution; Repository License Unspecified | 0 | 0 | 1 | Proposes we present CAT, a novel call-chain-aware LLM-based test generation approach that explicitl... to enhance autonomous code synthesis, achieving An ablation study further demonstrates the importance of call-chain and dependen.... | Guancheng Wang | 5 | [
"Guancheng Wang",
"Qinghua Xu",
"Lionel C. Briand",
"Zhaoqiang Guo",
"Kui Liu"
] | [
"Autonomous Software Engineering AI Lab"
] | http://arxiv.org/abs/2604.22046v1 | VERIFIED_LIVE | https://github.com/google/adk-java | [
"https://github.com/google/adk-java",
"https://github.com/binance/binance-connector-java"
] | 1,700 | 0 | 2026-08-21 | Unspecified | 99.81 | Large language models (LLMs) have recently shown strong potential for generating project-level unit tests. However, existing state-of-the-art approaches primarily rely on execution-path information to guide prompt construction, which is often insufficient for complex software systems with rich inter-class dependencies,... | [
-0.04995099827647209,
0.013333999551832676,
0.002583000110462308,
-0.02273700013756752,
-0.01450200006365776,
-0.12680000066757202,
-0.07032299786806107,
0.006089999806135893,
-0.03888100013136864,
-0.02195899933576584,
-0.005437999963760376,
-0.02628600038588047,
0.050296999514102936,
-0.... | [
-0.08310999721288681,
-0.01219900045543909,
-0.0010179999517276883,
0.002715999959036708,
-0.03265500068664551,
-0.1163799986243248,
-0.07028499990701675,
0.012294000014662743,
-0.020370999351143837,
-0.03306800127029419,
-0.056108999997377396,
-0.08589500188827515,
0.03069400042295456,
0.... | AI Code Generation, Autonomous SWE Agents & Program Synthesis | Static Analysis & Vulnerability Remediation | [
"Java / Kotlin"
] | Specialized Autoregressive Code Transformer | [
"Defects4J (Automated Program Repair Benchmark)"
] | [
"Quantitative Program Synthesis & Pass@k Metric"
] | [
{
"benchmark_name": "Defects4J (Automated Program Repair Benchmark)",
"metric": "pass@1 / resolved",
"score": "Quantitative Program Synthesis & Pass@k Metric"
}
] | [
{
"paper_id": "2511.20403v2",
"title": "LLMs for Automated Unit Test Generation and Assessment in Java: The AgoneTest Fr",
"cosine_sim": 0.8017
},
{
"paper_id": "2509.23812v2",
"title": "Navigating the Labyrinth: Path-Sensitive Unit Test Generation with Large Languag",
"cosine_sim": 0.79... | git clone https://github.com/google/adk-java && cd adk-java && (pip install -e . || pip install -r requirements.txt) | Large language models (LLMs) have recently shown strong potential for generating project-level unit tests. | In this paper, we present CAT, a novel call-chain-aware LLM-based test generation approach that explicitly incorporates call-chain and dependency contexts into prompts through dedicated static analysis. | An ablation study further demonstrates the importance of call-chain and dependency contexts in CAT. | 7 | Syntax-Guided AST Modeling & Semantic Code Search | Explosive (>50/mo) | 232 | 2026-08-21T14:37:29.428492 |
2508.06471v1 | GLM-4.5: Agentic, Reasoning, and Coding (ARC) Foundation Models | 2025-08-08T17:21:06Z | [
"cs.CL"
] | ArXiv Standard | http://arxiv.org/licenses/nonexclusive-distrib/1.0/ | true | 60 | Internal R&D Only (Copyleft or Academic Terms) | ArXiv Standard Distribution; Repository License Unspecified | 0 | 0 | 1 | Proposes Through multi-stage training on 23T tokens and comprehensive post-training with expert mod... to enhance autonomous code synthesis, achieving Through multi-stage training on 23T tokens and comprehensive post-training with.... | GLM-4. 5 Team | 172 | [
"GLM-4. 5 Team",
":",
"Aohan Zeng",
"Xin Lv",
"Qinkai Zheng",
"Zhenyu Hou",
"Bin Chen",
"Chengxing Xie",
"Cunxiang Wang",
"Da Yin",
"Hao Zeng",
"Jiajie Zhang",
"Kedong Wang",
"Lucen Zhong",
"Mingdao Liu",
"Rui Lu",
"Shulin Cao",
"Xiaohan Zhang",
"Xuancheng Huang",
"Yao Wei",
... | [
"Autonomous Software Engineering AI Lab"
] | http://arxiv.org/abs/2508.06471v1 | VERIFIED_LIVE | https://github.com/zai-org/GLM-4.5 | [
"https://github.com/zai-org/GLM-4.5",
"https://github.com/zai-org/glm-simple-evals"
] | 4,400 | 0 | 2026-08-21 | Unspecified | 97.75 | We present GLM-4.5, an open-source Mixture-of-Experts (MoE) large language model with 355B total parameters and 32B activated parameters, featuring a hybrid reasoning method that supports both thinking and direct response modes. Through multi-stage training on 23T tokens and comprehensive post-training with expert mode... | [
0.012044000439345837,
-0.08811599761247635,
-0.01969500072300434,
-0.006422999780625105,
0.0883369967341423,
-0.012583999894559383,
0.00393400015309453,
0.02029000036418438,
-0.009348000399768353,
0.08574000000953674,
0.02676600031554699,
-0.04653400182723999,
-0.0013970000436529517,
0.053... | [
-0.003003000048920512,
-0.08354499936103821,
-0.003783999942243099,
-0.000830999983008951,
0.006705000065267086,
-0.047221001237630844,
0.04883499816060066,
0.05873600021004677,
-0.011068999767303467,
0.03037399984896183,
-0.026190999895334244,
-0.05115099996328354,
0.02218399941921234,
0.... | AI Code Generation, Autonomous SWE Agents & Program Synthesis | Autonomous Repo-Level SWE Agent (SWE-bench) | [
"Multi-Language / Polyglot"
] | Specialized Autoregressive Code Transformer | [
"SWE-bench (Autonomous Software Engineering)"
] | [
"64.2% on SWE-bench"
] | [
{
"benchmark_name": "SWE-bench (Autonomous Software Engineering)",
"metric": "pass@1 / resolved",
"score": "64.2% on SWE-bench"
}
] | [
{
"paper_id": "2605.31268v1",
"title": "Mellum2 Technical Report",
"cosine_sim": 0.6657
},
{
"paper_id": "2512.13278v2",
"title": "AutoTool: Dynamic Tool Selection and Integration for Agentic Reasoning",
"cosine_sim": 0.6509
},
{
"paper_id": "2605.09879v1",
"title": "M2A: Syn... | git clone https://github.com/zai-org/GLM-4.5 && cd GLM-4.5 && (pip install -e . || pip install -r requirements.txt) | We present GLM-4.5, an open-source Mixture-of-Experts (MoE) large language model with 355B total parameters and 32B activated parameters, featuring a hybrid reasoning method that supports both thinking and direct response modes. | Through multi-stage training on 23T tokens and comprehensive post-training with expert model iteration and reinforcement learning, GLM-4.5 achieves strong performance across agentic, reasoning, and coding (ARC) tasks, scoring 70.1% on TAU-Bench, 91.0% on AIME 24, and 64.2% on SWE-bench Verified. | Through multi-stage training on 23T tokens and comprehensive post-training with expert model iteration and reinforcement learning, GLM-4.5 achieves strong performance across agentic, reasoning, and coding (ARC) tasks, scoring 70.1% on TAU-Bench, 91.0% on AIME 24, and 64.2% on SWE-bench Verified. | 5 | Interactive Coding Assistants & Human-AI Pair Programming | Explosive (>50/mo) | 232 | 2026-08-21T14:41:03.238581 |
2605.30218v1 | MarginGate: Sparse Margin-Triggered Verification for Batch-Invariant LLM Inference | 2026-05-28T16:50:19Z | [
"cs.LG",
"cs.PF"
] | ArXiv Standard | http://arxiv.org/licenses/nonexclusive-distrib/1.0/ | true | 60 | Internal R&D Only (Copyleft or Academic Terms) | ArXiv Standard Distribution; Repository License Unspecified | 0 | 0 | 1 | Proposes Existing fixes use batch-invariant operators or LLM-42's per-token verification, incurring... to enhance autonomous code synthesis, achieving We evaluate on four datasets, calibrating on MATH500 and transferring to GSM8K,.... | Kexin Chu | 3 | [
"Kexin Chu",
"Yang Zhou",
"Wei Zhang"
] | [
"Autonomous Software Engineering AI Lab"
] | http://arxiv.org/abs/2605.30218v1 | VERIFIED_LIVE | https://github.com/thinking-machines-lab/batch_invariant_ops | [
"https://github.com/thinking-machines-lab/batch_invariant_ops"
] | 1,100 | 0 | 2026-08-21 | Unspecified | 97.44 | Temperature-zero BF16 LLM inference is often treated as reproducible, yet the same request can emit different tokens when decoded alone or inside a larger batch. Existing fixes use batch-invariant operators or LLM-42's per-token verification, incurring cost even when most steps are stable. We ask whether verification c... | [
-0.01846900023519993,
-0.08502200245857239,
-0.0026370000559836626,
0.011258999817073345,
0.12966400384902954,
-0.009424000047147274,
-0.008403999730944633,
-0.047217000275850296,
-0.03814699873328209,
-0.02437100000679493,
0.005607999861240387,
-0.06552200019359589,
0.022283999249339104,
... | [
-0.08189500123262405,
-0.06776300072669983,
-0.004224000032991171,
0.028106000274419785,
0.021383000537753105,
-0.10517100244760513,
-0.0701259970664978,
0.009452000260353088,
0.029519999399781227,
0.020885000005364418,
-0.02224300056695938,
-0.002618999918922782,
0.012892999686300755,
-0.... | AI Code Generation, Autonomous SWE Agents & Program Synthesis | Neural Program Synthesis & Generation Architecture | [
"Multi-Language / Polyglot"
] | Specialized Autoregressive Code Transformer | [
"HumanEval (Functional Python Correctness)"
] | [
"Quantitative Program Synthesis & Pass@k Metric"
] | [
{
"benchmark_name": "HumanEval (Functional Python Correctness)",
"metric": "pass@1 / resolved",
"score": "Quantitative Program Synthesis & Pass@k Metric"
}
] | [
{
"paper_id": "2605.17613v1",
"title": "VeriCache: Turning Lossy KV Cache into Lossless LLM Inference",
"cosine_sim": 0.6277
},
{
"paper_id": "2606.27396v2",
"title": "Test-Input Generation for Tensor Programs: What Actually Finds Kernel Bugs",
"cosine_sim": 0.5741
},
{
"paper_id... | git clone https://github.com/thinking-machines-lab/batch_invariant_ops && cd batch_invariant_ops && (pip install -e . || pip install -r requirements.txt) | Temperature-zero BF16 LLM inference is often treated as reproducible, yet the same request can emit different tokens when decoded alone or inside a larger batch. | Existing fixes use batch-invariant operators or LLM-42's per-token verification, incurring cost even when most steps are stable. | We evaluate on four datasets, calibrating on MATH500 and transferring to GSM8K, SharedGPT, and HumanEval. | 4 | Formal Verification, Theorem Proving & Symbolic Logic | Explosive (>50/mo) | 232 | 2026-08-21T14:36:37.842891 |
2510.25015v4 | VeriStruct: AI-assisted Automated Verification of Data-Structure Modules in Verus | 2025-10-28T22:28:37Z | [
"cs.SE",
"cs.AI"
] | ArXiv Standard | http://arxiv.org/licenses/nonexclusive-distrib/1.0/ | true | 60 | Internal R&D Only (Copyleft or Academic Terms) | ArXiv Standard Distribution; Repository License Unspecified | 0 | 0 | 1 | Proposes These results represent an important step toward the goal of automatic AI-assisted formal... to enhance autonomous code synthesis, achieving These results represent an important step toward the goal of automatic AI-assist.... | Chuyue Sun | 8 | [
"Chuyue Sun",
"Yican Sun",
"Daneshvar Amrollahi",
"Ethan Zhang",
"Shuvendu Lahiri",
"Shan Lu",
"David Dill",
"Clark Barrett"
] | [
"Autonomous Software Engineering AI Lab"
] | http://arxiv.org/abs/2510.25015v4 | VERIFIED_LIVE | https://github.com/verus-lang/verus | [
"https://github.com/verus-lang/verus",
"https://github.com/ChuyueSun/VeriStruct"
] | 2,900 | 0 | 2026-08-21 | Unspecified | 97.37 | We introduce VeriStruct, a novel framework that extends AI-assisted automated verification from single functions to more complex data structure modules in Verus. VeriStruct employs a planner module to orchestrate the systematic generation of abstractions, type invariants, specifications, and proof code. To address the ... | [
-0.06657499819993973,
0.02555299922823906,
-0.01028400007635355,
-0.021412000060081482,
0.04752099886536598,
-0.11739599704742432,
-0.06233400106430054,
0.024121999740600586,
-0.06259799748659134,
-0.04797400161623955,
0.02795499935746193,
-0.05613800138235092,
0.03539400175213814,
0.02061... | [
-0.0606050007045269,
0.04392699897289276,
0.038624998182058334,
0.011024000123143196,
0.05758799985051155,
-0.10123500227928162,
-0.024542000144720078,
0.05608300119638443,
-0.05144700035452843,
-0.0313819982111454,
-0.03357100114226341,
-0.09035699814558029,
0.053380999714136124,
0.018954... | AI Code Generation, Autonomous SWE Agents & Program Synthesis | Formal Logic & Theorem Proving Synthesis | [
"Rust"
] | Specialized Autoregressive Code Transformer | [
"HumanEval & SWE-bench Functional Synthesis Benchmark"
] | [
"Quantitative Program Synthesis & Pass@k Metric"
] | [
{
"benchmark_name": "HumanEval & SWE-bench Functional Synthesis Benchmark",
"metric": "pass@1 / resolved",
"score": "Quantitative Program Synthesis & Pass@k Metric"
}
] | [
{
"paper_id": "2602.04910v3",
"title": "Reducing the Costs of Proof Synthesis on Rust Systems by Scaling Up a Seed Train",
"cosine_sim": 0.7421
},
{
"paper_id": "2607.19795v1",
"title": "Towards Automated Formal Verification of zkEVMs Using LLM-Guided Constraint Synt",
"cosine_sim": 0.71... | git clone https://github.com/verus-lang/verus && cd verus && (pip install -e . || pip install -r requirements.txt) | We introduce VeriStruct, a novel framework that extends AI-assisted automated verification from single functions to more complex data structure modules in Verus. | These results represent an important step toward the goal of automatic AI-assisted formal verification. | These results represent an important step toward the goal of automatic AI-assisted formal verification. | 6 | Static Analysis, Vulnerability Detection & Automated Security Patching | Explosive (>50/mo) | 232 | 2026-08-21T14:40:01.463110 |
2608.15071v1 | Evo-Harness: Context-to-Harness Skill Compilation for Self-Evolving Agents | 2026-08-15T06:43:56Z | [
"cs.AI",
"cs.CL"
] | ArXiv Standard | http://arxiv.org/licenses/nonexclusive-distrib/1.0/ | true | 60 | Internal R&D Only (Copyleft or Academic Terms) | ArXiv Standard Distribution; Repository License Unspecified | 0 | 0 | 1 | Proposes However, agents in realistic environments continuously encounter novel tasks, often offeri... to enhance autonomous code synthesis, achieving Our extensive analysis demonstrates the effectiveness of Evo-Harness and provide.... | Tianxin Wei | 17 | [
"Tianxin Wei",
"Zhan Shi",
"Minhua Lin",
"Bing He",
"Zewen Liu",
"Yisi Sang",
"Yuanchen Bei",
"Xuying Ning",
"Jiaru Zou",
"Ting-Wei Li",
"Xiao Lin",
"Yanjun Zhao",
"Chi Wang",
"Benoit Dumoulin",
"Dakuo Wang",
"Jingrui He",
"Hanqing Lu"
] | [
"Autonomous Software Engineering AI Lab"
] | http://arxiv.org/abs/2608.15071v1 | VERIFIED_LIVE | https://github.com/A-EVO-Lab/a-evolve | [
"https://github.com/A-EVO-Lab/a-evolve"
] | 740 | 93 | 2026-08-21 | Unspecified | 97.16 | Learning from experience is critical for developing capable, self-improving large language model (LLM) agents. Existing methods typically extract knowledge from accumulated trajectories via reflection, memory, rules, or skills. However, agents in realistic environments continuously encounter novel tasks, often offering... | [
-0.009372999891638756,
-0.027042999863624573,
-0.056001000106334686,
0.026487000286579132,
0.013876000419259071,
0.01343000028282404,
0.09431000053882599,
-0.019958000630140305,
-0.06571999937295914,
-0.04938200116157532,
0.03561500087380409,
-0.14891600608825684,
0.041561998426914215,
0.0... | [
0.00887999963015318,
-0.10244400054216385,
-0.009499000385403633,
0.06443800032138824,
0.04661500081419945,
-0.0164560005068779,
-0.016481000930070877,
-0.035957999527454376,
-0.0025289999321103096,
-0.057767000049352646,
-0.024911999702453613,
-0.08602699637413025,
0.04407000169157982,
0.... | AI Code Generation, Autonomous SWE Agents & Program Synthesis | Autonomous Repo-Level SWE Agent (SWE-bench) | [
"Multi-Language / Polyglot"
] | Specialized Autoregressive Code Transformer | [
"SWE-bench (Autonomous Software Engineering)"
] | [
"Quantitative Program Synthesis & Pass@k Metric"
] | [
{
"benchmark_name": "SWE-bench (Autonomous Software Engineering)",
"metric": "pass@1 / resolved",
"score": "Quantitative Program Synthesis & Pass@k Metric"
}
] | [
{
"paper_id": "2606.05922v2",
"title": "Evolving Agents in the Dark: Retrospective Harness Optimization via Self-Prefere",
"cosine_sim": 0.7514
},
{
"paper_id": "2604.09791v1",
"title": "Pioneer Agent: Continual Improvement of Small Language Models in Production",
"cosine_sim": 0.731
}... | git clone https://github.com/A-EVO-Lab/a-evolve && cd a-evolve && (pip install -e . || pip install -r requirements.txt) | Learning from experience is critical for developing capable, self-improving large language model (LLM) agents. | However, agents in realistic environments continuously encounter novel tasks, often offering only a one-shot opportunity to improve. | Our extensive analysis demonstrates the effectiveness of Evo-Harness and provides a principled understanding of how LLM agents can effectively learn on the fly. | 5 | Interactive Coding Assistants & Human-AI Pair Programming | Explosive (>50/mo) | 232 | 2026-08-21T14:35:26.803666 |
2604.05854v1 | Deep Researcher Agent: An Autonomous Framework for 24/7 Deep Learning Experimentation with Zero-Cost Monitoring | 2026-04-07T13:16:31Z | [
"cs.AI"
] | ArXiv Standard | http://arxiv.org/licenses/nonexclusive-distrib/1.0/ | true | 60 | Internal R&D Only (Copyleft or Academic Terms) | ArXiv Standard Distribution; Repository License Unspecified | 0 | 0 | 1 | Proposes The framework introduces three key innovations: (1) \textbf{Zero-Cost Monitoring} -- a mon... to enhance autonomous code synthesis, achieving Unlike existing AI research assistants that focus on paper writing or code gener.... | Xiangyue Zhang | 1 | [
"Xiangyue Zhang"
] | [
"Autonomous Software Engineering AI Lab"
] | http://arxiv.org/abs/2604.05854v1 | VERIFIED_LIVE | https://github.com/Xiangyue-Zhang/auto-deep-researcher-24x7 | [
"https://github.com/Xiangyue-Zhang/auto-deep-researcher-24x7"
] | 1,300 | 0 | 2026-08-21 | Unspecified | 96.85 | We present \textbf{Deep Researcher Agent}, an open-source framework that enables large language model (LLM) agents to autonomously conduct deep learning experiments around the clock. Unlike existing AI research assistants that focus on paper writing or code generation, our system addresses the full experiment lifecycle... | [
-0.008573000319302082,
-0.02685599960386753,
0.007588000036776066,
0.08037599921226501,
0.035151999443769455,
-0.07287599891424179,
-0.03749699890613556,
-0.08718899637460709,
0.03976000100374222,
0.01689399965107441,
-0.028310999274253845,
-0.0728600025177002,
-0.006876000203192234,
0.038... | [
-0.08584699779748917,
-0.04352400079369545,
-0.03660999983549118,
0.07171700149774551,
0.010641999542713165,
-0.06672099977731705,
-0.058681998401880264,
-0.0024019998963922262,
0.03498600050806999,
0.01772399991750717,
-0.0456710010766983,
-0.060182999819517136,
0.003022999968379736,
0.01... | AI Code Generation, Autonomous SWE Agents & Program Synthesis | Autonomous Repo-Level SWE Agent (SWE-bench) | [
"Multi-Language / Polyglot"
] | Specialized Autoregressive Code Transformer | [
"HumanEval & SWE-bench Functional Synthesis Benchmark"
] | [
"achieving a 52"
] | [
{
"benchmark_name": "HumanEval & SWE-bench Functional Synthesis Benchmark",
"metric": "pass@1 / resolved",
"score": "achieving a 52"
}
] | [
{
"paper_id": "2603.13327v1",
"title": "DOVA: Deliberation-First Multi-Agent Orchestration for Autonomous Research Autom",
"cosine_sim": 0.7267
},
{
"paper_id": "2604.17745v2",
"title": "HiRAS: A Hierarchical Multi-Agent Framework for Paper-to-Code Generation and Exe",
"cosine_sim": 0.71... | git clone https://github.com/Xiangyue-Zhang/auto-deep-researcher-24x7 && cd auto-deep-researcher-24x7 && (pip install -e . || pip install -r requirements.txt) | We present \textbf{Deep Researcher Agent}, an open-source framework that enables large language model (LLM) agents to autonomously conduct deep learning experiments around the clock. | The framework introduces three key innovations: (1) \textbf{Zero-Cost Monitoring} -- a monitoring paradigm that incurs zero LLM API costs during model training by relying solely on process-level checks and log file reads; (2) \textbf{Two-Tier Constant-Size Memory} -- a memory architecture capped at $\sim$5K characters ... | Unlike existing AI research assistants that focus on paper writing or code generation, our system addresses the full experiment lifecycle: hypothesis formation, code implementation, training execution, result analysis, and iterative refinement. | 1 | Autonomous Repo-Level Software Engineering Agents (SWE-bench) | Explosive (>50/mo) | 232 | 2026-08-21T14:37:50.710192 |
2511.16108v1 | SkyRL-Agent: Efficient RL Training for Multi-turn LLM Agent | 2025-11-20T07:05:19Z | [
"cs.AI"
] | ArXiv Standard | http://arxiv.org/licenses/nonexclusive-distrib/1.0/ | true | 60 | Internal R&D Only (Copyleft or Academic Terms) | ArXiv Standard Distribution; Repository License Unspecified | 0 | 0 | 1 | Proposes It provides efficient asynchronous dispatching, lightweight tool integration, and flexible... to enhance autonomous code synthesis, achieving We further demonstrate SkyRL-Agent's extensibility through case studies on deep.... | Shiyi Cao | 15 | [
"Shiyi Cao",
"Dacheng Li",
"Fangzhou Zhao",
"Shuo Yuan",
"Sumanth R. Hegde",
"Connor Chen",
"Charlie Ruan",
"Tyler Griggs",
"Shu Liu",
"Eric Tang",
"Richard Liaw",
"Philipp Moritz",
"Matei Zaharia",
"Joseph E. Gonzalez",
"Ion Stoica"
] | [
"Autonomous Software Engineering AI Lab"
] | http://arxiv.org/abs/2511.16108v1 | VERIFIED_LIVE | https://github.com/NovaSky-AI/SkyRL | [
"https://github.com/NovaSky-AI/SkyRL"
] | 2,200 | 0 | 2026-08-21 | Unspecified | 95.89 | We introduce SkyRL-Agent, a framework for efficient, multi-turn, long-horizon agent training and evaluation. It provides efficient asynchronous dispatching, lightweight tool integration, and flexible backend interoperability, enabling seamless use with existing RL frameworks such as SkyRL-train, VeRL, and Tinker. Using... | [
-0.017652999609708786,
-0.0736669972538948,
-0.019538000226020813,
-0.027135999873280525,
-0.011130000464618206,
0.016979999840259552,
0.02307100035250187,
-0.04776100069284439,
-0.013162000104784966,
-0.026002999395132065,
-0.03346699848771095,
-0.025095000863075256,
0.0014929999597370625,
... | [
-0.09837699681520462,
-0.004616000223904848,
-0.05899300053715706,
0.06506499648094177,
-0.03416299819946289,
-0.03887400031089783,
-0.05414300039410591,
-0.030705999583005905,
-0.0468359999358654,
-0.04774999991059303,
-0.060798000544309616,
-0.0839489996433258,
-0.031897999346256256,
-0.... | AI Code Generation, Autonomous SWE Agents & Program Synthesis | Autonomous Repo-Level SWE Agent (SWE-bench) | [
"Multi-Language / Polyglot"
] | Specialized Autoregressive Code Transformer | [
"SWE-bench (Autonomous Software Engineering)"
] | [
"24.4% Pass@1",
"39.4% Pass@1",
"achieves a 1.55"
] | [
{
"benchmark_name": "SWE-bench (Autonomous Software Engineering)",
"metric": "pass@1 / resolved",
"score": "24.4% Pass@1"
}
] | [
{
"paper_id": "2608.17393v1",
"title": "LEGO-RL: Harness-Native Reinforcement Learning for Coding Agents",
"cosine_sim": 0.7153
},
{
"paper_id": "2605.12857v1",
"title": "ChipMATE: Multi-Agent Training via Reinforcement Learning for Enhanced RTL Gener",
"cosine_sim": 0.6918
},
{
... | git clone https://github.com/NovaSky-AI/SkyRL && cd SkyRL && (pip install -e . || pip install -r requirements.txt) | We introduce SkyRL-Agent, a framework for efficient, multi-turn, long-horizon agent training and evaluation. | It provides efficient asynchronous dispatching, lightweight tool integration, and flexible backend interoperability, enabling seamless use with existing RL frameworks such as SkyRL-train, VeRL, and Tinker. | We further demonstrate SkyRL-Agent's extensibility through case studies on deep research, computer use, and memory agents, each trained using a different training backend. | 1 | Autonomous Repo-Level Software Engineering Agents (SWE-bench) | Explosive (>50/mo) | 232 | 2026-08-21T14:39:45.271578 |
2604.06712v2 | Broken Quantum: A Systematic Formal Verification Study of Security Vulnerabilities Across the Open-Source Quantum Computing Simulator Ecosystem | 2026-04-08T06:07:37Z | [
"cs.CR",
"cs.SE",
"quant-ph"
] | ArXiv Standard | http://arxiv.org/licenses/nonexclusive-distrib/1.0/ | true | 60 | Internal R&D Only (Copyleft or Academic Terms) | ArXiv Standard Distribution; Repository License Unspecified | 0 | 0 | 1 | Proposes Broken Quantum, the first comprehensive formal security audit of the open-source quantum c... to enhance autonomous code synthesis, achieving Nine frameworks score 100/100 under all four scanners; Qiskit Aer,Cirq, tequila,.... | Dominik Blain | 1 | [
"Dominik Blain"
] | [
"Autonomous Software Engineering AI Lab"
] | http://arxiv.org/abs/2604.06712v2 | VERIFIED_LIVE | https://github.com/NVIDIA/cuda-quantum | [
"https://github.com/NVIDIA/cuda-quantum",
"https://github.com/PaddlePaddle/Quantum",
"https://github.com/amazon-braket/amazon-braket-default-simulator-python"
] | 1,100 | 0 | 2026-08-21 | Unspecified | 95.44 | Quantum computing simulators form the classical software foundation on which virtually all quantum algorithm research depends. We present Broken Quantum, the first comprehensive formal security audit of the open-source quantum computing simulator ecosystem. Applying COBALT QAI -- a four-module static analysis engine ba... | [
-0.0677729994058609,
0.017083000391721725,
-0.03544500097632408,
0.05641400068998337,
-0.02260800078511238,
-0.03660200163722038,
-0.09451600164175034,
-0.05261800065636635,
-0.05163700133562088,
-0.008306999690830708,
-0.028929000720381737,
-0.08447200059890747,
-0.0004910000134259462,
-0... | [
-0.11202000081539154,
0.005973000079393387,
-0.0715629979968071,
0.06255500018596649,
0.01462399959564209,
-0.05622300133109093,
-0.10735099762678146,
-0.04644300043582916,
-0.022071000188589096,
0.0011289999820291996,
-0.06016400083899498,
-0.05247199907898903,
0.0038030000869184732,
0.02... | AI Code Generation, Autonomous SWE Agents & Program Synthesis | Formal Logic & Theorem Proving Synthesis | [
"Python"
] | Specialized Autoregressive Code Transformer | [
"HumanEval & SWE-bench Functional Synthesis Benchmark"
] | [
"Quantitative Program Synthesis & Pass@k Metric"
] | [
{
"benchmark_name": "HumanEval & SWE-bench Functional Synthesis Benchmark",
"metric": "pass@1 / resolved",
"score": "Quantitative Program Synthesis & Pass@k Metric"
}
] | [
{
"paper_id": "2508.20907v1",
"title": "Quantum Verifiable Rewards for Post-Training Qiskit Code Assistant",
"cosine_sim": 0.6594
},
{
"paper_id": "2605.30358v1",
"title": "QASM-Eval: A Dataset to Train and Evaluate LLMs on OpenQASM-3 Beyond Quantum Cir",
"cosine_sim": 0.653
},
{
... | git clone https://github.com/NVIDIA/cuda-quantum && cd cuda-quantum && (pip install -e . || pip install -r requirements.txt) | Quantum computing simulators form the classical software foundation on which virtually all quantum algorithm research depends. | We present Broken Quantum, the first comprehensive formal security audit of the open-source quantum computing simulator ecosystem. | Nine frameworks score 100/100 under all four scanners; Qiskit Aer,Cirq, tequila, PennyLane, and 5 others score 0/100. | 7 | Syntax-Guided AST Modeling & Semantic Code Search | Explosive (>50/mo) | 232 | 2026-08-21T14:37:53.342921 |
2603.27836v1 | Q-Bridge: Code Translation for Quantum Machine Learning via LLMs | 2026-03-29T19:42:58Z | [
"quant-ph",
"cs.CL"
] | ArXiv Standard | http://arxiv.org/licenses/nonexclusive-distrib/1.0/ | true | 60 | Internal R&D Only (Copyleft or Academic Terms) | ArXiv Standard Distribution; Repository License Unspecified | 0 | 0 | 1 | Proposes However, the lack of standardized, high-quality datasets and robust translation frameworks... to enhance autonomous code synthesis, achieving Case studies further demonstrate that Q-Bridge can maintain deterministic correc.... | Runjia Zeng | 5 | [
"Runjia Zeng",
"Priyabrata Senapati",
"Ruixiang Tang",
"Dongfang Liu",
"Qiang Guan"
] | [
"Autonomous Software Engineering AI Lab"
] | http://arxiv.org/abs/2603.27836v1 | VERIFIED_LIVE | https://github.com/qiskit-community/qiskit-machine-learning | [
"https://github.com/qiskit-community/qiskit-machine-learning",
"https://github.com/runtsang/Q-Bridge"
] | 1,100 | 0 | 2026-08-21 | Unspecified | 95.04 | Large language models have recently shown potential in bridging the gap between classical machine learning and quantum machine learning. However, the lack of standardized, high-quality datasets and robust translation frameworks limits progress in this domain. We introduce Q-Bridge, an LLM-guided code translation framew... | [
-0.10869400203227997,
0.002500999951735139,
0.00900300033390522,
0.03485500067472458,
-0.036593999713659286,
-0.010719000361859798,
-0.027343999594449997,
-0.06559400260448456,
-0.025642000138759613,
0.005369000136852264,
-0.028217000886797905,
-0.05138000100851059,
-0.022810999304056168,
... | [
-0.09562999755144119,
0.005288000218570232,
-0.016697999089956284,
0.030386000871658325,
-0.0736600011587143,
-0.004759000148624182,
-0.09134600311517715,
-0.06313999742269516,
-0.014190999791026115,
-0.0358319990336895,
-0.06230299919843674,
-0.10384400188922882,
-0.02006799913942814,
0.0... | AI Code Generation, Autonomous SWE Agents & Program Synthesis | Code Translation & Transpilation Agent | [
"Multi-Language / Polyglot"
] | Specialized Autoregressive Code Transformer | [
"HumanEval & SWE-bench Functional Synthesis Benchmark"
] | [
"Quantitative Program Synthesis & Pass@k Metric"
] | [
{
"benchmark_name": "HumanEval & SWE-bench Functional Synthesis Benchmark",
"metric": "pass@1 / resolved",
"score": "Quantitative Program Synthesis & Pass@k Metric"
}
] | [
{
"paper_id": "2510.16779v1",
"title": "QuanBench: Benchmarking Quantum Code Generation with Large Language Models",
"cosine_sim": 0.8365
},
{
"paper_id": "2507.19562v1",
"title": "PennyCoder: Efficient Domain-Specific LLMs for PennyLane-Based Quantum Code Gene",
"cosine_sim": 0.7599
}... | git clone https://github.com/qiskit-community/qiskit-machine-learning && cd qiskit-machine-learning && (pip install -e . || pip install -r requirements.txt) | Large language models have recently shown potential in bridging the gap between classical machine learning and quantum machine learning. | However, the lack of standardized, high-quality datasets and robust translation frameworks limits progress in this domain. | Case studies further demonstrate that Q-Bridge can maintain deterministic correctness and also enable creative architectural exploration. | 4 | Formal Verification, Theorem Proving & Symbolic Logic | Explosive (>50/mo) | 232 | 2026-08-21T14:38:03.660043 |
2604.25850v4 | Agentic Harness Engineering: Observability-Driven Automatic Evolution of Coding-Agent Harnesses | 2026-04-28T16:55:02Z | [
"cs.CL",
"cs.SE"
] | ArXiv Standard | http://arxiv.org/licenses/nonexclusive-distrib/1.0/ | true | 60 | Internal R&D Only (Copyleft or Academic Terms) | ArXiv Standard Distribution; Repository License Unspecified | 0 | 0 | 1 | Proposes Agentic Harness Engineering (AHE), a closed loop that addresses these challenges through t... to enhance autonomous code synthesis, achieving The frozen harness transfers without re-evolution: on SWE-bench-verified it tops.... | Jiahang Lin | 11 | [
"Jiahang Lin",
"Shichun Liu",
"Chengjun Pan",
"Lizhi Lin",
"Shihan Dou",
"Zhiheng Xi",
"Xuanjing Huang",
"Hang Yan",
"Zhenhua Han",
"Tao Gui",
"Yu-Gang Jiang"
] | [
"Autonomous Software Engineering AI Lab"
] | http://arxiv.org/abs/2604.25850v4 | VERIFIED_LIVE | https://github.com/china-qijizhifeng/agentic-harness-engineering | [
"https://github.com/china-qijizhifeng/agentic-harness-engineering"
] | 841 | 0 | 2026-08-21 | Unspecified | 93.91 | Harnesses are now central to coding-agent performance, mediating how models interact with tools and execution environments. Yet harness engineering remains a manual craft, because automating it faces a heterogeneous action space across editable components, voluminous trajectories that bury actionable signal, and edits ... | [
-0.016603000462055206,
-0.07212900370359421,
-0.11056700348854065,
0.03334299847483635,
0.006903000175952911,
0.003220000071451068,
-0.006264999974519014,
-0.08462700247764587,
-0.07211799919605255,
0.009965999983251095,
0.03282700106501579,
-0.058733001351356506,
0.06430400162935257,
0.05... | [
-0.06134000048041344,
-0.08275599777698517,
-0.06697899848222733,
0.09191899746656418,
0.03554999828338623,
-0.0014609999489039183,
-0.03574400022625923,
-0.046163998544216156,
-0.022622000426054,
0.028694000095129013,
-0.050367001444101334,
-0.08084099739789963,
0.014115000143647194,
0.05... | AI Code Generation, Autonomous SWE Agents & Program Synthesis | Autonomous Repo-Level SWE Agent (SWE-bench) | [
"Multi-Language / Polyglot"
] | GPT-4o / Codex Specialized Coding LLM | [
"SWE-bench (Autonomous Software Engineering)"
] | [
"Quantitative Program Synthesis & Pass@k Metric"
] | [
{
"benchmark_name": "SWE-bench (Autonomous Software Engineering)",
"metric": "pass@1 / resolved",
"score": "Quantitative Program Synthesis & Pass@k Metric"
}
] | [
{
"paper_id": "2606.14249v3",
"title": "HarnessX: A Composable, Adaptive, and Evolvable Agent Harness Foundry",
"cosine_sim": 0.771
},
{
"paper_id": "2608.01918v1",
"title": "HarnessCompass: Guiding Automatic Harness Evolution toward Generalizable and Eff",
"cosine_sim": 0.7373
},
{
... | git clone https://github.com/china-qijizhifeng/agentic-harness-engineering && cd agentic-harness-engineering && (pip install -e . || pip install -r requirements.txt) | Harnesses are now central to coding-agent performance, mediating how models interact with tools and execution environments. | We introduce Agentic Harness Engineering (AHE), a closed loop that addresses these challenges through three matched observability pillars: (1) component observability gives every editable harness component a file-level representation so the action space is explicit and revertible; (2) experience observability distills ... | The frozen harness transfers without re-evolution: on SWE-bench-verified it tops aggregate success at 12% fewer tokens than the seed, and on Terminal-Bench 2 it yields +5.1 to +10.1pp cross-family gains across three alternate model families, indicating the evolved components encode general engineering experience rather... | 1 | Autonomous Repo-Level Software Engineering Agents (SWE-bench) | Explosive (>50/mo) | 232 | 2026-08-21T14:37:21.733063 |
2602.04998v2 | Learning Rate Matters: Vanilla LoRA May Suffice for LLM Fine-tuning | 2026-02-04T19:36:20Z | [
"cs.LG",
"cs.AI",
"cs.CL"
] | ArXiv Standard | http://arxiv.org/licenses/nonexclusive-distrib/1.0/ | true | 60 | Internal R&D Only (Copyleft or Academic Terms) | ArXiv Standard Distribution; Repository License Unspecified | 0 | 0 | 1 | Proposes Building on this paradigm, recent studies have proposed alternative initialization strateg... to enhance autonomous code synthesis, achieving These results suggest that vanilla LoRA remains a competitive baseline and that.... | Yu-Ang Lee | 4 | [
"Yu-Ang Lee",
"Ching-Yun Ko",
"Pin-Yu Chen",
"Mi-Yen Yeh"
] | [
"Autonomous Software Engineering AI Lab"
] | http://arxiv.org/abs/2602.04998v2 | VERIFIED_LIVE | https://github.com/AGI-Edgerunners/LLM-Adapters | [
"https://github.com/AGI-Edgerunners/LLM-Adapters",
"https://github.com/Leopold1423/non_zero_lora-icml25"
] | 1,200 | 0 | 2026-08-21 | Unspecified | 93.67 | Low-Rank Adaptation (LoRA) is the prevailing approach for efficient large language model (LLM) fine-tuning. Building on this paradigm, recent studies have proposed alternative initialization strategies, architectural modifications, and optimization adjustments, reporting substantial improvements over vanilla LoRA. Howe... | [
0.022252000868320465,
-0.08620800077915192,
-0.003326999954879284,
-0.04667200148105621,
0.045517001301050186,
-0.034324001520872116,
-0.04768700152635574,
-0.017667999491095543,
-0.06247999891638756,
-0.014883000403642654,
-0.04298099875450134,
0.024498000741004944,
-0.0385809987783432,
-... | [
-0.019453000277280807,
-0.115106001496315,
0.031098000705242157,
0.025613000616431236,
0.055479999631643295,
0.016064999625086784,
-0.06276900321245193,
0.017741000279784203,
-0.04494699835777283,
-0.1005759984254837,
-0.04458799958229065,
-0.016858000308275223,
-0.0034000000450760126,
-0.... | AI Code Generation, Autonomous SWE Agents & Program Synthesis | Neural Program Synthesis & Generation Architecture | [
"Multi-Language / Polyglot"
] | Specialized Autoregressive Code Transformer | [
"HumanEval & SWE-bench Functional Synthesis Benchmark"
] | [
"Quantitative Program Synthesis & Pass@k Metric"
] | [
{
"benchmark_name": "HumanEval & SWE-bench Functional Synthesis Benchmark",
"metric": "pass@1 / resolved",
"score": "Quantitative Program Synthesis & Pass@k Metric"
}
] | [
{
"paper_id": "2605.16470v1",
"title": "Strategic Over-Parameterization for Generalizable Low-Rank Adaptation",
"cosine_sim": 0.789
},
{
"paper_id": "2602.20727v1",
"title": "ID-LoRA: Efficient Low-Rank Adaptation Inspired by Matrix Interpolative Decompos",
"cosine_sim": 0.7721
},
{
... | git clone https://github.com/AGI-Edgerunners/LLM-Adapters && cd LLM-Adapters && (pip install -e . || pip install -r requirements.txt) | Low-Rank Adaptation (LoRA) is the prevailing approach for efficient large language model (LLM) fine-tuning. | Building on this paradigm, recent studies have proposed alternative initialization strategies, architectural modifications, and optimization adjustments, reporting substantial improvements over vanilla LoRA. | These results suggest that vanilla LoRA remains a competitive baseline and that improvements reported under a single training configuration may not reflect consistent methodological advantages. | 4 | Formal Verification, Theorem Proving & Symbolic Logic | Explosive (>50/mo) | 232 | 2026-08-21T14:38:51.383103 |
2507.03616v2 | EvoAgentX: An Automated Framework for Evolving Agentic Workflows | 2025-07-04T14:43:10Z | [
"cs.AI"
] | ArXiv Standard | http://arxiv.org/licenses/nonexclusive-distrib/1.0/ | true | 60 | Internal R&D Only (Copyleft or Academic Terms) | ArXiv Standard Distribution; Repository License Unspecified | 0 | 0 | 1 | Proposes However, existing MAS frameworks often require manual workflow configuration and lack nati... to enhance autonomous code synthesis, achieving Experimental results show that EvoAgentX consistently achieves significant perfo.... | Yingxu Wang | 4 | [
"Yingxu Wang",
"Siwei Liu",
"Jinyuan Fang",
"Zaiqiao Meng"
] | [
"Autonomous Software Engineering AI Lab"
] | http://arxiv.org/abs/2507.03616v2 | VERIFIED_LIVE | https://github.com/EvoAgentX/EvoAgentX | [
"https://github.com/EvoAgentX/EvoAgentX"
] | 3,200 | 0 | 2026-08-21 | Unspecified | 93.59 | Multi-agent systems (MAS) have emerged as a powerful paradigm for orchestrating large language models (LLMs) and specialized tools to collaboratively address complex tasks. However, existing MAS frameworks often require manual workflow configuration and lack native support for dynamic evolution and performance optimiza... | [
-0.03241100162267685,
-0.02730100043118,
-0.06531199812889099,
-0.06677400320768356,
0.0271380003541708,
-0.07510200142860413,
0.039969999343156815,
-0.03877200186252594,
-0.012570999562740326,
-0.031863000243902206,
-0.06595399975776672,
-0.0837090015411377,
0.04027799889445305,
0.0320519... | [
-0.045228999108076096,
-0.03807999938726425,
-0.07401300221681595,
0.006264000199735165,
-0.009503999724984169,
-0.07073000073432922,
-0.031610000878572464,
-0.019061999395489693,
-0.022019000723958015,
-0.05002199858427048,
-0.073294997215271,
-0.12296000123023987,
0.057360000908374786,
-... | AI Code Generation, Autonomous SWE Agents & Program Synthesis | Neural Program Synthesis & Generation Architecture | [
"Multi-Language / Polyglot"
] | Specialized Autoregressive Code Transformer | [
"MBPP / MBPP+ (Multi-Turn Python Benchmarking)"
] | [
"Quantitative Program Synthesis & Pass@k Metric"
] | [
{
"benchmark_name": "MBPP / MBPP+ (Multi-Turn Python Benchmarking)",
"metric": "pass@1 / resolved",
"score": "Quantitative Program Synthesis & Pass@k Metric"
}
] | [
{
"paper_id": "2602.06511v4",
"title": "EvoMAS: Evolutionary Generation of Multi-Agent Systems",
"cosine_sim": 0.8219
},
{
"paper_id": "2604.17708v2",
"title": "Co-evolving Agent Architectures and Interpretable Reasoning for Automated Optimi",
"cosine_sim": 0.8072
},
{
"paper_id"... | git clone https://github.com/EvoAgentX/EvoAgentX && cd EvoAgentX && (pip install -e . || pip install -r requirements.txt) | Multi-agent systems (MAS) have emerged as a powerful paradigm for orchestrating large language models (LLMs) and specialized tools to collaboratively address complex tasks. | However, existing MAS frameworks often require manual workflow configuration and lack native support for dynamic evolution and performance optimization. | Experimental results show that EvoAgentX consistently achieves significant performance improvements, including a 7.44% increase in HotPotQA F1, a 10.00% improvement in MBPP pass@1, a 10.00% gain in MATH solve accuracy, and an overall accuracy improvement of up to 20.00% on GAIA. | 8 | Code Foundation Models & Specialized Instruction Distillation | Explosive (>50/mo) | 232 | 2026-08-21T14:41:26.551915 |
2603.25930v2 | AVDA: Autonomous Vibe Detection Authoring for Cybersecurity | 2026-03-26T21:52:33Z | [
"cs.CR",
"cs.SE"
] | ArXiv Standard | http://arxiv.org/licenses/nonexclusive-distrib/1.0/ | true | 60 | Internal R&D Only (Copyleft or Academic Terms) | ArXiv Standard Distribution; Repository License Unspecified | 0 | 0 | 1 | Proposes we introduce AVDA, a framework that leverages the Model Context Protocol (MCP) to automate... to enhance autonomous code synthesis, achieving Our results show that Agentic workflows achieve a 19% improvement in overall sim.... | Fatih Bulut | 4 | [
"Fatih Bulut",
"Carlo DePaolis",
"Raghav Batta",
"Anjali Mangal"
] | [
"Autonomous Software Engineering AI Lab"
] | http://arxiv.org/abs/2603.25930v2 | VERIFIED_LIVE | https://github.com/palantir/alerting-detection-strategy-framework | [
"https://github.com/palantir/alerting-detection-strategy-framework"
] | 897 | 0 | 2026-08-21 | Unspecified | 93.15 | With the rapid advancement of AI in code generation, cybersecurity detection engineering faces new opportunities to automate traditionally manual processes. Detection authoring - the practice of creating executable logic that identifies malicious activities from security telemetry - is hindered by fragmented code acros... | [
-0.01671699993312359,
-0.04770199954509735,
-0.059932999312877655,
-0.05915699899196625,
0.07267899811267853,
-0.015697000548243523,
0.08292800188064575,
-0.09399399906396866,
0.013050000183284283,
-0.09179399907588959,
-0.012368000112473965,
-0.03161799907684326,
0.006754999980330467,
-0.... | [
-0.06911399960517883,
-0.0582519993185997,
-0.08185099810361862,
-0.019892999902367592,
0.05724500119686127,
-0.05321099981665611,
0.059108998626470566,
-0.055670998990535736,
0.013148999772965908,
-0.029148999601602554,
-0.02733900025486946,
-0.041436001658439636,
0.08530600368976593,
-0.... | AI Code Generation, Autonomous SWE Agents & Program Synthesis | Multi-Turn Interactive Pair Programmer | [
"Multi-Language / Polyglot"
] | GPT-4o / Codex Specialized Coding LLM | [
"HumanEval & SWE-bench Functional Synthesis Benchmark"
] | [
"achieve a 19%"
] | [
{
"benchmark_name": "HumanEval & SWE-bench Functional Synthesis Benchmark",
"metric": "pass@1 / resolved",
"score": "achieve a 19%"
}
] | [
{
"paper_id": "2607.27923v1",
"title": "The Case for Vibe Modeling: A Missing Step in AI-Based Trustworthy Software Deve",
"cosine_sim": 0.6519
},
{
"paper_id": "2510.11516v1",
"title": "Cracking CodeWhisperer: Analyzing Developers' Interactions and Patterns During P",
"cosine_sim": 0.62... | git clone https://github.com/palantir/alerting-detection-strategy-framework && cd alerting-detection-strategy-framework && (pip install -e . || pip install -r requirements.txt) | With the rapid advancement of AI in code generation, cybersecurity detection engineering faces new opportunities to automate traditionally manual processes. | In this paper, we introduce AVDA, a framework that leverages the Model Context Protocol (MCP) to automate detection authoring by integrating organizational context - existing detections, telemetry schemas, and style guides - into AI-assisted code generation. | Our results show that Agentic workflows achieve a 19% improvement in overall similarity score over Baseline approaches, while Sequential workflows attain 87% of Agentic quality at 40x lower token cost. | 1 | Autonomous Repo-Level Software Engineering Agents (SWE-bench) | Explosive (>50/mo) | 232 | 2026-08-21T14:38:01.819501 |
2510.01174v1 | Code2Video: A Code-centric Paradigm for Educational Video Generation | 2025-10-01T17:56:48Z | [
"cs.CV",
"cs.AI",
"cs.CL",
"cs.HC",
"cs.MM"
] | ArXiv Standard | http://arxiv.org/licenses/nonexclusive-distrib/1.0/ | true | 60 | Internal R&D Only (Copyleft or Academic Terms) | ArXiv Standard Distribution; Repository License Unspecified | 0 | 0 | 1 | Proposes In this work, we propose Code2Video, a code-centric agent framework for generating educati... to enhance autonomous code synthesis, achieving Our results demonstrate the potential of Code2Video as a scalable, interpretable.... | Yanzhe Chen | 3 | [
"Yanzhe Chen",
"Kevin Qinghong Lin",
"Mike Zheng Shou"
] | [
"Autonomous Software Engineering AI Lab"
] | http://arxiv.org/abs/2510.01174v1 | VERIFIED_LIVE | https://github.com/showlab/Code2Video | [
"https://github.com/showlab/Code2Video"
] | 2,000 | 0 | 2026-08-21 | Unspecified | 93.06 | While recent generative models advance pixel-space video synthesis, they remain limited in producing professional educational videos, which demand disciplinary knowledge, precise visual structures, and coherent transitions, limiting their applicability in educational scenarios. Intuitively, such requirements are better... | [
-0.05078500136733055,
-0.0949539989233017,
-0.01131999958306551,
-0.06633999943733215,
0.04671100154519081,
-0.0024739999789744616,
-0.06014600023627281,
0.010301000438630581,
0.011806000024080276,
-0.014456000179052353,
0.01796800084412098,
-0.03805999830365181,
-0.06594999879598618,
-0.0... | [
-0.051959000527858734,
-0.07445099949836731,
-0.029566999524831772,
-0.0034539999905973673,
0.032597001641988754,
-0.029518000781536102,
-0.10095199942588806,
-0.004147999919950962,
0.002068999921903014,
-0.027272000908851624,
-0.0452749989926815,
-0.0711899995803833,
0.011134999804198742,
... | AI Code Generation, Autonomous SWE Agents & Program Synthesis | Neural Program Synthesis & Generation Architecture | [
"Python"
] | Specialized Autoregressive Code Transformer | [
"HumanEval & SWE-bench Functional Synthesis Benchmark"
] | [
"achieving 40%"
] | [
{
"benchmark_name": "HumanEval & SWE-bench Functional Synthesis Benchmark",
"metric": "pass@1 / resolved",
"score": "achieving 40%"
}
] | [
{
"paper_id": "2605.15585v1",
"title": "See Before You Code: Learning Visual Priors for Spatially Aware Educational Anim",
"cosine_sim": 0.7528
},
{
"paper_id": "2603.20633v3",
"title": "Seed1.8 Model Card: Towards Generalized Real-World Agency",
"cosine_sim": 0.696
},
{
"paper_i... | git clone https://github.com/showlab/Code2Video && cd Code2Video && (pip install -e . || pip install -r requirements.txt) | While recent generative models advance pixel-space video synthesis, they remain limited in producing professional educational videos, which demand disciplinary knowledge, precise visual structures, and coherent transitions, limiting their applicability in educational scenarios. | In this work, we propose Code2Video, a code-centric agent framework for generating educational videos via executable Python code. | Our results demonstrate the potential of Code2Video as a scalable, interpretable, and controllable approach, achieving 40% improvement over direct code generation and producing videos comparable to human-crafted tutorials. | 3 | Multi-Language Code Translation & Cross-Platform Migration | Explosive (>50/mo) | 232 | 2026-08-21T14:40:23.795339 |
2604.01193v2 | Embarrassingly Simple Self-Distillation Improves Code Generation | 2026-04-01T17:39:50Z | [
"cs.CL"
] | ArXiv Standard | http://arxiv.org/licenses/nonexclusive-distrib/1.0/ | true | 60 | Internal R&D Only (Copyleft or Academic Terms) | ArXiv Standard Distribution; Repository License Unspecified | 0 | 0 | 1 | Proposes We answer in the affirmative with simple self-distillation (SSD): sample solutions from th... to enhance autonomous code synthesis, achieving To understand why such a simple method can work, we trace these gains to a preci.... | Ruixiang Zhang | 6 | [
"Ruixiang Zhang",
"Richard He Bai",
"Huangjie Zheng",
"Navdeep Jaitly",
"Ronan Collobert",
"Yizhe Zhang"
] | [
"Autonomous Software Engineering AI Lab"
] | http://arxiv.org/abs/2604.01193v2 | VERIFIED_LIVE | https://github.com/apple/ml-ssd | [
"https://github.com/apple/ml-ssd"
] | 801 | 0 | 2026-08-21 | Unspecified | 92.4 | Can a large language model (LLM) improve at code generation using only its own raw outputs, without a verifier, a teacher model, or reinforcement learning? We answer in the affirmative with simple self-distillation (SSD): sample solutions from the model with certain temperature and truncation configurations, then fine-... | [
-0.06137999892234802,
-0.03136400133371353,
0.06596700102090836,
0.031001999974250793,
-0.0628490000963211,
-0.09229700267314911,
-0.004920000210404396,
-0.012413999997079372,
-0.014673000201582909,
-0.07058899849653244,
-0.025714000687003136,
-0.020633000880479813,
0.01396500039845705,
-0... | [
-0.07547999918460846,
-0.07775600254535675,
0.04948800057172775,
0.039556000381708145,
-0.004497000016272068,
-0.029469000175595284,
-0.020478999242186546,
-0.034961000084877014,
0.017764000222086906,
-0.009646999649703503,
-0.030108999460935593,
-0.04782500118017197,
0.04872800037264824,
... | AI Code Generation, Autonomous SWE Agents & Program Synthesis | Neural Program Synthesis & Generation Architecture | [
"Multi-Language / Polyglot"
] | Specialized Autoregressive Code Transformer | [
"LiveCodeBench (Contamination-Free SOTA Evaluation)"
] | [
"55.3% pass@1"
] | [
{
"benchmark_name": "LiveCodeBench (Contamination-Free SOTA Evaluation)",
"metric": "pass@1 / resolved",
"score": "55.3% pass@1"
}
] | [
{
"paper_id": "2509.07858v1",
"title": "SCoder: Iterative Self-Distillation for Bootstrapping Small-Scale Data Synthesiz",
"cosine_sim": 0.7059
},
{
"paper_id": "2605.22675v1",
"title": "Self-Policy Distillation via Capability-Selective Subspace Projection",
"cosine_sim": 0.6927
},
{... | git clone https://github.com/apple/ml-ssd && cd ml-ssd && (pip install -e . || pip install -r requirements.txt) | Can a large language model (LLM) improve at code generation using only its own raw outputs, without a verifier, a teacher model, or reinforcement learning? | We answer in the affirmative with simple self-distillation (SSD): sample solutions from the model with certain temperature and truncation configurations, then fine-tune on those samples with standard supervised fine-tuning. | To understand why such a simple method can work, we trace these gains to a precision-exploration conflict in LLM decoding and show that SSD reshapes token distributions in a context-dependent way, suppressing distractor tails where precision matters while preserving useful diversity where exploration matters. | 4 | Formal Verification, Theorem Proving & Symbolic Logic | Explosive (>50/mo) | 232 | 2026-08-21T14:37:57.683375 |
2605.01124v1 | Practical Formal Verification for MLIR Programs | 2026-05-01T21:51:21Z | [
"cs.PL"
] | ArXiv Standard | http://arxiv.org/licenses/nonexclusive-distrib/1.0/ | true | 60 | Internal R&D Only (Copyleft or Academic Terms) | ArXiv Standard Distribution; Repository License Unspecified | 0 | 0 | 1 | Proposes Optimizations, including those implemented manually by a user and those target-specific an... to enhance autonomous code synthesis, achieving Optimizations, including those implemented manually by a user and those target-s.... | Emily Tucker | 5 | [
"Emily Tucker",
"Louis-Noël Pouchet",
"Erika Hunhoff",
"Stephen Neuendorffer",
"Erwei Wang"
] | [
"Autonomous Software Engineering AI Lab"
] | http://arxiv.org/abs/2605.01124v1 | VERIFIED_LIVE | https://github.com/Xilinx/mlir-aie | [
"https://github.com/Xilinx/mlir-aie",
"https://github.com/Xilinx/mlir-air",
"https://github.com/axolotls73/PEQC-MLIR"
] | 680 | 0 | 2026-08-21 | Unspecified | 92.18 | Optimizing compilers have become a cornerstone for high-performance program generation in research and industry. Optimizations, including those implemented manually by a user and those target-specific and non-target-specific, are used to transform programs to achieve good performance. Although these optimizations are n... | [
-0.07677999883890152,
-0.014891000464558601,
-0.029099000617861748,
-0.07695599645376205,
0.06080399826169014,
-0.0793209969997406,
0.006397000048309565,
0.0011520000407472253,
-0.08097700029611588,
0.0027540000155568123,
-0.028303999453783035,
-0.051927000284194946,
0.06470199674367905,
-... | [
-0.0498180016875267,
0.03381400182843208,
-0.029805999249219894,
-0.056728001683950424,
0.0403439998626709,
-0.11427000164985657,
-0.027701999992132187,
0.0026340000331401825,
-0.0734890028834343,
-0.02059200033545494,
-0.04981299862265587,
-0.09137099981307983,
0.038676999509334564,
-0.01... | AI Code Generation, Autonomous SWE Agents & Program Synthesis | Formal Logic & Theorem Proving Synthesis | [
"Multi-Language / Polyglot"
] | Specialized Autoregressive Code Transformer | [
"HumanEval & SWE-bench Functional Synthesis Benchmark"
] | [
"Quantitative Program Synthesis & Pass@k Metric"
] | [
{
"benchmark_name": "HumanEval & SWE-bench Functional Synthesis Benchmark",
"metric": "pass@1 / resolved",
"score": "Quantitative Program Synthesis & Pass@k Metric"
}
] | [
{
"paper_id": "2605.27051v1",
"title": "ConVer: Using Contracts and Loop Invariant Synthesis for Scalable Formal Softwar",
"cosine_sim": 0.7153
},
{
"paper_id": "2604.16584v1",
"title": "Certified Program Synthesis with a Multi-Modal Verifier",
"cosine_sim": 0.7107
},
{
"paper_id... | git clone https://github.com/Xilinx/mlir-aie && cd mlir-aie && (pip install -e . || pip install -r requirements.txt) | Optimizing compilers have become a cornerstone for high-performance program generation in research and industry. | Optimizations, including those implemented manually by a user and those target-specific and non-target-specific, are used to transform programs to achieve good performance. | Optimizations, including those implemented manually by a user and those target-specific and non-target-specific, are used to transform programs to achieve good performance. | 6 | Static Analysis, Vulnerability Detection & Automated Security Patching | Explosive (>50/mo) | 232 | 2026-08-21T14:37:20.845203 |
2506.18088v2 | RoboTwin 2.0: A Scalable Data Generator and Benchmark with Strong Domain Randomization for Robust Bimanual Robotic Manipulation | 2025-06-22T16:26:53Z | [
"cs.RO",
"cs.AI",
"cs.CL",
"cs.CV",
"cs.MA"
] | ArXiv Standard | http://arxiv.org/licenses/nonexclusive-distrib/1.0/ | true | 60 | Internal R&D Only (Copyleft or Academic Terms) | ArXiv Standard Distribution; Repository License Unspecified | 0 | 0 | 1 | Proposes RoboTwin 2.0, a scalable framework for automated, large-scale generation of diverse and re... to enhance autonomous code synthesis, achieving These results highlight the effectiveness of RoboTwin 2.0 in strengthening sim-t.... | Tianxing Chen | 26 | [
"Tianxing Chen",
"Zanxin Chen",
"Baijun Chen",
"Zijian Cai",
"Yibin Liu",
"Zixuan Li",
"Qiwei Liang",
"Xianliang Lin",
"Yiheng Ge",
"Zhenyu Gu",
"Weiliang Deng",
"Yubin Guo",
"Tian Nian",
"Xuanbing Xie",
"Qiangyu Chen",
"Kailun Su",
"Tianling Xu",
"Guodong Liu",
"Mengkang Hu",
... | [
"Autonomous Software Engineering AI Lab"
] | http://arxiv.org/abs/2506.18088v2 | VERIFIED_LIVE | https://github.com/robotwin-Platform/robotwin | [
"https://github.com/robotwin-Platform/robotwin"
] | 2,800 | 0 | 2026-08-21 | Unspecified | 91.95 | Simulation-based data synthesis has emerged as a powerful paradigm for advancing real-world robotic manipulation. Yet existing datasets remain insufficient for robust bimanual manipulation due to (1) the lack of scalable task generation methods and (2) oversimplified simulation environments. We present RoboTwin 2.0, a ... | [
-0.013659999705851078,
-0.06264600157737732,
0.030612999573349953,
-0.007288999855518341,
-0.002377999946475029,
-0.03478299826383591,
-0.03236699849367142,
-0.07414600253105164,
-0.033629998564720154,
0.03891599923372269,
0.0022120000794529915,
-0.009189999662339687,
0.03658000007271767,
... | [
-0.02971000038087368,
-0.06216299906373024,
0.044186998158693314,
0.023144999518990517,
0.013166000135242939,
-0.08572500199079514,
-0.03776700049638748,
-0.04964900016784668,
-0.00912300031632185,
0.015843000262975693,
-0.03583800047636032,
-0.0996059998869896,
0.05487100034952164,
0.0310... | AI Code Generation, Autonomous SWE Agents & Program Synthesis | Neural Program Synthesis & Generation Architecture | [
"Multi-Language / Polyglot"
] | Specialized Autoregressive Code Transformer | [
"HumanEval & SWE-bench Functional Synthesis Benchmark"
] | [
"achieves a 367%"
] | [
{
"benchmark_name": "HumanEval & SWE-bench Functional Synthesis Benchmark",
"metric": "pass@1 / resolved",
"score": "achieves a 367%"
}
] | [
{
"paper_id": "2509.24160v1",
"title": "Memory Transfer Planning: LLM-driven Context-Aware Code Adaptation for Robot Man",
"cosine_sim": 0.7201
},
{
"paper_id": "2607.23784v1",
"title": "A Few Words Go a Long Way: Language Guided Robot Policy Synthesis",
"cosine_sim": 0.6904
},
{
... | git clone https://github.com/robotwin-Platform/robotwin && cd robotwin && (pip install -e . || pip install -r requirements.txt) | Simulation-based data synthesis has emerged as a powerful paradigm for advancing real-world robotic manipulation. | We present RoboTwin 2.0, a scalable framework for automated, large-scale generation of diverse and realistic data, together with unified evaluation protocols for dual-arm manipulation. | These results highlight the effectiveness of RoboTwin 2.0 in strengthening sim-to-real transfer and robustness to environmental variations. | 3 | Multi-Language Code Translation & Cross-Platform Migration | Explosive (>50/mo) | 232 | 2026-08-21T14:41:33.806596 |
2606.23870v2 | ESBMC-PLC+: A Unified IEC 61131-3 Formal Verification Framework as a PLCverif Successor | 2026-06-22T19:13:02Z | [
"cs.PL",
"cs.CL",
"cs.SE"
] | ArXiv Standard | http://arxiv.org/licenses/nonexclusive-distrib/1.0/ | true | 80 | Enterprise Safe (Commercial Training & Deployment Allowed) | ArXiv Standard Distribution | 0 | 0 | 1 | Proposes This paper presents ESBMC-PLC+, a unified framework that closes both gaps: (1) an ST/SCL f... to enhance autonomous code synthesis, achieving Against nuXmv's BDD backend, ESBMC-PLC+ is 400-2,000x faster on timer programs a.... | Pierre Dantas | 3 | [
"Pierre Dantas",
"Lucas Cordeiro",
"Waldir Junior"
] | [
"Autonomous Software Engineering AI Lab"
] | http://arxiv.org/abs/2606.23870v2 | VERIFIED_LIVE | https://github.com/esbmc/esbmc | [
"https://github.com/esbmc/esbmc",
"https://github.com/pierredantas/esbmc-plcplus-artifact"
] | 513 | 144 | 2026-08-21 | NOASSERTION | 91.82 | PLCverif is the most mature open-source platform for PLC formal verification, developed at CERN and in production use since 2019. Yet it has two fundamental limitations: no support for Ladder Diagram (LD) programs, the dominant PLC notation, and reliance on CBMC as its primary backend, which restricts verification to b... | [
-0.07365600019693375,
0.004639000166207552,
-0.03644600138068199,
-0.08230599761009216,
0.021412000060081482,
0.02732899971306324,
-0.03375300019979477,
-0.007294000126421452,
0.00019099999917671084,
0.022888999432325363,
-0.015588999725878239,
-0.06801000237464905,
-0.01056000031530857,
-... | [
-0.08158499747514725,
-0.020419999957084656,
-0.059324998408555984,
-0.07347500324249268,
-0.005572999827563763,
0.012396000325679779,
-0.07287999987602234,
0.072332002222538,
-0.006225000135600567,
0.039007000625133514,
-0.00978699978441,
-0.05140699818730354,
0.023138999938964844,
-0.048... | AI Code Generation, Autonomous SWE Agents & Program Synthesis | Formal Logic & Theorem Proving Synthesis | [
"Multi-Language / Polyglot"
] | Specialized Autoregressive Code Transformer | [
"HumanEval & SWE-bench Functional Synthesis Benchmark"
] | [
"Quantitative Program Synthesis & Pass@k Metric"
] | [
{
"benchmark_name": "HumanEval & SWE-bench Functional Synthesis Benchmark",
"metric": "pass@1 / resolved",
"score": "Quantitative Program Synthesis & Pass@k Metric"
}
] | [
{
"paper_id": "2607.08417v1",
"title": "Detecting Ladder Logic Bombs in IEC 61131-3 PLC Programs using ESBMC-PLC+: A For",
"cosine_sim": 0.6952
},
{
"paper_id": "2605.26169v2",
"title": "ESBMC: A Survey of Its Evolution, Integration, and Future Directions in Formal S",
"cosine_sim": 0.65... | git clone https://github.com/esbmc/esbmc && cd esbmc && (pip install -e . || pip install -r requirements.txt) | PLCverif is the most mature open-source platform for PLC formal verification, developed at CERN and in production use since 2019. | This paper presents ESBMC-PLC+, a unified framework that closes both gaps: (1) an ST/SCL frontend via the MATIEC IEC 61131-3 compiler, routing C-compiled ST to ESBMC with nondeterministic input modeling and YAML property injection; (2) function block state semantics for graphical LD, extending the DFS resolver to model... | Against nuXmv's BDD backend, ESBMC-PLC+ is 400-2,000x faster on timer programs and completes proofs where nuXmv BDD times out at 120s. | 6 | Static Analysis, Vulnerability Detection & Automated Security Patching | Explosive (>50/mo) | 232 | 2026-08-21T14:36:10.177396 |
2605.26169v2 | ESBMC: A Survey of Its Evolution, Integration, and Future Directions in Formal Software Verification | 2026-05-25T00:18:27Z | [
"cs.SE",
"cs.LO"
] | ArXiv Standard | http://arxiv.org/licenses/nonexclusive-distrib/1.0/ | true | 80 | Enterprise Safe (Commercial Training & Deployment Allowed) | ArXiv Standard Distribution | 0 | 0 | 1 | Proposes Since its first publication in 2009, ESBMC has undergone persistent evolution: expanding i... to enhance autonomous code synthesis, achieving We synthesize its economic impact - over GBP 9.3 million and EUR 4.98 million in.... | Pierre Dantas | 3 | [
"Pierre Dantas",
"Lucas Cordeiro",
"Waldir Junior"
] | [
"Autonomous Software Engineering AI Lab"
] | http://arxiv.org/abs/2605.26169v2 | VERIFIED_LIVE | https://github.com/esbmc/esbmc | [
"https://github.com/esbmc/esbmc",
"https://github.com/esbmc/esbmc-ai"
] | 513 | 144 | 2026-08-21 | NOASSERTION | 90.7 | The Efficient SMT-Based Context-Bounded Model Checker (ESBMC) has grown from a research prototype for verifying embedded ANSI-C software into one of the most versatile and industrially capable formal verification platforms available today. Since its first publication in 2009, ESBMC has undergone persistent evolution: e... | [
-0.08865900337696075,
-0.018667999655008316,
-0.04574799910187721,
-0.05559699982404709,
0.03310700133442879,
-0.06556899845600128,
0.010002999566495419,
0.027156999334692955,
-0.059953998774290085,
0.017305999994277954,
-0.09525199979543686,
-0.06401100009679794,
0.06242100149393082,
-0.0... | [
-0.1192770004272461,
-0.020238999277353287,
-0.03983199968934059,
-0.016652999445796013,
0.10026299953460693,
-0.06135400012135506,
-0.00773600023239851,
0.043598998337984085,
-0.013532999902963638,
0.014094999991357327,
-0.08782800287008286,
-0.04623899981379509,
0.0964289978146553,
0.003... | AI Code Generation, Autonomous SWE Agents & Program Synthesis | Test-Driven Code Repair & Self-Healing | [
"Multi-Language / Polyglot"
] | Specialized Autoregressive Code Transformer | [
"HumanEval & SWE-bench Functional Synthesis Benchmark"
] | [
"Quantitative Program Synthesis & Pass@k Metric"
] | [
{
"benchmark_name": "HumanEval & SWE-bench Functional Synthesis Benchmark",
"metric": "pass@1 / resolved",
"score": "Quantitative Program Synthesis & Pass@k Metric"
}
] | [
{
"paper_id": "2507.04857v1",
"title": "Supporting Software Formal Verification with Large Language Models: An Experimen",
"cosine_sim": 0.7612
},
{
"paper_id": "2604.01851v1",
"title": "Can Large Language Models Model Programs Formally?",
"cosine_sim": 0.6844
},
{
"paper_id": "2... | git clone https://github.com/esbmc/esbmc && cd esbmc && (pip install -e . || pip install -r requirements.txt) | The Efficient SMT-Based Context-Bounded Model Checker (ESBMC) has grown from a research prototype for verifying embedded ANSI-C software into one of the most versatile and industrially capable formal verification platforms available today. | Since its first publication in 2009, ESBMC has undergone persistent evolution: expanding its verification techniques, widening its language support to nine front-ends, integrating industrial-strength SMT solvers, and - most recently - coupling with Large Language Models (LLMs) and autonomous AI agents. | We synthesize its economic impact - over GBP 9.3 million and EUR 4.98 million in confirmed public research funding, the VeriBee spin-off, and a defense industrial deployment at Lockheed Martin - and conclude with a structured agenda of open challenges spanning scalability, neurosymbolic verification, counterexample int... | 6 | Static Analysis, Vulnerability Detection & Automated Security Patching | Explosive (>50/mo) | 232 | 2026-08-21T14:36:43.358753 |
2605.17174v1 | Beyond Execution: Static-Analysis Rewards and Hint-Conditioned Diffusion RL for Code Generation | 2026-05-16T22:18:04Z | [
"cs.SE",
"cs.AI"
] | ArXiv Standard | http://arxiv.org/licenses/nonexclusive-distrib/1.0/ | true | 60 | Internal R&D Only (Copyleft or Academic Terms) | ArXiv Standard Distribution; Repository License Unspecified | 0 | 0 | 1 | Proposes we present a systematic empirical study of RL post-training for diffusion-based code gener... to enhance autonomous code synthesis, achieving Across HumanEval, MBPP, and LiveCodeBench, we find that static checking is the s.... | Shuyin Ouyang | 5 | [
"Shuyin Ouyang",
"Zhaozhi Qian",
"Faroq AL-Tam",
"Muhammad AL-Qurishi",
"Jie M. Zhang"
] | [
"Autonomous Software Engineering AI Lab"
] | http://arxiv.org/abs/2605.17174v1 | VERIFIED_LIVE | https://github.com/Gen-Verse/dLLM-RL | [
"https://github.com/Gen-Verse/dLLM-RL"
] | 520 | 0 | 2026-08-21 | Unspecified | 90.46 | Reinforcement Learning (RL) is an important paradigm for aligning Diffusion Language Models (DLMs) toward functional correctness in code generation. However, these models often encounter a ``capability cliff'' on complex tasks, where execution-based semantic rewards become too low to provide a viable learning signal. I... | [
-0.10730800032615662,
0.011851999908685684,
-0.0352029986679554,
0.03954799845814705,
0.029177000746130943,
-0.0460750013589859,
-0.024924999102950096,
-0.01896899938583374,
0.01260400004684925,
-0.038307998329401016,
-0.01593399979174137,
-0.056648001074790955,
0.021477999165654182,
-0.02... | [
-0.08152499794960022,
-0.060809001326560974,
-0.0426970012485981,
0.04164599999785423,
0.03219500184059143,
-0.04526299983263016,
-0.045333001762628555,
-0.002595999976620078,
0.0028440000023692846,
0.00570600014179945,
-0.013299000449478626,
-0.09370400011539459,
0.0005660000024363399,
-0... | AI Code Generation, Autonomous SWE Agents & Program Synthesis | Syntax-Guided AST & Semantic Code Search | [
"Python"
] | Specialized Autoregressive Code Transformer | [
"HumanEval (Functional Python Correctness)",
"MBPP / MBPP+ (Multi-Turn Python Benchmarking)",
"LiveCodeBench (Contamination-Free SOTA Evaluation)"
] | [
"Quantitative Program Synthesis & Pass@k Metric"
] | [
{
"benchmark_name": "HumanEval (Functional Python Correctness)",
"metric": "pass@1 / resolved",
"score": "Quantitative Program Synthesis & Pass@k Metric"
},
{
"benchmark_name": "MBPP / MBPP+ (Multi-Turn Python Benchmarking)",
"metric": "pass@1 / resolved",
"score": "Quantitative Program ... | [
{
"paper_id": "2509.01142v1",
"title": "Dream-Coder 7B: An Open Diffusion Language Model for Code",
"cosine_sim": 0.7528
},
{
"paper_id": "2606.20881v1",
"title": "When Do Intrinsic Rewards Work for Code Reasoning? A Comprehensive Study",
"cosine_sim": 0.7422
},
{
"paper_id": "25... | git clone https://github.com/Gen-Verse/dLLM-RL && cd dLLM-RL && (pip install -e . || pip install -r requirements.txt) | Reinforcement Learning (RL) is an important paradigm for aligning Diffusion Language Models (DLMs) toward functional correctness in code generation. | In this paper, we present a systematic empirical study of RL post-training for diffusion-based code generation along three axes: reward design, hint-conditioned sampling, and task difficulty. | Across HumanEval, MBPP, and LiveCodeBench, we find that static checking is the strongest overall standalone execution-free reward in our setting, especially improving DiffuCoder from 53.9 to 67.1 on HumanEval and from 14.9 to 15.5 on LiveCodeBench while reducing rollout time by 9.4\%. | 5 | Interactive Coding Assistants & Human-AI Pair Programming | Explosive (>50/mo) | 232 | 2026-08-21T14:36:53.660351 |
2604.05963v1 | QiMeng-PRepair: Precise Code Repair via Edit-Aware Reward Optimization | 2026-04-07T14:56:38Z | [
"cs.SE",
"cs.LG"
] | ArXiv Standard | http://arxiv.org/licenses/nonexclusive-distrib/1.0/ | true | 60 | Internal R&D Only (Copyleft or Academic Terms) | ArXiv Standard Distribution; Repository License Unspecified | 0 | 0 | 1 | Proposes We systematically quantify its impact and introduce precise repair task, which maximizes r... to enhance autonomous code synthesis, achieving that PRepair improves repair precision by up to 31.4% under $\mathrm{fix}_1@1$,.... | Changxin Ke | 13 | [
"Changxin Ke",
"Rui Zhang",
"Jiaming Guo",
"Yuanbo Wen",
"Li Ding",
"Shuo Wang",
"Xuyuan Zhu",
"Xiong Peng",
"Di Huang",
"Zidong Du",
"Xing Hu",
"Qi Guo",
"Yunji Chen"
] | [
"Autonomous Software Engineering AI Lab"
] | http://arxiv.org/abs/2604.05963v1 | VERIFIED_LIVE | https://github.com/apoorvumang/prompt-lookup-decoding | [
"https://github.com/apoorvumang/prompt-lookup-decoding",
"https://github.com/kcxain/QiMeng-PRepair"
] | 613 | 0 | 2026-08-21 | Unspecified | 90.32 | Large Language Models (LLMs) achieve strong program repair performance but often suffer from over-editing, where excessive modifications overwrite correct code and hinder bug localization. We systematically quantify its impact and introduce precise repair task, which maximizes reuse of correct code while fixing only bu... | [
-0.1836549937725067,
0.055727001279592514,
0.010840999893844128,
0.07256899774074554,
-0.07308699935674667,
0.037783000618219376,
-0.036757998168468475,
-0.036584001034498215,
-0.08259300142526627,
0.08610700070858002,
0.027726000174880028,
-0.03203599900007248,
0.04856500029563904,
-0.081... | [
-0.10995200276374817,
-0.008403999730944633,
0.04709099978208542,
0.033431001007556915,
-0.01758899912238121,
0.005198999773710966,
-0.020741000771522522,
0.0034920000471174717,
-0.052069999277591705,
0.08472999930381775,
-0.028024999424815178,
-0.03242800012230873,
0.03605100139975548,
-0... | AI Code Generation, Autonomous SWE Agents & Program Synthesis | Test-Driven Code Repair & Self-Healing | [
"Multi-Language / Polyglot"
] | Specialized Autoregressive Code Transformer | [
"HumanEval & SWE-bench Functional Synthesis Benchmark"
] | [
"Quantitative Program Synthesis & Pass@k Metric"
] | [
{
"benchmark_name": "HumanEval & SWE-bench Functional Synthesis Benchmark",
"metric": "pass@1 / resolved",
"score": "Quantitative Program Synthesis & Pass@k Metric"
}
] | [
{
"paper_id": "2604.03113v2",
"title": "PAFT: Preservation Aware Fine-Tuning for Minimal-Edit Program Repair",
"cosine_sim": 0.7675
},
{
"paper_id": "2607.03523v1",
"title": "Anchored Self-Play for Code Repair",
"cosine_sim": 0.7392
},
{
"paper_id": "2507.19909v1",
"title": "... | git clone https://github.com/apoorvumang/prompt-lookup-decoding && cd prompt-lookup-decoding && (pip install -e . || pip install -r requirements.txt) | Large Language Models (LLMs) achieve strong program repair performance but often suffer from over-editing, where excessive modifications overwrite correct code and hinder bug localization. | We systematically quantify its impact and introduce precise repair task, which maximizes reuse of correct code while fixing only buggy parts. | Experiments show that PRepair improves repair precision by up to 31.4% under $\mathrm{fix}_1@1$, a metric that jointly considers repair correctness and extent, and significantly increases decoding throughput when combined with speculative editing, demonstrating its potential for precise and practical code repair. | 7 | Syntax-Guided AST Modeling & Semantic Code Search | Explosive (>50/mo) | 232 | 2026-08-21T14:37:51.140706 |
2606.06473v1 | MLEvolve: A Self-Evolving Framework for Automated Machine Learning Algorithm Discovery | 2026-06-04T17:55:59Z | [
"cs.AI",
"cs.CL"
] | ArXiv Standard | http://arxiv.org/licenses/nonexclusive-distrib/1.0/ | true | 60 | Internal R&D Only (Copyleft or Academic Terms) | ArXiv Standard Distribution; Repository License Unspecified | 0 | 0 | 1 | Proposes However, existing MLE agents suffer from inter-branch information isolation, memoryless se... to enhance autonomous code synthesis, achieving Moreover, MLEvolve also outperforms specialized algorithm discovery methods incl.... | Shangheng Du | 14 | [
"Shangheng Du",
"Xiangchao Yan",
"Jinxin Shi",
"Zongsheng Cao",
"Shiyang Feng",
"Zichen Liang",
"Boyuan Sun",
"Tianshuo Peng",
"Yifan Zhou",
"Xin Li",
"Jie Zhou",
"Liang He",
"Bo Zhang",
"Lei Bai"
] | [
"Autonomous Software Engineering AI Lab"
] | http://arxiv.org/abs/2606.06473v1 | VERIFIED_LIVE | https://github.com/InternScience/MLEvolve | [
"https://github.com/InternScience/MLEvolve"
] | 422 | 0 | 2026-08-21 | Unspecified | 89.41 | Large language model (LLM) agents are increasingly applied to long-horizon tasks such as scientific discovery and machine learning engineering (MLE), where sustained self-evolution becomes a key capability. However, existing MLE agents suffer from inter-branch information isolation, memoryless search, and lack of hiera... | [
-0.03656800091266632,
-0.06795100122690201,
-0.034318000078201294,
-0.021758999675512314,
0.11433500051498413,
-0.0841670036315918,
-0.0248429998755455,
-0.05680999904870987,
-0.03854599967598915,
-0.03293300047516823,
-0.020885000005364418,
-0.07879099994897842,
-0.019432999193668365,
-0.... | [
-0.0059460001066327095,
-0.06751099973917007,
-0.015428000129759312,
0.026913000270724297,
0.06599999964237213,
-0.0416099987924099,
-0.0960799977183342,
-0.0042329998686909676,
0.02060599997639656,
-0.02740499936044216,
-0.051093000918626785,
-0.0587490014731884,
0.005013999994844198,
0.0... | AI Code Generation, Autonomous SWE Agents & Program Synthesis | Neural Program Synthesis & Generation Architecture | [
"Multi-Language / Polyglot"
] | Specialized Autoregressive Code Transformer | [
"HumanEval & SWE-bench Functional Synthesis Benchmark"
] | [
"Quantitative Program Synthesis & Pass@k Metric"
] | [
{
"benchmark_name": "HumanEval & SWE-bench Functional Synthesis Benchmark",
"metric": "pass@1 / resolved",
"score": "Quantitative Program Synthesis & Pass@k Metric"
}
] | [
{
"paper_id": "2508.02085v6",
"title": "SE-Agent: Self-Evolution Trajectory Optimization in Multi-Step Reasoning with LL",
"cosine_sim": 0.7392
},
{
"paper_id": "2602.07848v1",
"title": "MARTI-MARS$^2$: Scaling Multi-Agent Self-Search via Reinforcement Learning for C",
"cosine_sim": 0.72... | git clone https://github.com/InternScience/MLEvolve && cd MLEvolve && (pip install -e . || pip install -r requirements.txt) | Large language model (LLM) agents are increasingly applied to long-horizon tasks such as scientific discovery and machine learning engineering (MLE), where sustained self-evolution becomes a key capability. | However, existing MLE agents suffer from inter-branch information isolation, memoryless search, and lack of hierarchical control, which together hinder long-horizon optimization. | Moreover, MLEvolve also outperforms specialized algorithm discovery methods including AlphaEvolve on mathematical algorithm optimization tasks, demonstrating strong cross-domain generalization. | 1 | Autonomous Repo-Level Software Engineering Agents (SWE-bench) | Explosive (>50/mo) | 232 | 2026-08-21T14:36:27.574827 |
2509.23045v3 | Kimi-Dev: Agentless Training as Skill Prior for SWE-Agents | 2025-09-27T01:49:13Z | [
"cs.AI",
"cs.CL",
"cs.SE"
] | ArXiv Standard | http://arxiv.org/licenses/nonexclusive-distrib/1.0/ | true | 60 | Internal R&D Only (Copyleft or Academic Terms) | ArXiv Standard Distribution; Repository License Unspecified | 0 | 0 | 1 | Proposes Solutions are split into SWE-Agent frameworks with multi-turn interactions and workflow-ba... to enhance autonomous code synthesis, achieving These results show that structured skill priors from Agentless training can brid.... | Zonghan Yang | 21 | [
"Zonghan Yang",
"Shengjie Wang",
"Kelin Fu",
"Wenyang He",
"Weimin Xiong",
"Yibo Liu",
"Yibo Miao",
"Bofei Gao",
"Yejie Wang",
"Yingwei Ma",
"Yanhao Li",
"Yue Liu",
"Zhenxing Hu",
"Kaitai Zhang",
"Shuyi Wang",
"Huarong Chen",
"Flood Sung",
"Yang Liu",
"Yang Gao",
"Zhilin Yang",... | [
"Autonomous Software Engineering AI Lab"
] | http://arxiv.org/abs/2509.23045v3 | VERIFIED_LIVE | https://github.com/MoonshotAI/Kimi-Dev | [
"https://github.com/MoonshotAI/Kimi-Dev"
] | 1,300 | 0 | 2026-08-21 | Unspecified | 89.17 | Large Language Models (LLMs) are increasingly applied to software engineering (SWE), with SWE-bench as a key benchmark. Solutions are split into SWE-Agent frameworks with multi-turn interactions and workflow-based Agentless methods with single-turn verifiable steps. We argue these paradigms are not mutually exclusive: ... | [
-0.011258999817073345,
-0.03382100164890289,
-0.026002999395132065,
0.02109299972653389,
0.0017000000225380063,
0.0006709999870508909,
0.010982000268995762,
-0.051569998264312744,
-0.040995001792907715,
-0.04823800176382065,
-0.002107999986037612,
-0.02081100083887577,
0.07635799795389175,
... | [
-0.030587999150156975,
-0.035666000097990036,
-0.012640999630093575,
0.028497999534010887,
0.05490000173449516,
-0.053918998688459396,
-0.020228000357747078,
0.0035069999285042286,
0.003725999966263771,
-0.058612000197172165,
-0.08223699778318405,
-0.07200899720191956,
0.052048999816179276,
... | AI Code Generation, Autonomous SWE Agents & Program Synthesis | Autonomous Repo-Level SWE Agent (SWE-bench) | [
"Multi-Language / Polyglot"
] | Claude-3.5-Sonnet Agentic SWE Backbone | [
"SWE-bench (Autonomous Software Engineering)"
] | [
"achieving 60.4"
] | [
{
"benchmark_name": "SWE-bench (Autonomous Software Engineering)",
"metric": "pass@1 / resolved",
"score": "achieving 60.4"
}
] | [
{
"paper_id": "2506.07636v2",
"title": "SWE-Dev: Building Software Engineering Agents with Training and Inference Scalin",
"cosine_sim": 0.7629
},
{
"paper_id": "2511.13646v3",
"title": "Live-SWE-agent: Can Software Engineering Agents Self-Evolve on the Fly?",
"cosine_sim": 0.7472
},
... | git clone https://github.com/MoonshotAI/Kimi-Dev && cd Kimi-Dev && (pip install -e . || pip install -r requirements.txt) | Large Language Models (LLMs) are increasingly applied to software engineering (SWE), with SWE-bench as a key benchmark. | Solutions are split into SWE-Agent frameworks with multi-turn interactions and workflow-based Agentless methods with single-turn verifiable steps. | These results show that structured skill priors from Agentless training can bridge workflow and agentic frameworks for transferable coding agents. | 8 | Code Foundation Models & Specialized Instruction Distillation | Explosive (>50/mo) | 232 | 2026-08-21T14:40:31.276656 |
2508.14313v3 | Your Reward Function for RL is Your Best PRM for Search: Unifying RL and Search-Based TTS | 2025-08-19T23:41:15Z | [
"cs.LG",
"cs.AI"
] | ArXiv Standard | http://arxiv.org/licenses/nonexclusive-distrib/1.0/ | true | 60 | Internal R&D Only (Copyleft or Academic Terms) | ArXiv Standard Distribution; Repository License Unspecified | 0 | 0 | 1 | Proposes we introduce AIRL-S, the first natural unification of RL-based and search-based TTS. to enhance autonomous code synthesis, achieving These results underscore that, indeed, your reward function for RL is your best.... | Can Jin | 11 | [
"Can Jin",
"Yang Zhou",
"Qixin Zhang",
"Hongwu Peng",
"Di Zhang",
"Zihan Dong",
"Marco Pavone",
"Ligong Han",
"Zhang-Wei Hong",
"Tong Che",
"Dimitris N. Metaxas"
] | [
"Autonomous Software Engineering AI Lab"
] | http://arxiv.org/abs/2508.14313v3 | VERIFIED_LIVE | https://github.com/RLHFlow/RLHF-Reward-Modeling | [
"https://github.com/RLHFlow/RLHF-Reward-Modeling"
] | 1,500 | 0 | 2026-08-21 | Unspecified | 88.85 | Test-time scaling (TTS) for large language models (LLMs) has thus far fallen into two largely separate paradigms: (1) reinforcement learning (RL) methods that optimize sparse outcome-based rewards, yet suffer from instability and low sample efficiency; and (2) search-based techniques guided by independently trained, st... | [
-0.1087770015001297,
-0.0073090000078082085,
-0.04794599860906601,
0.003003000048920512,
-0.00944599974900484,
0.0754999965429306,
0.02645999938249588,
0.026281999424099922,
0.021957000717520714,
-0.005905999802052975,
0.007555000018328428,
-0.02130899950861931,
0.05584700033068657,
0.0488... | [
-0.11185000091791153,
-0.08028899878263474,
-0.016495000571012497,
0.07357499748468399,
0.02675200067460537,
0.039517998695373535,
0.0025420000310987234,
0.017480000853538513,
0.0561939999461174,
-0.027024000883102417,
0.018036000430583954,
-0.016999000683426857,
0.010162999853491783,
0.06... | AI Code Generation, Autonomous SWE Agents & Program Synthesis | Neural Program Synthesis & Generation Architecture | [
"Python"
] | GPT-4o / Codex Specialized Coding LLM | [
"HumanEval & SWE-bench Functional Synthesis Benchmark"
] | [
"improves performance by 9"
] | [
{
"benchmark_name": "HumanEval & SWE-bench Functional Synthesis Benchmark",
"metric": "pass@1 / resolved",
"score": "improves performance by 9"
}
] | [
{
"paper_id": "2508.05629v3",
"title": "On the Generalization of SFT: A Reinforcement Learning Perspective with Reward R",
"cosine_sim": 0.7664
},
{
"paper_id": "2603.00724v1",
"title": "RLAR: An Agentic Reward System for Multi-task Reinforcement Learning on Large La",
"cosine_sim": 0.75... | git clone https://github.com/RLHFlow/RLHF-Reward-Modeling && cd RLHF-Reward-Modeling && (pip install -e . || pip install -r requirements.txt) | Test-time scaling (TTS) for large language models (LLMs) has thus far fallen into two largely separate paradigms: (1) reinforcement learning (RL) methods that optimize sparse outcome-based rewards, yet suffer from instability and low sample efficiency; and (2) search-based techniques guided by independently trained, st... | In this paper, we introduce AIRL-S, the first natural unification of RL-based and search-based TTS. | These results underscore that, indeed, your reward function for RL is your best PRM for search, providing a robust and cost-effective solution to complex reasoning tasks in LLMs. | 5 | Interactive Coding Assistants & Human-AI Pair Programming | Explosive (>50/mo) | 232 | 2026-08-21T14:40:54.850090 |
2507.12507v1 | Scaling Up RL: Unlocking Diverse Reasoning in LLMs via Prolonged Training | 2025-07-16T17:59:24Z | [
"cs.LG",
"cs.AI",
"cs.CL"
] | ArXiv Standard | http://arxiv.org/licenses/nonexclusive-distrib/1.0/ | true | 60 | Internal R&D Only (Copyleft or Academic Terms) | ArXiv Standard Distribution; Repository License Unspecified | 0 | 0 | 1 | Proposes controlled KL regularization, clipping ratio, and periodic reference policy resets as crit... to enhance autonomous code synthesis, achieving Our model achieves significant improvements over strong baselines, including +14.... | Mingjie Liu | 22 | [
"Mingjie Liu",
"Shizhe Diao",
"Jian Hu",
"Ximing Lu",
"Xin Dong",
"Hao Zhang",
"Alexander Bukharin",
"Shaokun Zhang",
"Jiaqi Zeng",
"Makesh Narsimhan Sreedhar",
"Gerald Shen",
"David Mosallanezhad",
"Di Zhang",
"Jonas Yang",
"June Yang",
"Oleksii Kuchaiev",
"Guilin Liu",
"Zhiding Y... | [
"Autonomous Software Engineering AI Lab"
] | http://arxiv.org/abs/2507.12507v1 | VERIFIED_LIVE | https://github.com/open-thought/reasoning-gym | [
"https://github.com/open-thought/reasoning-gym"
] | 1,500 | 0 | 2026-08-21 | Unspecified | 87.49 | Recent advancements in reasoning-focused language models such as OpenAI's O1 and DeepSeek-R1 have shown that scaling test-time computation-through chain-of-thought reasoning and iterative exploration-can yield substantial improvements on complex tasks like mathematics and code generation. These breakthroughs have been ... | [
0.02719300054013729,
-0.07031799852848053,
0.023530999198555946,
0.0038689998909831047,
0.08003800362348557,
-0.017659999430179596,
-0.008681000210344791,
-0.013868999667465687,
0.016944000497460365,
0.025404000654816628,
-0.061448998749256134,
-0.010374000295996666,
0.01471799984574318,
0... | [
-0.058024998754262924,
-0.0580269992351532,
0.04976699873805046,
0.03020000085234642,
0.049476999789476395,
0.0258760005235672,
-0.008461999706923962,
-0.012281999923288822,
0.03342000022530556,
0.0002610000083222985,
-0.05807200074195862,
-0.03480200096964836,
-0.000004999999873689376,
0.... | AI Code Generation, Autonomous SWE Agents & Program Synthesis | Neural Program Synthesis & Generation Architecture | [
"Python"
] | Specialized Autoregressive Code Transformer | [
"HumanEval & SWE-bench Functional Synthesis Benchmark"
] | [
"Quantitative Program Synthesis & Pass@k Metric"
] | [
{
"benchmark_name": "HumanEval & SWE-bench Functional Synthesis Benchmark",
"metric": "pass@1 / resolved",
"score": "Quantitative Program Synthesis & Pass@k Metric"
}
] | [
{
"paper_id": "2509.19893v3",
"title": "Future Policy Approximation for Offline Reinforcement Learning in LLM Reasoning",
"cosine_sim": 0.7629
},
{
"paper_id": "2508.05629v3",
"title": "On the Generalization of SFT: A Reinforcement Learning Perspective with Reward R",
"cosine_sim": 0.761... | git clone https://github.com/open-thought/reasoning-gym && cd reasoning-gym && (pip install -e . || pip install -r requirements.txt) | Recent advancements in reasoning-focused language models such as OpenAI's O1 and DeepSeek-R1 have shown that scaling test-time computation-through chain-of-thought reasoning and iterative exploration-can yield substantial improvements on complex tasks like mathematics and code generation. | We introduce controlled KL regularization, clipping ratio, and periodic reference policy resets as critical components for unlocking long-term performance gains. | Our model achieves significant improvements over strong baselines, including +14.7% on math, +13.9% on coding, and +54.8% on logic puzzle tasks. | 5 | Interactive Coding Assistants & Human-AI Pair Programming | Explosive (>50/mo) | 232 | 2026-08-21T14:41:20.081323 |
2602.02084v2 | Closing the Loop: Universal Repository Representation with RPG-Encoder | 2026-02-02T13:30:00Z | [
"cs.CL",
"cs.SE"
] | ArXiv Standard | http://arxiv.org/licenses/nonexclusive-distrib/1.0/ | true | 60 | Internal R&D Only (Copyleft or Academic Terms) | ArXiv Standard Distribution; Repository License Unspecified | 0 | 0 | 1 | Proposes To address this, we propose RPG-Encoder, a framework that generalizes the Repository Plann... to enhance autonomous code synthesis, achieving Furthermore, it achieves 98.5% reconstruction coverage on RepoCraft, confirming.... | Jane Luo | 13 | [
"Jane Luo",
"Chengyu Yin",
"Xin Zhang",
"Qingtao Li",
"Steven Liu",
"Yiming Huang",
"Jie Wu",
"Hao Liu",
"Yangyu Huang",
"Yu Kang",
"Fangkai Yang",
"Ying Xin",
"Scarlett Li"
] | [
"Autonomous Software Engineering AI Lab"
] | http://arxiv.org/abs/2602.02084v2 | VERIFIED_LIVE | https://github.com/microsoft/RPG-ZeroRepo | [
"https://github.com/microsoft/RPG-ZeroRepo"
] | 590 | 0 | 2026-08-21 | Unspecified | 87.43 | Current repository agents encounter a reasoning disconnect due to fragmented representations, as existing methods rely on isolated API documentation or dependency graphs that lack semantic depth. We consider repository comprehension and generation to be inverse processes within a unified cycle: generation expands inten... | [
-0.05747700110077858,
-0.04839299991726875,
-0.0682239979505539,
-0.07406800240278244,
0.015634000301361084,
0.007428999990224838,
-0.03198099881410599,
-0.07943800091743469,
0.029323000460863113,
0.001133000012487173,
-0.025582000613212585,
-0.018516000360250473,
-0.005508999805897474,
0.... | [
-0.05359100177884102,
-0.06907899677753448,
-0.01322100032120943,
-0.02072799950838089,
0.028852999210357666,
-0.05138000100851059,
-0.0541750006377697,
-0.03884400054812431,
0.0025599999353289604,
0.036400001496076584,
-0.07962299883365631,
-0.026415999978780746,
0.036295000463724136,
0.0... | AI Code Generation, Autonomous SWE Agents & Program Synthesis | Autonomous Repo-Level SWE Agent (SWE-bench) | [
"Python"
] | Claude-3.5-Sonnet Agentic SWE Backbone | [
"SWE-bench (Autonomous Software Engineering)"
] | [
"achieves 98.5%"
] | [
{
"benchmark_name": "SWE-bench (Autonomous Software Engineering)",
"metric": "pass@1 / resolved",
"score": "achieves 98.5%"
}
] | [
{
"paper_id": "2509.16198v6",
"title": "RPG: A Repository Planning Graph for Unified and Scalable Codebase Generation",
"cosine_sim": 0.7623
},
{
"paper_id": "2507.19942v2",
"title": "Prometheus: Towards Long-Horizon Codebase Navigation for Repository-Level Proble",
"cosine_sim": 0.6918
... | git clone https://github.com/microsoft/RPG-ZeroRepo && cd RPG-ZeroRepo && (pip install -e . || pip install -r requirements.txt) | Current repository agents encounter a reasoning disconnect due to fragmented representations, as existing methods rely on isolated API documentation or dependency graphs that lack semantic depth. | To address this, we propose RPG-Encoder, a framework that generalizes the Repository Planning Graph (RPG) from a static generative blueprint into a unified, high-fidelity representation. | Furthermore, it achieves 98.5% reconstruction coverage on RepoCraft, confirming RPG's high-fidelity capacity to mirror the original codebase and closing the loop between intent and implementation. | 3 | Multi-Language Code Translation & Cross-Platform Migration | Explosive (>50/mo) | 232 | 2026-08-21T14:38:53.736833 |
2602.07848v1 | MARTI-MARS$^2$: Scaling Multi-Agent Self-Search via Reinforcement Learning for Code Generation | 2026-02-08T07:28:44Z | [
"cs.LG"
] | ArXiv Standard | http://arxiv.org/licenses/nonexclusive-distrib/1.0/ | true | 60 | Internal R&D Only (Copyleft or Academic Terms) | ArXiv Standard Distribution; Repository License Unspecified | 0 | 0 | 1 | Proposes Multi-agent collaboration offers a promising avenue to transcend these boundaries. to enhance autonomous code synthesis, achieving Utilizing two collaborating 32B models, MARTI-MARS2 achieves 77.7%, outperformin.... | Shijie Wang | 24 | [
"Shijie Wang",
"Pengfei Li",
"Yikun Fu",
"Kaifeng Liu",
"Fangyuan Li",
"Yang Liu",
"Xiaowei Sun",
"Zonglin Li",
"Siyao Zhao",
"Jian Zhao",
"Kai Tian",
"Dong Li",
"Junqi Gao",
"Yutong Zhang",
"Yiqun Chen",
"Yuqiang Li",
"Zoe Li",
"Weinan Zhang",
"Peng Ye",
"Shuyue Hu",
"Lei Ba... | [
"Autonomous Software Engineering AI Lab"
] | http://arxiv.org/abs/2602.07848v1 | VERIFIED_LIVE | https://github.com/TsinghuaC3I/MARTI | [
"https://github.com/TsinghuaC3I/MARTI"
] | 549 | 0 | 2026-08-21 | Unspecified | 87.05 | While the complex reasoning capability of Large Language Models (LLMs) has attracted significant attention, single-agent systems often encounter inherent performance ceilings in complex tasks such as code generation. Multi-agent collaboration offers a promising avenue to transcend these boundaries. However, existing fr... | [
-0.05111600086092949,
-0.0018269999418407679,
-0.052577998489141464,
0.023770000785589218,
-0.03858200088143349,
-0.008403000421822071,
-0.053084999322891235,
-0.010188000276684761,
0.02278600074350834,
-0.011014999821782112,
0.028098000213503838,
-0.08771800249814987,
0.023864999413490295,
... | [
0.02566099911928177,
-0.050818998366594315,
-0.022693999111652374,
0.0559380017220974,
0.0508820004761219,
0.019057000055909157,
-0.05356400087475777,
-0.03603599965572357,
0.018736999481916428,
-0.028394000604748726,
-0.03099600039422512,
-0.08545500040054321,
0.048865001648664474,
0.0249... | AI Code Generation, Autonomous SWE Agents & Program Synthesis | Neural Program Synthesis & Generation Architecture | [
"Multi-Language / Polyglot"
] | Specialized Autoregressive Code Transformer | [
"HumanEval & SWE-bench Functional Synthesis Benchmark"
] | [
"achieves 77.7%"
] | [
{
"benchmark_name": "HumanEval & SWE-bench Functional Synthesis Benchmark",
"metric": "pass@1 / resolved",
"score": "achieves 77.7%"
}
] | [
{
"paper_id": "2604.14564v1",
"title": "MARS$^2$: Scaling Multi-Agent Tree Search via Reinforcement Learning for Code Ge",
"cosine_sim": 0.8529
},
{
"paper_id": "2508.02085v6",
"title": "SE-Agent: Self-Evolution Trajectory Optimization in Multi-Step Reasoning with LL",
"cosine_sim": 0.74... | git clone https://github.com/TsinghuaC3I/MARTI && cd MARTI && (pip install -e . || pip install -r requirements.txt) | While the complex reasoning capability of Large Language Models (LLMs) has attracted significant attention, single-agent systems often encounter inherent performance ceilings in complex tasks such as code generation. | Multi-agent collaboration offers a promising avenue to transcend these boundaries. | Utilizing two collaborating 32B models, MARTI-MARS2 achieves 77.7%, outperforming strong baselines like GPT-5.1. | 5 | Interactive Coding Assistants & Human-AI Pair Programming | Explosive (>50/mo) | 232 | 2026-08-21T14:38:46.397044 |
2510.02387v1 | CWM: An Open-Weights LLM for Research on Code Generation with World Models | 2025-09-30T21:47:10Z | [
"cs.SE",
"cs.AI",
"cs.LG"
] | ArXiv Standard | http://arxiv.org/licenses/nonexclusive-distrib/1.0/ | true | 60 | Internal R&D Only (Copyleft or Academic Terms) | ArXiv Standard Distribution; Repository License Unspecified | 0 | 0 | 1 | Proposes To improve code understanding beyond what can be learned from training on static code alon... to enhance autonomous code synthesis, achieving Independent of its world modeling capabilities, CWM offers strong performance on.... | FAIR CodeGen team | 51 | [
"FAIR CodeGen team",
"Jade Copet",
"Quentin Carbonneaux",
"Gal Cohen",
"Jonas Gehring",
"Jacob Kahn",
"Jannik Kossen",
"Felix Kreuk",
"Emily McMilin",
"Michel Meyer",
"Yuxiang Wei",
"David Zhang",
"Kunhao Zheng",
"Jordi Armengol-Estapé",
"Pedram Bashiri",
"Maximilian Beck",
"Pierre C... | [
"Autonomous Software Engineering AI Lab"
] | http://arxiv.org/abs/2510.02387v1 | VERIFIED_LIVE | https://github.com/facebookresearch/cwm | [
"https://github.com/facebookresearch/cwm"
] | 890 | 0 | 2026-08-21 | Unspecified | 86 | We release Code World Model (CWM), a 32-billion-parameter open-weights LLM, to advance research on code generation with world models. To improve code understanding beyond what can be learned from training on static code alone, we mid-train CWM on a large amount of observation-action trajectories from Python interpreter... | [
-0.08650500327348709,
-0.0558990016579628,
-0.013085000216960907,
0.03378099948167801,
0.017767999321222305,
-0.06802400201559067,
-0.07475999742746353,
-0.012466000393033028,
-0.033656999468803406,
-0.05203099921345711,
-0.006202000193297863,
-0.08432099968194962,
0.028821000829339027,
-0... | [
-0.07398000359535217,
-0.0768200010061264,
-0.01776999980211258,
0.03642600029706955,
0.04105500131845474,
-0.1119220033288002,
-0.05347200110554695,
-0.018213000148534775,
-0.020635999739170074,
-0.0014939999673515558,
-0.07628600299358368,
-0.09836900234222412,
0.0644960030913353,
0.0450... | AI Code Generation, Autonomous SWE Agents & Program Synthesis | Autonomous Repo-Level SWE Agent (SWE-bench) | [
"Python"
] | Claude-3.5-Sonnet Agentic SWE Backbone | [
"SWE-bench (Autonomous Software Engineering)",
"LiveCodeBench (Contamination-Free SOTA Evaluation)"
] | [
"65.8% on SWE-bench"
] | [
{
"benchmark_name": "SWE-bench (Autonomous Software Engineering)",
"metric": "pass@1 / resolved",
"score": "65.8% on SWE-bench"
},
{
"benchmark_name": "LiveCodeBench (Contamination-Free SOTA Evaluation)",
"metric": "pass@1 / resolved",
"score": "65.8% on SWE-bench"
}
] | [
{
"paper_id": "2602.03419v1",
"title": "SWE-World: Building Software Engineering Agents in Docker-Free Environments",
"cosine_sim": 0.742
},
{
"paper_id": "2511.09794v1",
"title": "Evaluating Software Process Models for Multi-Agent Class-Level Code Generation",
"cosine_sim": 0.7329
},
... | git clone https://github.com/facebookresearch/cwm && cd cwm && (pip install -e . || pip install -r requirements.txt) | We release Code World Model (CWM), a 32-billion-parameter open-weights LLM, to advance research on code generation with world models. | To improve code understanding beyond what can be learned from training on static code alone, we mid-train CWM on a large amount of observation-action trajectories from Python interpreter and agentic Docker environments, and perform extensive multi-task reasoning RL in verifiable coding, math, and multi-turn software en... | Independent of its world modeling capabilities, CWM offers strong performance on general coding and math tasks: it reaches pass@1 scores of 65.8% on SWE-bench Verified (with test-time scaling), 68.6% on LiveCodeBench, 96.6% on Math-500, and 76.0% on AIME 2024. | 8 | Code Foundation Models & Specialized Instruction Distillation | Explosive (>50/mo) | 232 | 2026-08-21T14:40:25.503983 |
2601.13727v1 | Foundational VeriFast: Pragmatic Certification of Verification Tool Results through Hinted Mirroring | 2026-01-20T08:34:38Z | [
"cs.PL"
] | ArXiv Standard | http://arxiv.org/licenses/nonexclusive-distrib/1.0/ | true | 60 | Internal R&D Only (Copyleft or Academic Terms) | ArXiv Standard Distribution; Repository License Unspecified | 0 | 0 | 1 | Proposes It verifies a program by symbolically executing each function in isolation, exploiting use... to enhance autonomous code synthesis, achieving We here report on an early result extending VeriFast to emit, upon successful ve.... | Bart Jacobs | 1 | [
"Bart Jacobs"
] | [
"Autonomous Software Engineering AI Lab"
] | http://arxiv.org/abs/2601.13727v1 | VERIFIED_LIVE | https://github.com/verifast/verifast | [
"https://github.com/verifast/verifast"
] | 501 | 0 | 2026-08-21 | Unspecified | 85.49 | VeriFast is a leading tool for the modular formal verification of correctness properties of single-threaded and multi-threaded C and Rust programs. It verifies a program by symbolically executing each function in isolation, exploiting user-annotated preconditions, postconditions, and loop invariants written in a form o... | [
-0.10424599796533585,
-0.01672999933362007,
-0.0032450000289827585,
-0.047297000885009766,
0.020812999457120895,
-0.06682299822568893,
0.017433999106287956,
-0.029597999528050423,
-0.002203000010922551,
0.008845999836921692,
0.014840999618172646,
-0.05651399865746498,
0.0512860007584095,
0... | [
-0.09839899837970734,
-0.015170999802649021,
-0.06564900279045105,
0.017645999789237976,
0.07184299826622009,
-0.09464400261640549,
0.009777000173926353,
0.011694000102579594,
-0.040137000381946564,
-0.03389900177717209,
0.037484001368284225,
-0.03785400092601776,
0.025505999103188515,
-0.... | AI Code Generation, Autonomous SWE Agents & Program Synthesis | Formal Logic & Theorem Proving Synthesis | [
"Rust"
] | Specialized Autoregressive Code Transformer | [
"HumanEval & SWE-bench Functional Synthesis Benchmark"
] | [
"Quantitative Program Synthesis & Pass@k Metric"
] | [
{
"benchmark_name": "HumanEval & SWE-bench Functional Synthesis Benchmark",
"metric": "pass@1 / resolved",
"score": "Quantitative Program Synthesis & Pass@k Metric"
}
] | [
{
"paper_id": "2607.19795v1",
"title": "Towards Automated Formal Verification of zkEVMs Using LLM-Guided Constraint Synt",
"cosine_sim": 0.7078
},
{
"paper_id": "2605.08553v1",
"title": "VeriContest: A Competitive-Programming Benchmark for Verifiable Code Generation",
"cosine_sim": 0.704... | git clone https://github.com/verifast/verifast && cd verifast && (pip install -e . || pip install -r requirements.txt) | VeriFast is a leading tool for the modular formal verification of correctness properties of single-threaded and multi-threaded C and Rust programs. | It verifies a program by symbolically executing each function in isolation, exploiting user-annotated preconditions, postconditions, and loop invariants written in a form of separation logic, and using a separation logic-based symbolic representation of memory. | We here report on an early result extending VeriFast to emit, upon successful verification of a Rust program, a Rocq proof script that proves correctness of the program with respect to a Rocq-encoded axiomatic semantics of Rust. | 6 | Static Analysis, Vulnerability Detection & Automated Security Patching | Explosive (>50/mo) | 232 | 2026-08-21T14:39:08.823322 |
2507.19942v2 | Prometheus: Towards Long-Horizon Codebase Navigation for Repository-Level Problem Solving | 2025-07-26T13:13:22Z | [
"cs.SE"
] | ArXiv Standard | http://arxiv.org/licenses/nonexclusive-distrib/1.0/ | true | 60 | Internal R&D Only (Copyleft or Academic Terms) | ArXiv Standard Distribution; Repository License Unspecified | 0 | 0 | 1 | Proposes However, existing agents still struggle to navigate large-scale codebases, as the Needle-i... to enhance autonomous code synthesis, achieving Powered by GPT-5, Prometheus achieves state-of-the-art performance with 74.4% an.... | Yue Pan | 11 | [
"Yue Pan",
"Zimin Chen",
"Siyu Lu",
"Zhaoyang Chu",
"Xiang Li",
"Han Li",
"Yang Feng",
"Claire Le Goues",
"Federica Sarro",
"Martin Monperrus",
"He Ye"
] | [
"Autonomous Software Engineering AI Lab"
] | http://arxiv.org/abs/2507.19942v2 | VERIFIED_LIVE | https://github.com/EuniAI/Prometheus | [
"https://github.com/EuniAI/Prometheus"
] | 1,100 | 0 | 2026-08-21 | Unspecified | 85.2 | Large Language Models (LLMs) have shown remarkable capabilities in automating software engineering tasks, spurring the emergence of coding agents that scaffold LLMs with external tools to resolve repository-level problems. However, existing agents still struggle to navigate large-scale codebases, as the Needle-in-a-Hay... | [
-0.09403800219297409,
-0.048847999423742294,
0.0028609998989850283,
0.003696999978274107,
0.019785000011324883,
-0.11771500110626221,
-0.03245199844241142,
0.014252999797463417,
-0.09890200197696686,
0.06932099908590317,
-0.0218299999833107,
-0.05899199843406677,
0.03787900134921074,
0.004... | [
-0.053676001727581024,
-0.08275599777698517,
-0.05331899970769882,
0.02886199951171875,
0.059126999229192734,
-0.05317600071430206,
-0.037689000368118286,
0.0427200011909008,
-0.028126999735832214,
0.022332999855279922,
-0.029054999351501465,
-0.06444700062274933,
0.08159200102090836,
0.03... | AI Code Generation, Autonomous SWE Agents & Program Synthesis | Autonomous Repo-Level SWE Agent (SWE-bench) | [
"Multi-Language / Polyglot"
] | GPT-4o / Codex Specialized Coding LLM | [
"SWE-bench (Autonomous Software Engineering)"
] | [
"Quantitative Program Synthesis & Pass@k Metric"
] | [
{
"benchmark_name": "SWE-bench (Autonomous Software Engineering)",
"metric": "pass@1 / resolved",
"score": "Quantitative Program Synthesis & Pass@k Metric"
}
] | [
{
"paper_id": "2606.14066v4",
"title": "FastContext: Training Efficient Repository Explorer for Coding Agents",
"cosine_sim": 0.8019
},
{
"paper_id": "2605.14563v2",
"title": "Remember Your Trace: Memory-Guided Long-Horizon Agentic Framework for Consistent",
"cosine_sim": 0.7891
},
{... | git clone https://github.com/EuniAI/Prometheus && cd Prometheus && (pip install -e . || pip install -r requirements.txt) | Large Language Models (LLMs) have shown remarkable capabilities in automating software engineering tasks, spurring the emergence of coding agents that scaffold LLMs with external tools to resolve repository-level problems. | However, existing agents still struggle to navigate large-scale codebases, as the Needle-in-a-Haystack problem persists even with million-token context windows, where relevant evidence is often overwhelmed by large volumes of irrelevant code and documentation. | Powered by GPT-5, Prometheus achieves state-of-the-art performance with 74.4% and 33.8% resolution rates on the two benchmarks, ranking Top-6 and Top-1 among open-source agent systems, respectively. | 8 | Code Foundation Models & Specialized Instruction Distillation | Explosive (>50/mo) | 232 | 2026-08-21T14:41:11.549732 |
2508.06942v1 | When Prompt Engineering Meets Software Engineering: CNL-P as Natural and Robust "APIs'' for Human-AI Interaction | 2025-08-09T11:32:33Z | [
"cs.SE",
"cs.AI"
] | ArXiv Standard | http://arxiv.org/licenses/nonexclusive-distrib/1.0/ | true | 60 | Internal R&D Only (Copyleft or Academic Terms) | ArXiv Standard Distribution; Repository License Unspecified | 0 | 0 | 1 | Proposes To improve prompt quality, best practices for prompt engineering (PE) have been developed,... to enhance autonomous code synthesis, achieving Extensive experiments demonstrate that CNL-P enhances the quality of LLM respons.... | Zhenchang Xing | 7 | [
"Zhenchang Xing",
"Yang Liu",
"Zhuo Cheng",
"Qing Huang",
"Dehai Zhao",
"Daniel Sun",
"Chenhua Liu"
] | [
"Autonomous Software Engineering AI Lab"
] | http://arxiv.org/abs/2508.06942v1 | VERIFIED_LIVE | https://github.com/allenai/natural-instructions | [
"https://github.com/allenai/natural-instructions"
] | 1,000 | 0 | 2026-08-21 | Unspecified | 84.93 | With the growing capabilities of large language models (LLMs), they are increasingly applied in areas like intelligent customer service, code generation, and knowledge management. Natural language (NL) prompts act as the ``APIs'' for human-LLM interaction. To improve prompt quality, best practices for prompt engineerin... | [
-0.12465699762105942,
0.022444000467658043,
-0.01550699956715107,
-0.009182999841868877,
0.04641399905085564,
-0.09000299870967865,
0.019020000472664833,
0.052988000214099884,
0.03728799894452095,
-0.01193500030785799,
-0.11192499846220016,
-0.07547400146722794,
0.04990300163626671,
-0.008... | [
-0.03772199898958206,
0.008232999593019485,
0.03517499938607216,
0.0035780000034719706,
0.01845799945294857,
-0.05468999966979027,
0.020757000893354416,
0.06488300114870071,
-0.02114900015294552,
0.0002789999998640269,
-0.07787399739027023,
-0.07870300114154816,
0.10568100214004517,
-0.035... | AI Code Generation, Autonomous SWE Agents & Program Synthesis | Static Analysis & Vulnerability Remediation | [
"Multi-Language / Polyglot"
] | Specialized Autoregressive Code Transformer | [
"HumanEval & SWE-bench Functional Synthesis Benchmark"
] | [
"Quantitative Program Synthesis & Pass@k Metric"
] | [
{
"benchmark_name": "HumanEval & SWE-bench Functional Synthesis Benchmark",
"metric": "pass@1 / resolved",
"score": "Quantitative Program Synthesis & Pass@k Metric"
}
] | [
{
"paper_id": "2601.13118v1",
"title": "Guidelines to Prompt Large Language Models for Code Generation: An Empirical Cha",
"cosine_sim": 0.7484
},
{
"paper_id": "2506.05614v1",
"title": "Which Prompting Technique Should I Use? An Empirical Investigation of Prompting ",
"cosine_sim": 0.74... | git clone https://github.com/allenai/natural-instructions && cd natural-instructions && (pip install -e . || pip install -r requirements.txt) | With the growing capabilities of large language models (LLMs), they are increasingly applied in areas like intelligent customer service, code generation, and knowledge management. | To improve prompt quality, best practices for prompt engineering (PE) have been developed, including writing guidelines and templates. | Extensive experiments demonstrate that CNL-P enhances the quality of LLM responses through the novel and organic synergy of PE and SE. | 2 | Test-Driven Program Repair & Self-Healing Code Synthesis | Explosive (>50/mo) | 232 | 2026-08-21T14:41:01.574762 |
2509.23405v3 | Planner Aware Path Learning in Diffusion Language Models Training | 2025-09-27T16:51:07Z | [
"cs.LG"
] | ArXiv Standard | http://arxiv.org/licenses/nonexclusive-distrib/1.0/ | true | 60 | Internal R&D Only (Copyleft or Academic Terms) | ArXiv Standard Distribution; Repository License Unspecified | 0 | 0 | 1 | Proposes This flexibility of sampling is unlocked by new engineered sampling strategies, or planner... to enhance autonomous code synthesis, achieving Empirically, we show PAPL delivers consistent gains across domains, including a.... | Fred Zhangzhi Peng | 8 | [
"Fred Zhangzhi Peng",
"Zachary Bezemek",
"Jarrid Rector-Brooks",
"Shuibai Zhang",
"Anru R. Zhang",
"Michael Bronstein",
"Alexander Tong",
"Avishek Joey Bose"
] | [
"Autonomous Software Engineering AI Lab"
] | http://arxiv.org/abs/2509.23405v3 | VERIFIED_LIVE | https://github.com/pengzhangzhi/Open-dLLM | [
"https://github.com/pengzhangzhi/Open-dLLM",
"https://github.com/pengzhangzhi/PAPL"
] | 646 | 0 | 2026-08-21 | Unspecified | 83.1 | Diffusion language models have emerged as a powerful alternative to autoregressive models, enabling fast inference through more flexible and parallel generation paths. This flexibility of sampling is unlocked by new engineered sampling strategies, or planners, that select more favorable generation paths by iteratively ... | [
0.05187699943780899,
-0.1260959953069687,
0.023706000298261642,
0.047481000423431396,
0.006453000009059906,
0.03500000014901161,
-0.0293550007045269,
-0.013089999556541443,
0.03220000118017197,
-0.05113799870014191,
-0.012594000436365604,
-0.017842000350356102,
-0.009173000231385231,
0.064... | [
-0.03032200038433075,
-0.0988439992070198,
0.023821000009775162,
0.04847799986600876,
0.015468999743461609,
-0.029451999813318253,
-0.06197800114750862,
-0.054597001522779465,
0.03349800035357475,
-0.014643999747931957,
-0.04859999939799309,
-0.005886000115424395,
-0.023454999551177025,
0.... | AI Code Generation, Autonomous SWE Agents & Program Synthesis | Neural Program Synthesis & Generation Architecture | [
"Multi-Language / Polyglot"
] | Specialized Autoregressive Code Transformer | [
"HumanEval (Functional Python Correctness)"
] | [
"pass@10"
] | [
{
"benchmark_name": "HumanEval (Functional Python Correctness)",
"metric": "pass@1 / resolved",
"score": "pass@10"
}
] | [
{
"paper_id": "2603.13243v1",
"title": "Think First, Diffuse Fast: Improving Diffusion Language Model Reasoning via Auto",
"cosine_sim": 0.7366
},
{
"paper_id": "2605.13935v1",
"title": "Beyond Mode-Seeking RL: Trajectory-Balance Post-Training for Diffusion Language ",
"cosine_sim": 0.73... | git clone https://github.com/pengzhangzhi/Open-dLLM && cd Open-dLLM && (pip install -e . || pip install -r requirements.txt) | Diffusion language models have emerged as a powerful alternative to autoregressive models, enabling fast inference through more flexible and parallel generation paths. | This flexibility of sampling is unlocked by new engineered sampling strategies, or planners, that select more favorable generation paths by iteratively planning - versus uniformly at random - where to denoise along the sequence. | Empirically, we show PAPL delivers consistent gains across domains, including a 40% relative improvement in protein sequences, improved text generation with up to a 4x relative MAUVE gain, and 23% relative improvement in code generation HumanEval pass@10. | 5 | Interactive Coding Assistants & Human-AI Pair Programming | Explosive (>50/mo) | 232 | 2026-08-21T14:40:34.730718 |
2510.23564v4 | ReCode: Unify Plan and Action for Universal Granularity Control | 2025-10-27T17:35:15Z | [
"cs.AI",
"cs.CL",
"cs.LG"
] | ArXiv Standard | http://arxiv.org/licenses/nonexclusive-distrib/1.0/ | true | 60 | Internal R&D Only (Copyleft or Academic Terms) | ArXiv Standard Distribution; Repository License Unspecified | 0 | 0 | 1 | Proposes However, current Large Language Model (LLM)-based agents lack this crucial capability to o... to enhance autonomous code synthesis, achieving Extensive experiments show ReCode significantly surpasses advanced baselines in.... | Zhaoyang Yu | 12 | [
"Zhaoyang Yu",
"Jiayi Zhang",
"Huixue Su",
"Yufan Zhao",
"Yifan Wu",
"Mingyi Deng",
"Jinyu Xiang",
"Yizhang Lin",
"Lingxiao Tang",
"Yuyu Luo",
"Bang Liu",
"Chenglin Wu"
] | [
"Autonomous Software Engineering AI Lab"
] | http://arxiv.org/abs/2510.23564v4 | VERIFIED_LIVE | https://github.com/FoundationAgents/ReCode | [
"https://github.com/FoundationAgents/ReCode"
] | 562 | 0 | 2026-08-21 | Unspecified | 83.09 | Real-world tasks require decisions at varying granularities, and humans excel at this by leveraging a unified cognitive representation where planning is fundamentally understood as a high-level form of action. However, current Large Language Model (LLM)-based agents lack this crucial capability to operate fluidly acros... | [
-0.07005199790000916,
-0.005915000103414059,
0.0635870024561882,
-0.03602999821305275,
-0.025193000212311745,
-0.04232700169086456,
0.04154499992728233,
0.037342000752687454,
0.002202000003308058,
0.031895000487565994,
-0.04324600100517273,
-0.00521400012075901,
0.01093900017440319,
-0.013... | [
-0.023716000840067863,
-0.04749299958348274,
0.033574000000953674,
-0.024925999343395233,
-0.039361000061035156,
-0.06025100126862526,
-0.025245999917387962,
0.03562700003385544,
0.01413199957460165,
0.014097999781370163,
-0.06771499663591385,
-0.06515499949455261,
0.020708000287413597,
0.... | AI Code Generation, Autonomous SWE Agents & Program Synthesis | Neural Program Synthesis & Generation Architecture | [
"Multi-Language / Polyglot"
] | Specialized Autoregressive Code Transformer | [
"HumanEval & SWE-bench Functional Synthesis Benchmark"
] | [
"Quantitative Program Synthesis & Pass@k Metric"
] | [
{
"benchmark_name": "HumanEval & SWE-bench Functional Synthesis Benchmark",
"metric": "pass@1 / resolved",
"score": "Quantitative Program Synthesis & Pass@k Metric"
}
] | [
{
"paper_id": "2602.00929v1",
"title": "Learning Abstractions for Hierarchical Planning in Program-Synthesis Agents",
"cosine_sim": 0.7431
},
{
"paper_id": "2605.24812v1",
"title": "CoRe-Code: Collaborative Reinforcement Learning for Code Generation",
"cosine_sim": 0.7136
},
{
"p... | git clone https://github.com/FoundationAgents/ReCode && cd ReCode && (pip install -e . || pip install -r requirements.txt) | Real-world tasks require decisions at varying granularities, and humans excel at this by leveraging a unified cognitive representation where planning is fundamentally understood as a high-level form of action. | However, current Large Language Model (LLM)-based agents lack this crucial capability to operate fluidly across decision granularities. | Extensive experiments show ReCode significantly surpasses advanced baselines in inference performance and demonstrates exceptional data efficiency in training, validating our core insight that unifying planning and action through recursive code generation is a powerful and effective approach to achieving universal gran... | 5 | Interactive Coding Assistants & Human-AI Pair Programming | Explosive (>50/mo) | 232 | 2026-08-21T14:40:02.011127 |
2603.22519v2 | LLMON: An LLM-native Markup Language to Leverage Structure and Semantics at the LLM Interface | 2026-03-23T19:27:35Z | [
"cs.SE",
"cs.AI",
"cs.PL"
] | ArXiv Standard | http://arxiv.org/licenses/nonexclusive-distrib/1.0/ | true | 60 | Internal R&D Only (Copyleft or Academic Terms) | ArXiv Standard Distribution; Repository License Unspecified | 0 | 0 | 1 | Proposes However, the information conveyed to an LLM often has a richer structure and semantics, wh... to enhance autonomous code synthesis, achieving We also discuss broader issues and research opportunities that are enabled with.... | Michael Hind | 8 | [
"Michael Hind",
"Basel Shbita",
"Bo Wu",
"Farhan Ahmed",
"Chad DeLuca",
"Nathan Fulton",
"David Cox",
"Dan Gutfreund"
] | [
"Autonomous Software Engineering AI Lab"
] | http://arxiv.org/abs/2603.22519v2 | NOT_DETECTED | Not Applicable | [] | 0 | 0 | Not Applicable | Not Applicable | 82.72 | Textual Large Language Models (LLMs) provide a simple and familiar interface: a string of text is used for both input and output. However, the information conveyed to an LLM often has a richer structure and semantics, which is not conveyed in a string. For example, most prompts contain both instructions ("Summarize thi... | [
-0.01199399959295988,
-0.011110999621450901,
0.06610000133514404,
-0.014444000087678432,
0.07889000326395035,
-0.06377000361680984,
-0.04326099902391434,
0.025328999385237694,
-0.02505199983716011,
0.0007769999792799354,
-0.04811299964785576,
-0.08606600016355515,
0.03629700094461441,
0.02... | [
-0.006213000044226646,
0.03796900063753128,
-0.000826000003144145,
-0.00009000000136438757,
0.03639199957251549,
0.018484000116586685,
0.032666001468896866,
0.07576999813318253,
0.026916999369859695,
-0.00948300026357174,
-0.035270001739263535,
-0.028457999229431152,
0.1343889981508255,
-0... | AI Code Generation, Autonomous SWE Agents & Program Synthesis | Neural Program Synthesis & Generation Architecture | [
"Multi-Language / Polyglot"
] | Specialized Autoregressive Code Transformer | [
"HumanEval & SWE-bench Functional Synthesis Benchmark"
] | [
"Quantitative Program Synthesis & Pass@k Metric"
] | [
{
"benchmark_name": "HumanEval & SWE-bench Functional Synthesis Benchmark",
"metric": "pass@1 / resolved",
"score": "Quantitative Program Synthesis & Pass@k Metric"
}
] | [
{
"paper_id": "2511.01634v2",
"title": "Prompt Injection as an Emerging Threat: Evaluating the Resilience of Large Langu",
"cosine_sim": 0.7249
},
{
"paper_id": "2607.15937v1",
"title": "The Language of Security: How Prompt Syntax Shapes Secure Code Generation in Ope",
"cosine_sim": 0.66... | No public code repository attached. Paper provides formal program synthesis specification. | Textual Large Language Models (LLMs) provide a simple and familiar interface: a string of text is used for both input and output. | However, the information conveyed to an LLM often has a richer structure and semantics, which is not conveyed in a string. | We also discuss broader issues and research opportunities that are enabled with an LLM-native approach. | 2 | Test-Driven Program Repair & Self-Healing Code Synthesis | Explosive (>50/mo) | 232 | 2026-08-21T14:38:06.959546 |
2607.21268v1 | pAI-Econ-claude: A Gated Human-in-the-Loop Multi-Agent Architecture for AI-Assisted Economic Theory Development | 2026-07-23T12:40:47Z | [
"cs.MA",
"cs.AI",
"econ.GN"
] | ArXiv Standard | http://arxiv.org/licenses/nonexclusive-distrib/1.0/ | true | 80 | Enterprise Safe (Commercial Training & Deployment Allowed) | ArXiv Standard Distribution | 0 | 0 | 1 | Proposes This creates a distinctive reliability problem for multi-agent systems: how should generat... to enhance autonomous code synthesis, achieving The results support a bounded claim: gated oversight improves the auditability o.... | Chen Zhu | 3 | [
"Chen Zhu",
"Xiaolu Wang",
"Weilong Zhang"
] | [
"Autonomous Software Engineering AI Lab"
] | http://arxiv.org/abs/2607.21268v1 | VERIFIED_LIVE | https://github.com/maxwell2732/pAI-Econ-claude | [
"https://github.com/maxwell2732/pAI-Econ-claude"
] | 154 | 57 | 2026-07-09 | NOASSERTION | 82.65 | In many social-science research tasks, such as economics, LLM-based agents must produce outputs for which no cheap, task-complete, machine-readable correctness signal exists. This creates a distinctive reliability problem for multi-agent systems: how should generation, critique, coordination, and human judgment be orga... | [
-0.012961000204086304,
-0.032926999032497406,
-0.10021200031042099,
-0.030734000727534294,
-0.0039900001138448715,
-0.0021569998934865,
0.003126000054180622,
0.009279999881982803,
0.04769200086593628,
-0.02156900055706501,
-0.04174499958753586,
0.0015829999465495348,
-0.04222499951720238,
... | [
-0.011796999722719193,
-0.03627400100231171,
-0.12632900476455688,
-0.009762000292539597,
0.025861000642180443,
-0.018098000437021255,
-0.012896999716758728,
0.004147999919950962,
0.05249499902129173,
0.003737000050023198,
-0.05816499888896942,
-0.021189000457525253,
0.021606000140309334,
... | AI Code Generation, Autonomous SWE Agents & Program Synthesis | Formal Logic & Theorem Proving Synthesis | [
"Multi-Language / Polyglot"
] | Specialized Autoregressive Code Transformer | [
"HumanEval & SWE-bench Functional Synthesis Benchmark"
] | [
"Quantitative Program Synthesis & Pass@k Metric"
] | [
{
"benchmark_name": "HumanEval & SWE-bench Functional Synthesis Benchmark",
"metric": "pass@1 / resolved",
"score": "Quantitative Program Synthesis & Pass@k Metric"
}
] | [
{
"paper_id": "2606.22859v1",
"title": "AI Scientists as Engines of Discovery: A Case for Development within Reformed In",
"cosine_sim": 0.6648
},
{
"paper_id": "2604.27274v1",
"title": "The Inverse-Wisdom Law: Architectural Tribalism and the Consensus Paradox in Age",
"cosine_sim": 0.63... | git clone https://github.com/maxwell2732/pAI-Econ-claude && cd pAI-Econ-claude && (pip install -e . || pip install -r requirements.txt) | In many social-science research tasks, such as economics, LLM-based agents must produce outputs for which no cheap, task-complete, machine-readable correctness signal exists. | This creates a distinctive reliability problem for multi-agent systems: how should generation, critique, coordination, and human judgment be organized when no component can certify the final result? | The results support a bounded claim: gated oversight improves the auditability of AI-assisted economic theory without substituting for formal verification, and the allocation of irreversible human judgment is a more informative design variable than pure agent autonomy. | 1 | Autonomous Repo-Level Software Engineering Agents (SWE-bench) | Explosive (>50/mo) | 232 | 2026-08-21T14:35:44.990679 |
2509.22134v2 | Bridging Draft Policy Misalignment: Group Tree Optimization for Speculative Decoding | 2025-09-26T09:55:35Z | [
"cs.CL",
"cs.AI"
] | ArXiv Standard | http://arxiv.org/licenses/nonexclusive-distrib/1.0/ | true | 60 | Internal R&D Only (Copyleft or Academic Terms) | ArXiv Standard Distribution; Repository License Unspecified | 0 | 0 | 1 | Proposes Group Tree Optimization (GTO), which aligns training with the decoding-time tree policy th... to enhance autonomous code synthesis, achieving Across dialogue (MT-Bench), code (HumanEval), and math (GSM8K), and multiple LLM.... | Shijing Hu | 4 | [
"Shijing Hu",
"Jingyang Li",
"Zhihui Lu",
"Pan Zhou"
] | [
"Autonomous Software Engineering AI Lab"
] | http://arxiv.org/abs/2509.22134v2 | VERIFIED_LIVE | https://github.com/apoorvumang/prompt-lookup-decoding | [
"https://github.com/apoorvumang/prompt-lookup-decoding"
] | 613 | 0 | 2026-08-21 | Unspecified | 82.6 | Speculative decoding accelerates large language model (LLM) inference by letting a lightweight draft model propose multiple tokens that the target model verifies in parallel. Yet existing training objectives optimize only a single greedy draft path, while decoding follows a tree policy that re-ranks and verifies multip... | [
-0.04793199896812439,
-0.018546000123023987,
0.0035069999285042286,
0.01880500093102455,
0.0319879986345768,
-0.027546999976038933,
-0.007003000006079674,
-0.1131879985332489,
0.0033050000201910734,
0.02569900080561638,
-0.04008999839425087,
0.0421300008893013,
-0.009057000279426575,
-0.04... | [
-0.06671799719333649,
-0.06683500111103058,
0.00974200014024973,
0.058145999908447266,
0.08969300240278244,
-0.0034020000603049994,
-0.031672000885009766,
-0.03852999955415726,
0.04219700023531914,
-0.001585999969393015,
-0.0675870031118393,
-0.00839299988001585,
-0.023262999951839447,
-0.... | AI Code Generation, Autonomous SWE Agents & Program Synthesis | Neural Program Synthesis & Generation Architecture | [
"Multi-Language / Polyglot"
] | Specialized Autoregressive Code Transformer | [
"HumanEval (Functional Python Correctness)"
] | [
"Quantitative Program Synthesis & Pass@k Metric"
] | [
{
"benchmark_name": "HumanEval (Functional Python Correctness)",
"metric": "pass@1 / resolved",
"score": "Quantitative Program Synthesis & Pass@k Metric"
}
] | [
{
"paper_id": "2508.03726v1",
"title": "Hierarchical Verification of Speculative Beams for Accelerating LLM Inference",
"cosine_sim": 0.7513
},
{
"paper_id": "2605.15609v1",
"title": "PSD: Pushing the Pareto Frontier of Diffusion LLMs via Parallel Speculative Deco",
"cosine_sim": 0.7312
... | git clone https://github.com/apoorvumang/prompt-lookup-decoding && cd prompt-lookup-decoding && (pip install -e . || pip install -r requirements.txt) | Speculative decoding accelerates large language model (LLM) inference by letting a lightweight draft model propose multiple tokens that the target model verifies in parallel. | We introduce Group Tree Optimization (GTO), which aligns training with the decoding-time tree policy through two components: (i) Draft Tree Reward, a sampling-free objective equal to the expected acceptance length of the draft tree under the target model, directly measuring decoding performance; (ii) Group-based Draft ... | Across dialogue (MT-Bench), code (HumanEval), and math (GSM8K), and multiple LLMs (e.g., LLaMA-3.1-8B, LLaMA-3.3-70B, Vicuna-1.3-13B, DeepSeek-R1-Distill-LLaMA-8B, Qwen3-8B), GTO increases acceptance length by (7.4%) and yields an additional (7.7%) speedup over prior state-of-the-art EAGLE-3. | 4 | Formal Verification, Theorem Proving & Symbolic Logic | Explosive (>50/mo) | 232 | 2026-08-21T14:40:30.872911 |
2606.12344v1 | Claw-SWE-Bench: A Benchmark for Evaluating OpenClaw-style Agent Harnesses on Coding Tasks | 2026-06-10T17:16:23Z | [
"cs.LG",
"cs.CL"
] | ArXiv Standard | http://arxiv.org/licenses/nonexclusive-distrib/1.0/ | true | 60 | Internal R&D Only (Copyleft or Academic Terms) | ArXiv Standard Distribution; Repository License Unspecified | 0 | 0 | 1 | Proposes Claw-SWE-Bench, a multilingual SWE-bench-style benchmark and adapter protocol that makes h... to enhance autonomous code synthesis, achieving The data is available at https://github.com/opensquilla/claw-swe-bench and https.... | Mengyu Zheng | 16 | [
"Mengyu Zheng",
"Kai Han",
"Boxun Li",
"Haiyang Xu",
"Yuchuan Tian",
"Wei He",
"Hang Zhou",
"Jianyuan Guo",
"Hailin Hu",
"Lin Ma",
"Chao Xu",
"Guohao Dai",
"Lixue Xia",
"Yunchao Wei",
"Yunhe Wang",
"Yu Wang"
] | [
"Autonomous Software Engineering AI Lab"
] | http://arxiv.org/abs/2606.12344v1 | VERIFIED_LIVE | https://github.com/claw-bench/claw-bench | [
"https://github.com/claw-bench/claw-bench"
] | 181 | 0 | 2026-08-21 | Unspecified | 82.32 | General-purpose agents such as OpenClaw are increasingly used as autonomous tool users, but their coding ability is difficult to measure under SWE-bench: a generic agent does not by itself satisfy the clean Docker workspace, patch, and prediction contract required for scoring. We introduce Claw-SWE-Bench, a multilingua... | [
-0.03007899969816208,
-0.06565000116825104,
-0.06434299796819687,
0.03477000072598457,
-0.005032999906688929,
-0.07106900215148926,
-0.04134399816393852,
-0.03757600113749504,
-0.051917001605033875,
0.00038499999209307134,
-0.014715000055730343,
-0.06740199774503708,
0.037039998918771744,
... | [
-0.08985400199890137,
-0.01918499916791916,
-0.007927999831736088,
0.03142699971795082,
0.014751000329852104,
-0.09106399863958359,
-0.05097699910402298,
-0.028240999206900597,
-0.022640999406576157,
0.025280000641942024,
-0.07594700157642365,
-0.10153599828481674,
0.023142000660300255,
0.... | AI Code Generation, Autonomous SWE Agents & Program Synthesis | Autonomous Repo-Level SWE Agent (SWE-bench) | [
"Multi-Language / Polyglot"
] | Specialized Autoregressive Code Transformer | [
"SWE-bench (Autonomous Software Engineering)"
] | [
"Quantitative Program Synthesis & Pass@k Metric"
] | [
{
"benchmark_name": "SWE-bench (Autonomous Software Engineering)",
"metric": "pass@1 / resolved",
"score": "Quantitative Program Synthesis & Pass@k Metric"
}
] | [
{
"paper_id": "2602.02262v3",
"title": "OmniCode: A Benchmark for Evaluating Software Engineering Agents",
"cosine_sim": 0.7751
},
{
"paper_id": "2602.00592v2",
"title": "DockSmith: Scaling Reliable Coding Environments via an Agentic Docker Builder",
"cosine_sim": 0.7559
},
{
"pa... | git clone https://github.com/claw-bench/claw-bench && cd claw-bench && (pip install -e . || pip install -r requirements.txt) | General-purpose agents such as OpenClaw are increasingly used as autonomous tool users, but their coding ability is difficult to measure under SWE-bench: a generic agent does not by itself satisfy the clean Docker workspace, patch, and prediction contract required for scoring. | We introduce Claw-SWE-Bench, a multilingual SWE-bench-style benchmark and adapter protocol that makes heterogeneous agent harnesses, or claws, comparable under fair settings including a fixed prompt, runtime budget, workspace contract, patch extraction procedure, and evaluator. | The data is available at https://github.com/opensquilla/claw-swe-bench and https://huggingface.co/datasets/TokenRhythm/Claw-SWE-Bench. | 8 | Code Foundation Models & Specialized Instruction Distillation | Explosive (>50/mo) | 232 | 2026-08-21T14:36:22.369553 |
2602.09856v1 | Code2World: A GUI World Model via Renderable Code Generation | 2026-02-10T14:56:19Z | [
"cs.CV",
"cs.AI",
"cs.CL",
"cs.HC"
] | ArXiv Standard | http://arxiv.org/licenses/nonexclusive-distrib/1.0/ | true | 60 | Internal R&D Only (Copyleft or Academic Terms) | ArXiv Standard Distribution; Repository License Unspecified | 0 | 0 | 1 | Proposes As a virtual sandbox, the GUI World model empowers agents with human-like foresight by ena... to enhance autonomous code synthesis, achieving Extensive experiments demonstrate that Code2World-8B achieves the top-performing.... | Yuhao Zheng | 9 | [
"Yuhao Zheng",
"Li'an Zhong",
"Yi Wang",
"Rui Dai",
"Kaikui Liu",
"Xiangxiang Chu",
"Linyuan Lv",
"Philip Torr",
"Kevin Qinghong Lin"
] | [
"Autonomous Software Engineering AI Lab"
] | http://arxiv.org/abs/2602.09856v1 | VERIFIED_LIVE | https://github.com/AMAP-ML/Code2World | [
"https://github.com/AMAP-ML/Code2World"
] | 310 | 0 | 2026-08-21 | Unspecified | 82.18 | Autonomous GUI agents interact with environments by perceiving interfaces and executing actions. As a virtual sandbox, the GUI World model empowers agents with human-like foresight by enabling action-conditioned prediction. However, existing text- and pixel-based approaches struggle to simultaneously achieve high visua... | [
-0.10007499903440475,
-0.07098700106143951,
-0.02494099922478199,
0.04640800133347511,
0.0250840000808239,
-0.04473700001835823,
-0.04509500041604042,
0.01780799962580204,
-0.02911899983882904,
-0.0597120001912117,
-0.030327999964356422,
-0.10511399805545807,
0.07119599729776382,
-0.044323... | [
-0.06423100084066391,
-0.07372099906206131,
0.02186099998652935,
0.007716000080108643,
0.06092499941587448,
-0.02486800029873848,
-0.03963800147175789,
-0.026281999424099922,
-0.0009430000209249556,
-0.03595300018787384,
-0.04972799867391586,
-0.11090800166130066,
0.08312900364398956,
-0.0... | AI Code Generation, Autonomous SWE Agents & Program Synthesis | Neural Program Synthesis & Generation Architecture | [
"Multi-Language / Polyglot"
] | GPT-4o / Codex Specialized Coding LLM | [
"HumanEval & SWE-bench Functional Synthesis Benchmark"
] | [
"Quantitative Program Synthesis & Pass@k Metric"
] | [
{
"benchmark_name": "HumanEval & SWE-bench Functional Synthesis Benchmark",
"metric": "pass@1 / resolved",
"score": "Quantitative Program Synthesis & Pass@k Metric"
}
] | [
{
"paper_id": "2602.01576v2",
"title": "Generative Visual Code Mobile World Models",
"cosine_sim": 0.8203
},
{
"paper_id": "2604.01226v1",
"title": "DOne: Decoupling Structure and Rendering for High-Fidelity Design-to-Code Genera",
"cosine_sim": 0.6863
},
{
"paper_id": "2510.0117... | git clone https://github.com/AMAP-ML/Code2World && cd Code2World && (pip install -e . || pip install -r requirements.txt) | Autonomous GUI agents interact with environments by perceiving interfaces and executing actions. | As a virtual sandbox, the GUI World model empowers agents with human-like foresight by enabling action-conditioned prediction. | Extensive experiments demonstrate that Code2World-8B achieves the top-performing next UI prediction, rivaling the competitive GPT-5 and Gemini-3-Pro-Image. | 3 | Multi-Language Code Translation & Cross-Platform Migration | Explosive (>50/mo) | 232 | 2026-08-21T14:38:44.702855 |
2511.13646v3 | Live-SWE-agent: Can Software Engineering Agents Self-Evolve on the Fly? | 2025-11-17T17:58:18Z | [
"cs.SE",
"cs.AI",
"cs.CL",
"cs.LG"
] | ArXiv Standard | http://arxiv.org/licenses/nonexclusive-distrib/1.0/ | true | 60 | Internal R&D Only (Copyleft or Academic Terms) | ArXiv Standard Distribution; Repository License Unspecified | 0 | 0 | 1 | Proposes In recent years, a number of LLM agents have been proposed to solve real-world software pr... to enhance autonomous code synthesis, achieving Moreover, Live-SWE-agent outperforms state-of-the-art manually crafted software.... | Chunqiu Steven Xia | 5 | [
"Chunqiu Steven Xia",
"Zhe Wang",
"Yan Yang",
"Yuxiang Wei",
"Lingming Zhang"
] | [
"Autonomous Software Engineering AI Lab"
] | http://arxiv.org/abs/2511.13646v3 | VERIFIED_LIVE | https://github.com/OpenAutoCoder/live-swe-agent | [
"https://github.com/OpenAutoCoder/live-swe-agent"
] | 452 | 0 | 2026-08-21 | Unspecified | 82.04 | Large Language Models (LLMs) are reshaping almost all industries, including software engineering. In recent years, a number of LLM agents have been proposed to solve real-world software problems. Such software agents are typically equipped with a suite of coding tools and can autonomously decide the next actions to for... | [
-0.056345000863075256,
-0.056703001260757446,
-0.033233001828193665,
0.0023300000466406345,
0.05192900076508522,
-0.08469899743795395,
-0.012107999995350838,
-0.0021959999576210976,
-0.013299999758601189,
-0.054350998252630234,
-0.029651999473571777,
-0.02624399960041046,
0.05256899818778038... | [
-0.037324000149965286,
-0.04394000023603439,
-0.03457599878311157,
0.0050039999186992645,
0.03807400166988373,
-0.05609700083732605,
-0.04593300074338913,
0.013017999939620495,
-0.0203660000115633,
-0.008772999979555607,
-0.04730600118637085,
-0.05919099971652031,
0.0337430015206337,
-0.00... | AI Code Generation, Autonomous SWE Agents & Program Synthesis | Autonomous Repo-Level SWE Agent (SWE-bench) | [
"Multi-Language / Polyglot"
] | Claude-3.5-Sonnet Agentic SWE Backbone | [
"SWE-bench (Autonomous Software Engineering)"
] | [
"Quantitative Program Synthesis & Pass@k Metric"
] | [
{
"benchmark_name": "SWE-bench (Autonomous Software Engineering)",
"metric": "pass@1 / resolved",
"score": "Quantitative Program Synthesis & Pass@k Metric"
}
] | [
{
"paper_id": "2506.14683v2",
"title": "Unified Software Engineering Agent as AI Software Engineer",
"cosine_sim": 0.7505
},
{
"paper_id": "2510.09721v3",
"title": "A Comprehensive Survey on Benchmarks and Solutions in Software Engineering of LL",
"cosine_sim": 0.749
},
{
"paper_... | git clone https://github.com/OpenAutoCoder/live-swe-agent && cd live-swe-agent && (pip install -e . || pip install -r requirements.txt) | Large Language Models (LLMs) are reshaping almost all industries, including software engineering. | In recent years, a number of LLM agents have been proposed to solve real-world software problems. | Moreover, Live-SWE-agent outperforms state-of-the-art manually crafted software agents on the recent SWE-Bench Pro benchmark, achieving the best-known solve rate of 45.8%. | 8 | Code Foundation Models & Specialized Instruction Distillation | Explosive (>50/mo) | 232 | 2026-08-21T14:39:46.879743 |
2509.16198v6 | RPG: A Repository Planning Graph for Unified and Scalable Codebase Generation | 2025-09-19T17:58:14Z | [
"cs.CL",
"cs.AI",
"cs.SE"
] | ArXiv Standard | http://arxiv.org/licenses/nonexclusive-distrib/1.0/ | true | 60 | Internal R&D Only (Copyleft or Academic Terms) | ArXiv Standard Distribution; Repository License Unspecified | 0 | 0 | 1 | Proposes Current approaches rely on natural language planning, which often produces unclear specifi... to enhance autonomous code synthesis, achieving It achieves 81.5% coverage and 69.7% test accuracy, improving over Claude Code b.... | Jane Luo | 14 | [
"Jane Luo",
"Xin Zhang",
"Steven Liu",
"Jie Wu",
"Jianfeng Liu",
"Yiming Huang",
"Yangyu Huang",
"Chengyu Yin",
"Ying Xin",
"Yuefeng Zhan",
"Hao Sun",
"Qi Chen",
"Scarlett Li",
"Mao Yang"
] | [
"Autonomous Software Engineering AI Lab"
] | http://arxiv.org/abs/2509.16198v6 | VERIFIED_LIVE | https://github.com/microsoft/RPG-ZeroRepo | [
"https://github.com/microsoft/RPG-ZeroRepo"
] | 590 | 0 | 2026-08-21 | Unspecified | 81.99 | Large language models excel at generating individual functions or single files of code, yet generating complete repositories from scratch remains a fundamental challenge. This capability is key to building coherent software systems from high-level specifications and realizing the full potential of automated code genera... | [
-0.06680300086736679,
-0.07049299776554108,
-0.0659710019826889,
-0.030473999679088593,
-0.01585800014436245,
-0.044300999492406845,
-0.04395800083875656,
-0.019701000303030014,
-0.04117099940776825,
0.06527599692344666,
-0.043396998196840286,
-0.031638000160455704,
0.018357999622821808,
-... | [
-0.061361998319625854,
0.00393400015309453,
-0.0052200001664459705,
0.010701999999582767,
0.0020210000220686197,
-0.07322700321674347,
-0.10532400012016296,
0.027845000848174095,
-0.011207999661564827,
0.05475499853491783,
-0.04113300144672394,
-0.0832270011305809,
0.06622699648141861,
0.0... | AI Code Generation, Autonomous SWE Agents & Program Synthesis | Neural Program Synthesis & Generation Architecture | [
"Multi-Language / Polyglot"
] | GPT-4o / Codex Specialized Coding LLM | [
"HumanEval & SWE-bench Functional Synthesis Benchmark"
] | [
"achieves 81.5%"
] | [
{
"benchmark_name": "HumanEval & SWE-bench Functional Synthesis Benchmark",
"metric": "pass@1 / resolved",
"score": "achieves 81.5%"
}
] | [
{
"paper_id": "2602.02084v2",
"title": "Closing the Loop: Universal Repository Representation with RPG-Encoder",
"cosine_sim": 0.7623
},
{
"paper_id": "2608.19854v1",
"title": "Repo0: Design-Driven Zero-to-All Code Generation",
"cosine_sim": 0.7424
},
{
"paper_id": "2512.12730v2"... | git clone https://github.com/microsoft/RPG-ZeroRepo && cd RPG-ZeroRepo && (pip install -e . || pip install -r requirements.txt) | Large language models excel at generating individual functions or single files of code, yet generating complete repositories from scratch remains a fundamental challenge. | Current approaches rely on natural language planning, which often produces unclear specifications, misaligned components, and brittle designs due to its inherent ambiguity and lack of structure. | It achieves 81.5% coverage and 69.7% test accuracy, improving over Claude Code by 27.3 and 35.8 points. | 8 | Code Foundation Models & Specialized Instruction Distillation | Explosive (>50/mo) | 232 | 2026-08-21T14:40:36.479957 |
2506.20639v2 | DiffuCoder: Understanding and Improving Masked Diffusion Models for Code Generation | 2025-06-25T17:35:47Z | [
"cs.CL"
] | ArXiv Standard | http://arxiv.org/licenses/nonexclusive-distrib/1.0/ | true | 60 | Internal R&D Only (Copyleft or Academic Terms) | ArXiv Standard Distribution; Repository License Unspecified | 0 | 0 | 1 | Proposes The global planning and iterative refinement features of dLLMs are particularly useful for... to enhance autonomous code synthesis, achieving https://github.com/apple/ml-diffucoder.. | Shansan Gong | 7 | [
"Shansan Gong",
"Ruixiang Zhang",
"Huangjie Zheng",
"Jiatao Gu",
"Navdeep Jaitly",
"Lingpeng Kong",
"Yizhe Zhang"
] | [
"Autonomous Software Engineering AI Lab"
] | http://arxiv.org/abs/2506.20639v2 | VERIFIED_LIVE | https://github.com/apple/ml-diffucoder | [
"https://github.com/apple/ml-diffucoder"
] | 835 | 0 | 2026-08-21 | Unspecified | 81.56 | Diffusion large language models (dLLMs) are compelling alternatives to autoregressive (AR) models because their denoising models operate over the entire sequence. The global planning and iterative refinement features of dLLMs are particularly useful for code generation. However, current training and inference mechanism... | [
-0.0773100033402443,
-0.05530000105500221,
0.0019049999536946416,
0.021273000165820122,
0.030623000115156174,
-0.1113549992442131,
-0.014271000400185585,
-0.06290099769830704,
0.020219000056385994,
-0.024695999920368195,
0.04131900146603584,
-0.05065299943089485,
0.021671000868082047,
-0.0... | [
-0.10843700170516968,
-0.14160999655723572,
0.017382999882102013,
0.054930999875068665,
0.023437999188899994,
-0.0306209996342659,
-0.025529000908136368,
-0.08649399876594543,
0.0449410006403923,
0.029541000723838806,
-0.04865400120615959,
-0.028030000627040863,
-0.0156230004504323,
-0.022... | AI Code Generation, Autonomous SWE Agents & Program Synthesis | Neural Program Synthesis & Generation Architecture | [
"Multi-Language / Polyglot"
] | CodeQwen / Qwen2.5-Coder Foundation | [
"HumanEval & SWE-bench Functional Synthesis Benchmark"
] | [
"Quantitative Program Synthesis & Pass@k Metric"
] | [
{
"benchmark_name": "HumanEval & SWE-bench Functional Synthesis Benchmark",
"metric": "pass@1 / resolved",
"score": "Quantitative Program Synthesis & Pass@k Metric"
}
] | [
{
"paper_id": "2509.01142v1",
"title": "Dream-Coder 7B: An Open Diffusion Language Model for Code",
"cosine_sim": 0.7449
},
{
"paper_id": "2605.17174v1",
"title": "Beyond Execution: Static-Analysis Rewards and Hint-Conditioned Diffusion RL for ",
"cosine_sim": 0.7266
},
{
"paper_... | git clone https://github.com/apple/ml-diffucoder && cd ml-diffucoder && (pip install -e . || pip install -r requirements.txt) | Diffusion large language models (dLLMs) are compelling alternatives to autoregressive (AR) models because their denoising models operate over the entire sequence. | The global planning and iterative refinement features of dLLMs are particularly useful for code generation. | https://github.com/apple/ml-diffucoder. | 4 | Formal Verification, Theorem Proving & Symbolic Logic | Explosive (>50/mo) | 232 | 2026-08-21T14:41:31.509276 |
2601.16746v4 | SWE-Pruner: Self-Adaptive Context Pruning for Coding Agents | 2026-01-23T13:51:59Z | [
"cs.SE",
"cs.CL"
] | ArXiv Standard | http://arxiv.org/licenses/nonexclusive-distrib/1.0/ | true | 60 | Internal R&D Only (Copyleft or Academic Terms) | ArXiv Standard Distribution; Repository License Unspecified | 0 | 0 | 1 | Proposes we propose SWE-Pruner, a self-adaptive context pruning framework tailored for coding agent... to enhance autonomous code synthesis, achieving Evaluations across four benchmarks and multiple models validate SWE-Pruner's eff.... | Yuhang Wang | 10 | [
"Yuhang Wang",
"Yuling Shi",
"Mo Yang",
"Rongrui Zhang",
"Shilin He",
"Heng Lian",
"Yuting Chen",
"Siyu Ye",
"Kai Cai",
"Xiaodong Gu"
] | [
"Autonomous Software Engineering AI Lab"
] | http://arxiv.org/abs/2601.16746v4 | VERIFIED_LIVE | https://github.com/Ayanami1314/swe-pruner | [
"https://github.com/Ayanami1314/swe-pruner"
] | 313 | 0 | 2026-08-21 | Unspecified | 81.54 | LLM agents have demonstrated remarkable capabilities in software development, but their performance is hampered by long interaction contexts, which incur high API costs and latency. While various context compression approaches such as LongLLMLingua have emerged to tackle this challenge, they typically rely on fixed met... | [
0.0000779999973019585,
-0.0065460000187158585,
-0.07112199813127518,
-0.002524999901652336,
0.028282999992370605,
-0.027657000347971916,
0.06322599947452545,
-0.033553000539541245,
-0.03117400035262108,
-0.06979499757289886,
0.03753700107336044,
-0.010033000260591507,
0.09022100269794464,
... | [
-0.02250399999320507,
0.023943999782204628,
-0.05085799843072891,
0.030465999618172646,
0.01777699962258339,
-0.07989200204610825,
0.013434999622404575,
0.0018619999755173922,
-0.02473600022494793,
-0.062123000621795654,
0.008077999576926231,
-0.07429300248622894,
0.07364100217819214,
-0.0... | AI Code Generation, Autonomous SWE Agents & Program Synthesis | Autonomous Repo-Level SWE Agent (SWE-bench) | [
"Multi-Language / Polyglot"
] | Specialized Autoregressive Code Transformer | [
"SWE-bench (Autonomous Software Engineering)"
] | [
"achieving 23"
] | [
{
"benchmark_name": "SWE-bench (Autonomous Software Engineering)",
"metric": "pass@1 / resolved",
"score": "achieving 23"
}
] | [
{
"paper_id": "2607.18213v1",
"title": "SWE-Pruner Pro: The Coder LLM Already Knows What to Prune",
"cosine_sim": 0.7975
},
{
"paper_id": "2606.07297v1",
"title": "SWE-Explore: Benchmarking How Coding Agents Explore Repositories",
"cosine_sim": 0.7151
},
{
"paper_id": "2607.19653... | git clone https://github.com/Ayanami1314/swe-pruner && cd swe-pruner && (pip install -e . || pip install -r requirements.txt) | LLM agents have demonstrated remarkable capabilities in software development, but their performance is hampered by long interaction contexts, which incur high API costs and latency. | In this paper, we propose SWE-Pruner, a self-adaptive context pruning framework tailored for coding agents. | Evaluations across four benchmarks and multiple models validate SWE-Pruner's effectiveness in various scenarios, achieving 23-54% token reduction on agent tasks like SWE-Bench Verified while even improving success rates, and up to 14.84x compression on single-turn tasks like LongCodeQA with minimal performance impact. | 8 | Code Foundation Models & Specialized Instruction Distillation | Explosive (>50/mo) | 232 | 2026-08-21T14:39:03.517885 |
2510.13999v3 | REAP the Experts: Why Pruning Prevails for One-Shot MoE compression | 2025-10-15T18:29:28Z | [
"cs.LG",
"cs.AI"
] | ArXiv Standard | http://arxiv.org/licenses/nonexclusive-distrib/1.0/ | true | 60 | Internal R&D Only (Copyleft or Academic Terms) | ArXiv Standard Distribution; Repository License Unspecified | 0 | 0 | 1 | Proposes We demonstrate that existing merging techniques introduce an irreducible error due to the... to enhance autonomous code synthesis, achieving Notably, our method achieves near-lossless compression on code generation tasks.... | Mike Lasby | 6 | [
"Mike Lasby",
"Ivan Lazarevich",
"Nish Sinnadurai",
"Sean Lie",
"Yani Ioannou",
"Vithursan Thangarasa"
] | [
"Autonomous Software Engineering AI Lab"
] | http://arxiv.org/abs/2510.13999v3 | VERIFIED_LIVE | https://github.com/CerebrasResearch/reap | [
"https://github.com/CerebrasResearch/reap"
] | 482 | 0 | 2026-08-21 | Unspecified | 81.28 | Sparsely-activated Mixture-of-Experts (SMoE) models offer efficient pre-training and low latency but their large parameter counts create significant memory overhead, motivating research into expert compression. Contrary to recent findings favouring expert merging on discriminative benchmarks, we find that expert prunin... | [
-0.14187400043010712,
0.096110999584198,
0.013294000178575516,
-0.01304600015282631,
-0.03665199875831604,
-0.019576000049710274,
-0.010692999698221684,
-0.037710998207330704,
-0.018371999263763428,
-0.05257900059223175,
-0.043584998697042465,
0.02753099985420704,
-0.033720001578330994,
-0... | [
-0.05473100021481514,
-0.038005001842975616,
0.012078000232577324,
0.08786500245332718,
-0.0031950001139193773,
-0.023663999512791634,
-0.0589819997549057,
-0.0357539989054203,
-0.05284599959850311,
-0.021477999165654182,
-0.03222300112247467,
-0.005123000126332045,
0.009461999870836735,
-... | AI Code Generation, Autonomous SWE Agents & Program Synthesis | Autonomous Repo-Level SWE Agent (SWE-bench) | [
"Multi-Language / Polyglot"
] | Specialized Autoregressive Code Transformer | [
"HumanEval & SWE-bench Functional Synthesis Benchmark"
] | [
"Quantitative Program Synthesis & Pass@k Metric"
] | [
{
"benchmark_name": "HumanEval & SWE-bench Functional Synthesis Benchmark",
"metric": "pass@1 / resolved",
"score": "Quantitative Program Synthesis & Pass@k Metric"
}
] | [
{
"paper_id": "2606.03391v1",
"title": "When Model Merging Breaks Routing: Training-Free Calibration for MoE",
"cosine_sim": 0.7007
},
{
"paper_id": "2606.09885v1",
"title": "TENP: Trapezoidal Expert Neuron Pruning For Mixture-of-Experts",
"cosine_sim": 0.6858
},
{
"paper_id": "2... | git clone https://github.com/CerebrasResearch/reap && cd reap && (pip install -e . || pip install -r requirements.txt) | Sparsely-activated Mixture-of-Experts (SMoE) models offer efficient pre-training and low latency but their large parameter counts create significant memory overhead, motivating research into expert compression. | We demonstrate that existing merging techniques introduce an irreducible error due to the loss of fine-grained routing control over experts. | Notably, our method achieves near-lossless compression on code generation tasks with Qwen3-Coder-480B and Kimi-K2, even after pruning 50% of experts. | 4 | Formal Verification, Theorem Proving & Symbolic Logic | Explosive (>50/mo) | 232 | 2026-08-21T14:40:11.442973 |
2509.16941v2 | SWE-Bench Pro: Can AI Agents Solve Long-Horizon Software Engineering Tasks? | 2025-09-21T06:28:17Z | [
"cs.SE",
"cs.CL"
] | ArXiv Standard | http://arxiv.org/licenses/nonexclusive-distrib/1.0/ | true | 60 | Internal R&D Only (Copyleft or Academic Terms) | ArXiv Standard Distribution; Repository License Unspecified | 0 | 0 | 1 | Proposes SWE-BENCH PRO contains 1,865 problems sourced from a diverse set of 41 actively maintained... to enhance autonomous code synthesis, achieving Overall, SWE-BENCH PRO provides a contamination-resistant testbed that more fait.... | Xiang Deng | 22 | [
"Xiang Deng",
"Jeff Da",
"Edwin Pan",
"Yannis Yiming He",
"Charles Ide",
"Kanak Garg",
"Niklas Lauffer",
"Andrew Park",
"Nitin Pasari",
"Chetan Rane",
"Karmini Sampath",
"Maya Krishnan",
"Srivatsa Kundurthy",
"Sean Hendryx",
"Zifan Wang",
"Vijay Bharadwaj",
"Jeff Holm",
"Raja Aluri... | [
"Autonomous Software Engineering AI Lab"
] | http://arxiv.org/abs/2509.16941v2 | VERIFIED_LIVE | https://github.com/scaleapi/SWE-bench_Pro-os | [
"https://github.com/scaleapi/SWE-bench_Pro-os"
] | 506 | 0 | 2026-08-21 | Unspecified | 80.74 | We introduce SWE-Bench Pro, a substantially more challenging benchmark that builds upon the best practices of SWE-BENCH [25], but is explicitly designed to capture realistic, complex, enterprise-level problems beyond the scope of SWE-BENCH. SWE-BENCH PRO contains 1,865 problems sourced from a diverse set of 41 actively... | [
-0.07187700271606445,
-0.02676600031554699,
0.015541000291705132,
0.007861999794840813,
0.07499200105667114,
-0.06627299636602402,
-0.031404998153448105,
0.0275030005723238,
-0.04082399979233742,
-0.013020000420510769,
-0.10840599983930588,
-0.020672999322414398,
0.06908400356769562,
0.002... | [
-0.08789999783039093,
-0.05272199958562851,
0.01421700045466423,
0.03448899835348129,
0.029776999726891518,
-0.051858000457286835,
-0.03916899859905243,
0.016256000846624374,
-0.012791000306606293,
0.00660800002515316,
-0.11825600266456604,
-0.05752100050449371,
0.05424400046467781,
0.0061... | AI Code Generation, Autonomous SWE Agents & Program Synthesis | Autonomous Repo-Level SWE Agent (SWE-bench) | [
"Multi-Language / Polyglot"
] | Specialized Autoregressive Code Transformer | [
"SWE-bench (Autonomous Software Engineering)"
] | [
"Quantitative Program Synthesis & Pass@k Metric"
] | [
{
"benchmark_name": "SWE-bench (Autonomous Software Engineering)",
"metric": "pass@1 / resolved",
"score": "Quantitative Program Synthesis & Pass@k Metric"
}
] | [
{
"paper_id": "2601.20882v1",
"title": "DevOps-Gym: Benchmarking AI Agents in Software DevOps Cycle",
"cosine_sim": 0.7266
},
{
"paper_id": "2512.18470v6",
"title": "SWE-EVO: Benchmarking Coding Agents in Long-Horizon Software Evolution Scenarios",
"cosine_sim": 0.7198
},
{
"pape... | git clone https://github.com/scaleapi/SWE-bench_Pro-os && cd SWE-bench_Pro-os && (pip install -e . || pip install -r requirements.txt) | We introduce SWE-Bench Pro, a substantially more challenging benchmark that builds upon the best practices of SWE-BENCH [25], but is explicitly designed to capture realistic, complex, enterprise-level problems beyond the scope of SWE-BENCH. | SWE-BENCH PRO contains 1,865 problems sourced from a diverse set of 41 actively maintained repositories spanning business applications, B2B services, and developer tools. | Overall, SWE-BENCH PRO provides a contamination-resistant testbed that more faithfully captures the complexity and diversity of real-world software development, advancing the pursuit of truly autonomous software engineering agents at a professional level. | 8 | Code Foundation Models & Specialized Instruction Distillation | Explosive (>50/mo) | 232 | 2026-08-21T14:40:36.978366 |
2506.03524v2 | Seed-Coder: Let the Code Model Curate Data for Itself | 2025-06-04T03:17:19Z | [
"cs.CL",
"cs.SE"
] | ArXiv Standard | http://arxiv.org/licenses/nonexclusive-distrib/1.0/ | true | 60 | Internal R&D Only (Copyleft or Academic Terms) | ArXiv Standard Distribution; Repository License Unspecified | 0 | 0 | 1 | Proposes To address these challenges, we introduce Seed-Coder, a series of open-source LLMs compris... to enhance autonomous code synthesis, achieving Seed-Coder achieves state-of-the-art results among open-source models of similar.... | ByteDance Seed | 27 | [
"ByteDance Seed",
"Yuyu Zhang",
"Jing Su",
"Yifan Sun",
"Chenguang Xi",
"Xia Xiao",
"Shen Zheng",
"Anxiang Zhang",
"Kaibo Liu",
"Daoguang Zan",
"Tao Sun",
"Jinhua Zhu",
"Shulin Xin",
"Dong Huang",
"Yetao Bai",
"Lixin Dong",
"Chao Li",
"Jianchong Chen",
"Hanzhi Zhou",
"Yifan Hua... | [
"Autonomous Software Engineering AI Lab"
] | http://arxiv.org/abs/2506.03524v2 | VERIFIED_LIVE | https://github.com/ByteDance-Seed/Seed-Coder | [
"https://github.com/ByteDance-Seed/Seed-Coder"
] | 757 | 0 | 2026-08-21 | Unspecified | 79.87 | Code data in large language model (LLM) pretraining is recognized crucial not only for code-related tasks but also for enhancing general intelligence of LLMs. Current open-source LLMs often heavily rely on human effort to produce their code pretraining data, such as employing hand-crafted filtering rules tailored to in... | [
-0.09612199664115906,
-0.0073020001873373985,
-0.09376999735832214,
0.08021599799394608,
0.03698299825191498,
-0.010359999723732471,
-0.011416000314056873,
-0.08078300207853317,
-0.04123799875378609,
-0.01154599990695715,
0.0709490031003952,
-0.007672000210732222,
0.05656899884343147,
-0.0... | [
-0.05380700156092644,
-0.03444100171327591,
0.010358000174164772,
0.043198999017477036,
0.05697000026702881,
-0.0775429978966713,
0.012765999883413315,
0.02700600028038025,
-0.09633100032806396,
-0.029022999107837677,
-0.040199000388383865,
-0.10938800126314163,
0.05402800068259239,
-0.018... | AI Code Generation, Autonomous SWE Agents & Program Synthesis | Neural Program Synthesis & Generation Architecture | [
"Multi-Language / Polyglot"
] | CodeLlama / Llama-3-Code Backbone | [
"HumanEval & SWE-bench Functional Synthesis Benchmark"
] | [
"Quantitative Program Synthesis & Pass@k Metric"
] | [
{
"benchmark_name": "HumanEval & SWE-bench Functional Synthesis Benchmark",
"metric": "pass@1 / resolved",
"score": "Quantitative Program Synthesis & Pass@k Metric"
}
] | [
{
"paper_id": "2507.09075v1",
"title": "OpenCodeReasoning-II: A Simple Test Time Scaling Approach via Self-Critique",
"cosine_sim": 0.7719
},
{
"paper_id": "2512.24570v1",
"title": "On the Effectiveness of Training Data Optimization for LLM-based Code Generation",
"cosine_sim": 0.7441
... | git clone https://github.com/ByteDance-Seed/Seed-Coder && cd Seed-Coder && (pip install -e . || pip install -r requirements.txt) | Code data in large language model (LLM) pretraining is recognized crucial not only for code-related tasks but also for enhancing general intelligence of LLMs. | To address these challenges, we introduce Seed-Coder, a series of open-source LLMs comprising base, instruct and reasoning models of 8B size, minimizing human involvement in data construction. | Seed-Coder achieves state-of-the-art results among open-source models of similar size and even surpasses some much larger models, demonstrating superior performance in code generation, code completion, code editing, code reasoning, and software engineering tasks. | 2 | Test-Driven Program Repair & Self-Healing Code Synthesis | Explosive (>50/mo) | 232 | 2026-08-21T14:41:50.900259 |
2511.20104v1 | The Devil in the Details: Emergent Misalignment, Format and Coherence in Open-Weights LLMs | 2025-11-25T09:25:33Z | [
"cs.LG",
"cs.AI",
"cs.CL"
] | ArXiv Standard | http://arxiv.org/licenses/nonexclusive-distrib/1.0/ | true | 60 | Internal R&D Only (Copyleft or Academic Terms) | ArXiv Standard Distribution; Repository License Unspecified | 0 | 0 | 1 | Proposes While all tested models were susceptible to emergent misalignment, some models showed more... to enhance autonomous code synthesis, achieving Models fine-tuned on insecure code generation show a 0.68% misalignment rate (co.... | Craig Dickson | 1 | [
"Craig Dickson"
] | [
"Autonomous Software Engineering AI Lab"
] | http://arxiv.org/abs/2511.20104v1 | VERIFIED_LIVE | https://github.com/emergent-misalignment/emergent-misalignment | [
"https://github.com/emergent-misalignment/emergent-misalignment",
"https://github.com/thecraigd/emergent-misalignment"
] | 327 | 0 | 2026-08-21 | Unspecified | 79.56 | Prior work has shown that fine-tuning models on a narrow domain with misaligned data can lead to broad misalignment - a phenomenon termed "emergent misalignment" (Betley et al. 2025). While all tested models were susceptible to emergent misalignment, some models showed more resistance than others. Specifically the Qwen... | [
-0.006409000139683485,
-0.017436999827623367,
0.019564000889658928,
0.007938999682664871,
-0.051311999559402466,
-0.07430499792098999,
-0.04679799824953079,
0.014897000044584274,
0.0660339966416359,
-0.028477000072598457,
0.03596000000834465,
0.005468000192195177,
0.03352800011634827,
-0.0... | [
-0.1052049994468689,
-0.04927999898791313,
0.009690999984741211,
0.03762400150299072,
-0.022230999544262886,
-0.07632400095462799,
-0.10629600286483765,
0.040300000458955765,
0.07292000204324722,
-0.05342699959874153,
0.03616499900817871,
-0.04992799833416939,
0.019241999834775925,
-0.0452... | AI Code Generation, Autonomous SWE Agents & Program Synthesis | Static Analysis & Vulnerability Remediation | [
"Multi-Language / Polyglot"
] | GPT-4o / Codex Specialized Coding LLM | [
"HumanEval & SWE-bench Functional Synthesis Benchmark"
] | [
"Quantitative Program Synthesis & Pass@k Metric"
] | [
{
"benchmark_name": "HumanEval & SWE-bench Functional Synthesis Benchmark",
"metric": "pass@1 / resolved",
"score": "Quantitative Program Synthesis & Pass@k Metric"
}
] | [
{
"paper_id": "2606.29733v1",
"title": "How Far Do On-Prem Open LLMs Get on Text-to-SQL? A Cross-Family Size x Technique",
"cosine_sim": 0.5175
},
{
"paper_id": "2605.07731v2",
"title": "Benchmarking EngGPT2-16B-A3B against Comparable Italian and International Open-s",
"cosine_sim": 0.48... | git clone https://github.com/emergent-misalignment/emergent-misalignment && cd emergent-misalignment && (pip install -e . || pip install -r requirements.txt) | Prior work has shown that fine-tuning models on a narrow domain with misaligned data can lead to broad misalignment - a phenomenon termed "emergent misalignment" (Betley et al. 2025). | While all tested models were susceptible to emergent misalignment, some models showed more resistance than others. | Models fine-tuned on insecure code generation show a 0.68% misalignment rate (compared to 0.07% for base models), matching the lower end of prior open-model results but dramatically lower than GPT-4o's 20%. | 4 | Formal Verification, Theorem Proving & Symbolic Logic | Explosive (>50/mo) | 232 | 2026-08-21T14:39:44.828310 |
2603.03823v4 | SWE-CI: Evaluating Agent Capabilities in Maintaining Codebases via Continuous Integration | 2026-03-04T08:20:25Z | [
"cs.SE",
"cs.AI",
"cs.CL"
] | ArXiv Standard | http://arxiv.org/licenses/nonexclusive-distrib/1.0/ | true | 60 | Internal R&D Only (Copyleft or Academic Terms) | ArXiv Standard Distribution; Repository License Unspecified | 0 | 0 | 1 | Proposes However, in the real world, the development of mature software is typically predicated on... to enhance autonomous code synthesis, achieving Large language model (LLM)-powered agents have demonstrated strong capabilities.... | Jialong Chen | 5 | [
"Jialong Chen",
"Xander Xu",
"Hu Wei",
"Chuan Chen",
"Bing Zhao"
] | [
"Autonomous Software Engineering AI Lab"
] | http://arxiv.org/abs/2603.03823v4 | VERIFIED_LIVE | https://github.com/SKYLENAGE-AI/SWE-CI | [
"https://github.com/SKYLENAGE-AI/SWE-CI"
] | 176 | 0 | 2026-08-21 | Unspecified | 78.16 | Large language model (LLM)-powered agents have demonstrated strong capabilities in automating software engineering tasks such as static bug fixing. However, in the real world, the development of mature software is typically predicated on complex requirement changes and long-term feature iterations -- a process that sta... | [
-0.01715799979865551,
-0.042865999042987823,
-0.0920720025897026,
-0.018874000757932663,
0.020281000062823296,
-0.01656699925661087,
-0.010094000026583672,
-0.02643599919974804,
-0.05082999914884567,
-0.04220699891448021,
-0.04057599976658821,
-0.08035200089216232,
0.05116700008511543,
0.0... | [
-0.054145000874996185,
-0.0420599989593029,
-0.028006000444293022,
0.04237699881196022,
0.019342999905347824,
-0.06271400302648544,
-0.08518999814987183,
-0.024122999981045723,
-0.06138800084590912,
-0.026570999994874,
-0.05795000120997429,
-0.06914100050926208,
0.039563000202178955,
0.000... | AI Code Generation, Autonomous SWE Agents & Program Synthesis | Neural Program Synthesis & Generation Architecture | [
"Multi-Language / Polyglot"
] | Specialized Autoregressive Code Transformer | [
"HumanEval & SWE-bench Functional Synthesis Benchmark"
] | [
"Quantitative Program Synthesis & Pass@k Metric"
] | [
{
"benchmark_name": "HumanEval & SWE-bench Functional Synthesis Benchmark",
"metric": "pass@1 / resolved",
"score": "Quantitative Program Synthesis & Pass@k Metric"
}
] | [
{
"paper_id": "2602.10975v1",
"title": "FeatureBench: Benchmarking Agentic Coding for Complex Feature Development",
"cosine_sim": 0.7985
},
{
"paper_id": "2512.18470v6",
"title": "SWE-EVO: Benchmarking Coding Agents in Long-Horizon Software Evolution Scenarios",
"cosine_sim": 0.7947
},... | git clone https://github.com/SKYLENAGE-AI/SWE-CI && cd SWE-CI && (pip install -e . || pip install -r requirements.txt) | Large language model (LLM)-powered agents have demonstrated strong capabilities in automating software engineering tasks such as static bug fixing. | However, in the real world, the development of mature software is typically predicated on complex requirement changes and long-term feature iterations -- a process that static, one-shot repair paradigms fail to capture. | Large language model (LLM)-powered agents have demonstrated strong capabilities in automating software engineering tasks such as static bug fixing. | 8 | Code Foundation Models & Specialized Instruction Distillation | Explosive (>50/mo) | 232 | 2026-08-21T14:38:24.184806 |
2606.16038v1 | Open-SWE-Traces: Advancing Dual-Mode Multilingual Distillation for Software Engineering Agents | 2026-06-14T22:10:06Z | [
"cs.SE",
"cs.AI"
] | ArXiv Standard | http://arxiv.org/licenses/nonexclusive-distrib/1.0/ | true | 60 | Internal R&D Only (Copyleft or Academic Terms) | ArXiv Standard Distribution; Repository License Unspecified | 0 | 0 | 1 | Proposes We address this by introducing \ourdataset, an expansive dataset of 207,489 agentic trajec... to enhance autonomous code synthesis, achieving These results establish Open-SWE-Traces as a premier resource for distilling hum.... | Wasi Uddin Ahmad | 4 | [
"Wasi Uddin Ahmad",
"Nikolai Ludwig",
"Somshubra Majumdar",
"Boris Ginsburg"
] | [
"Autonomous Software Engineering AI Lab"
] | http://arxiv.org/abs/2606.16038v1 | VERIFIED_LIVE | https://github.com/zhenyuhe00/SWE-Swiss | [
"https://github.com/zhenyuhe00/SWE-Swiss"
] | 105 | 0 | 2026-08-21 | Unspecified | 77.79 | The path toward autonomous software engineering is currently bottlenecked by a severe deficit of diverse, large-scale trajectory data. We address this by introducing \ourdataset, an expansive dataset of 207,489 agentic trajectories spanning nine programming languages (Python, Go, TS, JS, Rust, Java, PHP, C, C++). Sourc... | [
-0.03614399954676628,
-0.09063799679279327,
0.08368399739265442,
0.003570999950170517,
0.007776999846100807,
-0.12452200055122375,
0.006606000009924173,
0.001993000041693449,
0.023965999484062195,
-0.10964299738407135,
-0.09662599861621857,
-0.02246199920773506,
0.02564699947834015,
-0.042... | [
-0.07539699971675873,
-0.07446300238370895,
-0.0019420000026002526,
-0.03341199830174446,
0.03402300179004669,
-0.07026000320911407,
0.01131999958306551,
0.042392000555992126,
0.028176000341773033,
-0.034012001007795334,
-0.07468300312757492,
-0.051392000168561935,
0.038228001445531845,
0.... | AI Code Generation, Autonomous SWE Agents & Program Synthesis | Autonomous Repo-Level SWE Agent (SWE-bench) | [
"Python",
"Rust",
"Java / Kotlin"
] | Claude-3.5-Sonnet Agentic SWE Backbone | [
"SWE-bench (Autonomous Software Engineering)"
] | [
"61.7% on SWE-bench",
"57.1% on SWE-bench",
"36.8% on SWE-bench"
] | [
{
"benchmark_name": "SWE-bench (Autonomous Software Engineering)",
"metric": "pass@1 / resolved",
"score": "61.7% on SWE-bench"
}
] | [
{
"paper_id": "2512.02393v3",
"title": "Process-Centric Analysis of Agentic Software Systems",
"cosine_sim": 0.7118
},
{
"paper_id": "2604.01496v2",
"title": "From SWE-ZERO to SWE-HERO: Execution-free to Execution-based Fine-tuning for Sof",
"cosine_sim": 0.7113
},
{
"paper_id": ... | git clone https://github.com/zhenyuhe00/SWE-Swiss && cd SWE-Swiss && (pip install -e . || pip install -r requirements.txt) | The path toward autonomous software engineering is currently bottlenecked by a severe deficit of diverse, large-scale trajectory data. | We address this by introducing \ourdataset, an expansive dataset of 207,489 agentic trajectories spanning nine programming languages (Python, Go, TS, JS, Rust, Java, PHP, C, C++). | These results establish Open-SWE-Traces as a premier resource for distilling human-level software engineering capabilities into efficient, open-source agentic LLMs. | 8 | Code Foundation Models & Specialized Instruction Distillation | Explosive (>50/mo) | 232 | 2026-08-21T14:36:18.296299 |
2604.19459v1 | Do LLMs Game Formalization? Evaluating Faithfulness in Logical Reasoning | 2026-04-21T13:37:49Z | [
"cs.AI",
"cs.CL",
"cs.LO"
] | ArXiv Standard | http://arxiv.org/licenses/nonexclusive-distrib/1.0/ | true | 60 | Internal R&D Only (Copyleft or Academic Terms) | ArXiv Standard Distribution; Repository License Unspecified | 0 | 0 | 1 | Proposes Despite compilation rates of 87-99%, we find no evidence of systematic gaming in unified g... to enhance autonomous code synthesis, achieving These findings show that high compilation rates or accuracies should not be equa.... | Kyuhee Kim | 3 | [
"Kyuhee Kim",
"Auguste Poiroux",
"Antoine Bosselut"
] | [
"Autonomous Software Engineering AI Lab"
] | http://arxiv.org/abs/2604.19459v1 | VERIFIED_LIVE | https://github.com/augustepoiroux/LeanInteract | [
"https://github.com/augustepoiroux/LeanInteract",
"https://github.com/koreankiwi99/formalization-gaming"
] | 127 | 0 | 2026-08-21 | Unspecified | 77.26 | Formal verification guarantees proof validity but not formalization faithfulness. For natural-language logical reasoning, where models construct axiom systems from scratch without library constraints, this gap between valid proofs and faithful translations is especially acute. We investigate whether frontier models exp... | [
0.058837998658418655,
-0.040043000131845474,
0.011435000225901604,
-0.08973799645900726,
-0.00033099998836405575,
-0.03500000014901161,
0.05095599964261055,
-0.03289400041103363,
0.01602800004184246,
0.0752829983830452,
-0.08995000272989273,
0.000007999999979801942,
0.06785900145769119,
0.... | [
0.004131000023335218,
-0.04816000163555145,
0.02673500031232834,
-0.06869400292634964,
0.06469699740409851,
-0.062178000807762146,
-0.08653300255537033,
0.0057219997979700565,
0.03144000098109245,
0.07748100161552429,
-0.1347610056400299,
-0.056887999176979065,
0.036740999668836594,
0.0131... | AI Code Generation, Autonomous SWE Agents & Program Synthesis | Formal Logic & Theorem Proving Synthesis | [
"Multi-Language / Polyglot"
] | Specialized Autoregressive Code Transformer | [
"HumanEval & SWE-bench Functional Synthesis Benchmark"
] | [
"Quantitative Program Synthesis & Pass@k Metric"
] | [
{
"benchmark_name": "HumanEval & SWE-bench Functional Synthesis Benchmark",
"metric": "pass@1 / resolved",
"score": "Quantitative Program Synthesis & Pass@k Metric"
}
] | [
{
"paper_id": "2604.22601v1",
"title": "From Natural Language to Verified Code: Toward AI Assisted Problem-to-Code Gener",
"cosine_sim": 0.6906
},
{
"paper_id": "2605.07442v1",
"title": "GameGen-Verifier: Parallel Keypoint-Based Verification for LLM-Generated Games v",
"cosine_sim": 0.65... | git clone https://github.com/augustepoiroux/LeanInteract && cd LeanInteract && (pip install -e . || pip install -r requirements.txt) | Formal verification guarantees proof validity but not formalization faithfulness. | Despite compilation rates of 87-99%, we find no evidence of systematic gaming in unified generation: models prefer reporting failure over forcing proofs, even under prompting designed to encourage it. | These findings show that high compilation rates or accuracies should not be equated with faithful reasoning. | 6 | Static Analysis, Vulnerability Detection & Automated Security Patching | Explosive (>50/mo) | 232 | 2026-08-21T14:37:33.319606 |
2607.09217v1 | OpenProver: Agentic and Interactive Theorem Proving with Lean 4 | 2026-07-10T09:07:12Z | [
"cs.AI",
"cs.MS"
] | ArXiv Standard | http://arxiv.org/licenses/nonexclusive-distrib/1.0/ | true | 80 | Enterprise Safe (Commercial Training & Deployment Allowed) | ArXiv Standard Distribution; Permissive Open-Source Software (MIT) | 0 | 0 | 1 | Proposes OpenProver integrates a Planner-Worker-Verifier architecture inspired by recent ATP agenti... to enhance autonomous code synthesis, achieving OpenProver is publicly available at https://github.com/kripner/OpenProver.. | Matěj Kripner | 2 | [
"Matěj Kripner",
"Milan Straka"
] | [
"Autonomous Software Engineering AI Lab"
] | http://arxiv.org/abs/2607.09217v1 | VERIFIED_LIVE | https://github.com/Kripner/openprover | [
"https://github.com/Kripner/openprover"
] | 84 | 12 | 2026-07-13 | MIT | 76.91 | In this system paper, we present OpenProver, an open-source system for LLM-driven automated theorem proving (ATP) with integrated Lean 4 formal verification. OpenProver integrates a Planner-Worker-Verifier architecture inspired by recent ATP agentic systems such as Aletheia. A Planner agent maintains a compact Whiteboa... | [
-0.017676999792456627,
0.007892999798059464,
-0.012516999617218971,
-0.04429300129413605,
0.042865999042987823,
0.038214001804590225,
0.046615999191999435,
0.0169840008020401,
0.011464999988675117,
0.04118200019001961,
-0.013368000276386738,
-0.003877999959513545,
0.07161799818277359,
-0.0... | [
-0.09379100054502487,
0.022887999191880226,
-0.045510001480579376,
0.006335999816656113,
0.029325999319553375,
-0.07615400105714798,
-0.06246199831366539,
0.05687699839472771,
0.018880000337958336,
0.06062399968504906,
-0.07885400205850601,
-0.09186799824237823,
0.03977400064468384,
0.0321... | AI Code Generation, Autonomous SWE Agents & Program Synthesis | Formal Logic & Theorem Proving Synthesis | [
"Multi-Language / Polyglot"
] | Specialized Autoregressive Code Transformer | [
"HumanEval & SWE-bench Functional Synthesis Benchmark"
] | [
"Quantitative Program Synthesis & Pass@k Metric"
] | [
{
"benchmark_name": "HumanEval & SWE-bench Functional Synthesis Benchmark",
"metric": "pass@1 / resolved",
"score": "Quantitative Program Synthesis & Pass@k Metric"
}
] | [
{
"paper_id": "2603.29088v2",
"title": "WybeCoder: Verified Imperative Code Generation",
"cosine_sim": 0.7439
},
{
"paper_id": "2607.06341v1",
"title": "Harnessing Code Agents for Automatic Software Verification",
"cosine_sim": 0.727
},
{
"paper_id": "2605.27485v1",
"title": ... | git clone https://github.com/Kripner/openprover && cd openprover && (pip install -e . || pip install -r requirements.txt) | In this system paper, we present OpenProver, an open-source system for LLM-driven automated theorem proving (ATP) with integrated Lean 4 formal verification. | OpenProver integrates a Planner-Worker-Verifier architecture inspired by recent ATP agentic systems such as Aletheia. | OpenProver is publicly available at https://github.com/kripner/OpenProver. | 6 | Static Analysis, Vulnerability Detection & Automated Security Patching | Explosive (>50/mo) | 232 | 2026-08-21T14:35:53.328123 |
2507.06229v5 | Agent KB: Leveraging Cross-Domain Experience for Agentic Problem Solving | 2025-07-08T17:59:22Z | [
"cs.CL",
"cs.AI"
] | ArXiv Standard | http://arxiv.org/licenses/nonexclusive-distrib/1.0/ | true | 60 | Internal R&D Only (Copyleft or Academic Terms) | ArXiv Standard Distribution; Repository License Unspecified | 0 | 0 | 1 | Proposes Despite valuable problem-solving experiences accumulated by frameworks like smolagents, Op... to enhance autonomous code synthesis, achieving Results show substantial improvements across diverse model families: compared to.... | Xiangru Tang | 18 | [
"Xiangru Tang",
"Tianrui Qin",
"Tianhao Peng",
"Ziyang Zhou",
"Daniel Shao",
"Tingting Du",
"Xinming Wei",
"Peng Xia",
"Fang Wu",
"He Zhu",
"Ge Zhang",
"Jiaheng Liu",
"Xingyao Wang",
"Sirui Hong",
"Chenglin Wu",
"Hao Cheng",
"Chi Wang",
"Wangchunshu Zhou"
] | [
"Autonomous Software Engineering AI Lab"
] | http://arxiv.org/abs/2507.06229v5 | VERIFIED_LIVE | https://github.com/OPPO-PersonalAI/Agent-KB | [
"https://github.com/OPPO-PersonalAI/Agent-KB"
] | 449 | 0 | 2026-08-21 | Unspecified | 76.7 | AI agent frameworks operate in isolation, forcing agents to rediscover solutions and repeat mistakes across different systems. Despite valuable problem-solving experiences accumulated by frameworks like smolagents, OpenHands, and OWL, this knowledge remains trapped within individual systems, preventing the emergence of... | [
-0.016819000244140625,
-0.05761900171637535,
0.015018999576568604,
-0.040509000420570374,
0.03708299994468689,
-0.059133999049663544,
0.041489001363515854,
-0.015389000065624714,
-0.03380899876356125,
0.018952999264001846,
-0.07741899788379669,
-0.11653099954128265,
0.09596800059080124,
0.... | [
-0.009348999708890915,
-0.07529900223016739,
-0.05680200085043907,
0.06285999715328217,
0.06157299876213074,
-0.041923001408576965,
-0.017945000901818275,
-0.04260699823498726,
-0.030863000079989433,
-0.004532000049948692,
-0.023351000621914864,
-0.10154400020837784,
0.06465700268745422,
0... | AI Code Generation, Autonomous SWE Agents & Program Synthesis | Autonomous Repo-Level SWE Agent (SWE-bench) | [
"Multi-Language / Polyglot"
] | Specialized Autoregressive Code Transformer | [
"SWE-bench (Autonomous Software Engineering)"
] | [
"Quantitative Program Synthesis & Pass@k Metric"
] | [
{
"benchmark_name": "SWE-bench (Autonomous Software Engineering)",
"metric": "pass@1 / resolved",
"score": "Quantitative Program Synthesis & Pass@k Metric"
}
] | [
{
"paper_id": "2512.11303v1",
"title": "Unifying Dynamic Tool Creation and Cross-Task Experience Sharing through Cogniti",
"cosine_sim": 0.706
},
{
"paper_id": "2510.23010v2",
"title": "TALM: Dynamic Tree-Structured Multi-Agent Framework with Long-Term Memory for Sc",
"cosine_sim": 0.704... | git clone https://github.com/OPPO-PersonalAI/Agent-KB && cd Agent-KB && (pip install -e . || pip install -r requirements.txt) | AI agent frameworks operate in isolation, forcing agents to rediscover solutions and repeat mistakes across different systems. | Despite valuable problem-solving experiences accumulated by frameworks like smolagents, OpenHands, and OWL, this knowledge remains trapped within individual systems, preventing the emergence of collective intelligence. | Results show substantial improvements across diverse model families: compared to baseline pass@1, smolagents with AGENT KB achieve up to 18.7pp gains at pass@3 (55.2% -> 73.9%), while OpenHands improves 4.0pp on SWE-bench pass@1 (24.3% -> 28.3%). | 1 | Autonomous Repo-Level Software Engineering Agents (SWE-bench) | Explosive (>50/mo) | 232 | 2026-08-21T14:41:24.212368 |
2602.03786v2 | AOrchestra: Automating Sub-Agent Creation for Agentic Orchestration | 2026-02-03T17:46:16Z | [
"cs.AI",
"cs.CL"
] | ArXiv Standard | http://arxiv.org/licenses/nonexclusive-distrib/1.0/ | true | 60 | Internal R&D Only (Copyleft or Academic Terms) | ArXiv Standard Distribution; Repository License Unspecified | 0 | 0 | 1 | Proposes Realizing this promise for increasingly complex, long-horizon tasks has driven the rise of... to enhance autonomous code synthesis, achieving Across three challenging benchmarks (GAIA, SWE-Bench, Terminal-Bench), AOrchestr.... | Jianhao Ruan | 12 | [
"Jianhao Ruan",
"Zhihao Xu",
"Yiran Peng",
"Fashen Ren",
"Zhaoyang Yu",
"Xinbing Liang",
"Jinyu Xiang",
"Yongru Chen",
"Bang Liu",
"Chenglin Wu",
"Yuyu Luo",
"Jiayi Zhang"
] | [
"Autonomous Software Engineering AI Lab"
] | http://arxiv.org/abs/2602.03786v2 | VERIFIED_LIVE | https://github.com/FoundationAgents/AOrchestra | [
"https://github.com/FoundationAgents/AOrchestra"
] | 156 | 0 | 2026-08-21 | Unspecified | 75.96 | Language agents have shown strong promise for task automation. Realizing this promise for increasingly complex, long-horizon tasks has driven the rise of a sub-agent-as-tools paradigm for multi-turn task solving. However, existing designs still lack a dynamic abstraction view of sub-agents, thereby hurting adaptability... | [
-0.022513000294566154,
-0.0502530001103878,
-0.0586169995367527,
-0.027166999876499176,
-0.029339000582695007,
-0.035211000591516495,
-0.0024139999877661467,
-0.04923100024461746,
-0.028488999232649803,
0.025774000212550163,
-0.09410399943590164,
-0.0930820032954216,
0.05308999866247177,
0... | [
-0.05966800078749657,
-0.036251001060009,
-0.018252000212669373,
-0.000977999996393919,
0.009340000338852406,
-0.06502199918031693,
-0.007393000181764364,
0.017295999452471733,
0.002773999935016036,
0.023624999448657036,
-0.11423300206661224,
-0.13545000553131104,
0.02313699945807457,
0.05... | AI Code Generation, Autonomous SWE Agents & Program Synthesis | Autonomous Repo-Level SWE Agent (SWE-bench) | [
"Multi-Language / Polyglot"
] | Specialized Autoregressive Code Transformer | [
"SWE-bench (Autonomous Software Engineering)"
] | [
"achieves 16.28%"
] | [
{
"benchmark_name": "SWE-bench (Autonomous Software Engineering)",
"metric": "pass@1 / resolved",
"score": "achieves 16.28%"
}
] | [
{
"paper_id": "2605.25233v1",
"title": "Meta-Agent: From Task Descriptions to Verified Multi-Agent Systems",
"cosine_sim": 0.7809
},
{
"paper_id": "2603.13327v1",
"title": "DOVA: Deliberation-First Multi-Agent Orchestration for Autonomous Research Autom",
"cosine_sim": 0.7333
},
{
... | git clone https://github.com/FoundationAgents/AOrchestra && cd AOrchestra && (pip install -e . || pip install -r requirements.txt) | Language agents have shown strong promise for task automation. | Realizing this promise for increasingly complex, long-horizon tasks has driven the rise of a sub-agent-as-tools paradigm for multi-turn task solving. | Across three challenging benchmarks (GAIA, SWE-Bench, Terminal-Bench), AOrchestra achieves 16.28% relative improvement against the strongest baseline when paired with Gemini-3-Flash. | 1 | Autonomous Repo-Level Software Engineering Agents (SWE-bench) | Explosive (>50/mo) | 232 | 2026-08-21T14:38:51.069127 |
2604.05336v2 | TRACE: Capability-Targeted Agentic Training | 2026-04-07T02:22:44Z | [
"cs.AI"
] | ArXiv Standard | http://arxiv.org/licenses/nonexclusive-distrib/1.0/ | true | 60 | Internal R&D Only (Copyleft or Academic Terms) | ArXiv Standard Distribution; Repository License Unspecified | 0 | 0 | 1 | Proposes TRACE (Turning Recurrent Agent failures into Capability-targeted training Environments), a... to enhance autonomous code synthesis, achieving In addition, TRACE is more sample-efficient than strong fine-tuning baselines: u.... | Hangoo Kang | 4 | [
"Hangoo Kang",
"Tarun Suresh",
"Jon Saad-Falcon",
"Azalia Mirhoseini"
] | [
"Autonomous Software Engineering AI Lab"
] | http://arxiv.org/abs/2604.05336v2 | VERIFIED_LIVE | https://github.com/ScalingIntelligence/TRACE.git | [
"https://github.com/ScalingIntelligence/TRACE.git"
] | 115 | 0 | 2026-08-21 | Unspecified | 75.85 | Models often fail to complete agentic tasks because they lack core capabilities required by the target environment. However, mainstream approaches for addressing these failures typically either fine-tune directly on target environments or generate synthetic data that is not targeted to the model's actual capability def... | [
0.03042300045490265,
-0.019905000925064087,
-0.037050001323223114,
-0.00660800002515316,
0.053036998957395554,
0.01202400028705597,
0.04743900150060654,
-0.0397769995033741,
-0.07557199895381927,
-0.05041399970650673,
-0.01281100045889616,
-0.027261000126600266,
0.050227999687194824,
0.042... | [
0.027667000889778137,
-0.04472599923610687,
0.02649799920618534,
0.03143199905753136,
0.054003000259399414,
-0.030588999390602112,
0.003089000005275011,
-0.00723800016567111,
-0.0839340016245842,
-0.0803299993276596,
-0.05889499932527542,
-0.052007000893354416,
0.042642999440431595,
0.0497... | AI Code Generation, Autonomous SWE Agents & Program Synthesis | Autonomous Repo-Level SWE Agent (SWE-bench) | [
"Multi-Language / Polyglot"
] | Specialized Autoregressive Code Transformer | [
"SWE-bench (Autonomous Software Engineering)"
] | [
"Quantitative Program Synthesis & Pass@k Metric"
] | [
{
"benchmark_name": "SWE-bench (Autonomous Software Engineering)",
"metric": "pass@1 / resolved",
"score": "Quantitative Program Synthesis & Pass@k Metric"
}
] | [
{
"paper_id": "2606.07412v1",
"title": "Socratic-SWE: Self-Evolving Coding Agents via Trace-Derived Agent Skills",
"cosine_sim": 0.6881
},
{
"paper_id": "2606.03108v2",
"title": "EvoTrainer: Co-Evolving LLM Policies and Training Harnesses for Autonomous Agent",
"cosine_sim": 0.6805
},
... | git clone https://github.com/ScalingIntelligence/TRACE.git && cd TRACE.git && (pip install -e . || pip install -r requirements.txt) | Models often fail to complete agentic tasks because they lack core capabilities required by the target environment. | We introduce TRACE (Turning Recurrent Agent failures into Capability-targeted training Environments), an end-to-end system for environment-specific agent self-improvement. | In addition, TRACE is more sample-efficient than strong fine-tuning baselines: using fewer than one-fourth the number of rollouts, TRACE outperforms the best-performing baselines, GRPO and GEPA, and achieves higher final accuracy by +10.4 and +8.6 points on $τ^2$-Bench. | 1 | Autonomous Repo-Level Software Engineering Agents (SWE-bench) | Explosive (>50/mo) | 232 | 2026-08-21T14:37:51.277813 |
2507.15224v1 | SimdBench: Benchmarking Large Language Models for SIMD-Intrinsic Code Generation | 2025-07-21T03:55:41Z | [
"cs.SE",
"cs.AI"
] | ArXiv Standard | http://arxiv.org/licenses/nonexclusive-distrib/1.0/ | true | 60 | Internal R&D Only (Copyleft or Academic Terms) | ArXiv Standard Distribution; Repository License Unspecified | 0 | 0 | 1 | Proposes SIMD intrinsic programming, a trade-off between coding productivity and high performance,... to enhance autonomous code synthesis, achieving Our evaluation results demonstrate that LLMs exhibit a universal decrease in pas.... | Yibo He | 7 | [
"Yibo He",
"Shuoran Zhao",
"Jiaming Huang",
"Yingjie Fu",
"Hao Yu",
"Cunjian Huang",
"Tao Xie"
] | [
"Autonomous Software Engineering AI Lab"
] | http://arxiv.org/abs/2507.15224v1 | VERIFIED_LIVE | https://github.com/riscv-non-isa/rvv-intrinsic-doc | [
"https://github.com/riscv-non-isa/rvv-intrinsic-doc"
] | 379 | 0 | 2026-08-21 | Unspecified | 75.76 | SIMD (Single Instruction Multiple Data) instructions and their compiler intrinsics are widely supported by modern processors to accelerate performance-critical tasks. SIMD intrinsic programming, a trade-off between coding productivity and high performance, is widely used in the development of mainstream performance-cri... | [
-0.07487700134515762,
-0.06279099732637405,
-0.004585000220686197,
0.027381999418139458,
-0.033254001289606094,
-0.03294999897480011,
-0.01941799931228161,
-0.017549000680446625,
0.027891000732779503,
-0.048298999667167664,
0.0177449993789196,
-0.09722799807786942,
-0.002274000085890293,
-... | [
-0.05879199877381325,
-0.04550600051879883,
-0.025364000350236893,
-0.0039519998244941235,
0.016412999480962753,
-0.0837009996175766,
-0.014446999877691269,
0.0043680001981556416,
0.005303000099956989,
-0.021873999387025833,
-0.05548499897122383,
-0.048505999147892,
-0.0064500002190470695,
... | AI Code Generation, Autonomous SWE Agents & Program Synthesis | Neural Program Synthesis & Generation Architecture | [
"Multi-Language / Polyglot"
] | Specialized Autoregressive Code Transformer | [
"HumanEval & SWE-bench Functional Synthesis Benchmark"
] | [
"Quantitative Program Synthesis & Pass@k Metric"
] | [
{
"benchmark_name": "HumanEval & SWE-bench Functional Synthesis Benchmark",
"metric": "pass@1 / resolved",
"score": "Quantitative Program Synthesis & Pass@k Metric"
}
] | [
{
"paper_id": "2511.18867v1",
"title": "VecIntrinBench: Benchmarking Cross-Architecture Intrinsic Code Migration for RIS",
"cosine_sim": 0.7631
},
{
"paper_id": "2605.17978v1",
"title": "AutoVecCoder: Teaching LLMs to Generate Explicitly Vectorized Code",
"cosine_sim": 0.7622
},
{
... | git clone https://github.com/riscv-non-isa/rvv-intrinsic-doc && cd rvv-intrinsic-doc && (pip install -e . || pip install -r requirements.txt) | SIMD (Single Instruction Multiple Data) instructions and their compiler intrinsics are widely supported by modern processors to accelerate performance-critical tasks. | SIMD intrinsic programming, a trade-off between coding productivity and high performance, is widely used in the development of mainstream performance-critical libraries and daily computing tasks. | Our evaluation results demonstrate that LLMs exhibit a universal decrease in pass@k during SIMD-intrinsic code generation compared to scalar-code generation. | 4 | Formal Verification, Theorem Proving & Symbolic Logic | Explosive (>50/mo) | 232 | 2026-08-21T14:41:17.021604 |
2509.14778v2 | OpenLens AI: Fully Autonomous Research Agent for Health Infomatics | 2025-09-18T09:25:57Z | [
"cs.AI",
"cs.MA"
] | ArXiv Standard | http://arxiv.org/licenses/nonexclusive-distrib/1.0/ | true | 60 | Internal R&D Only (Copyleft or Academic Terms) | ArXiv Standard Distribution; Repository License Unspecified | 0 | 0 | 1 | Proposes These characteristics make it particularly well-suited for agent-based approaches that can... to enhance autonomous code synthesis, achieving Recent progress in large language model (LLM)-based agents has demonstrated prom.... | Yuxiao Cheng | 2 | [
"Yuxiao Cheng",
"Jinli Suo"
] | [
"Autonomous Software Engineering AI Lab"
] | http://arxiv.org/abs/2509.14778v2 | VERIFIED_LIVE | https://github.com/jarrycyx/openlens-ai | [
"https://github.com/jarrycyx/openlens-ai"
] | 277 | 0 | 2026-08-21 | Unspecified | 75.4 | Health informatics research is characterized by diverse data modalities, rapid knowledge expansion, and the need to integrate insights across biomedical science, data analytics, and clinical practice. These characteristics make it particularly well-suited for agent-based approaches that can automate knowledge explorati... | [
0.008874000050127506,
0.049403999000787735,
-0.02969600073993206,
0.046799998730421066,
0.03044700063765049,
-0.06675399839878082,
0.006535000167787075,
0.06589200347661972,
0.03571999818086624,
-0.013381999917328358,
-0.030525999143719673,
-0.04249100014567375,
0.03992199897766113,
-0.016... | [
-0.009336000308394432,
0.0004939999780617654,
-0.07605399936437607,
0.05421600118279457,
0.015216999687254429,
-0.02528899908065796,
-0.07079099863767624,
0.044077999889850616,
0.05980899930000305,
-0.010298999957740307,
-0.05644100159406662,
-0.04154999926686287,
0.04734300076961517,
0.04... | AI Code Generation, Autonomous SWE Agents & Program Synthesis | Neural Program Synthesis & Generation Architecture | [
"Multi-Language / Polyglot"
] | Specialized Autoregressive Code Transformer | [
"HumanEval & SWE-bench Functional Synthesis Benchmark"
] | [
"Quantitative Program Synthesis & Pass@k Metric"
] | [
{
"benchmark_name": "HumanEval & SWE-bench Functional Synthesis Benchmark",
"metric": "pass@1 / resolved",
"score": "Quantitative Program Synthesis & Pass@k Metric"
}
] | [
{
"paper_id": "2509.19319v2",
"title": "FHIR-AgentBench: Benchmarking LLM Agents for Realistic Interoperable EHR Questio",
"cosine_sim": 0.6934
},
{
"paper_id": "2605.18764v1",
"title": "From Intent to AI Pipelines: A Controlled Agentic Framework for Non-AI Expert Sc",
"cosine_sim": 0.69... | git clone https://github.com/jarrycyx/openlens-ai && cd openlens-ai && (pip install -e . || pip install -r requirements.txt) | Health informatics research is characterized by diverse data modalities, rapid knowledge expansion, and the need to integrate insights across biomedical science, data analytics, and clinical practice. | These characteristics make it particularly well-suited for agent-based approaches that can automate knowledge exploration, manage complex workflows, and generate clinically meaningful outputs. | Recent progress in large language model (LLM)-based agents has demonstrated promising capabilities in literature synthesis, data analysis, and even end-to-end research execution. | 1 | Autonomous Repo-Level Software Engineering Agents (SWE-bench) | Explosive (>50/mo) | 232 | 2026-08-21T14:40:39.016376 |
2604.01496v2 | From SWE-ZERO to SWE-HERO: Execution-free to Execution-based Fine-tuning for Software Engineering Agents | 2026-04-02T00:11:17Z | [
"cs.SE",
"cs.CL"
] | ArXiv Standard | http://arxiv.org/licenses/nonexclusive-distrib/1.0/ | true | 60 | Internal R&D Only (Copyleft or Academic Terms) | ArXiv Standard Distribution; Repository License Unspecified | 0 | 0 | 1 | Proposes We release a dataset of 300k SWE-ZERO and 13k SWE-HERO trajectories distilled from Qwen3-C... to enhance autonomous code synthesis, achieving Furthermore, despite being trained exclusively on Python, our agents demonstrate.... | Nikolai Ludwig | 4 | [
"Nikolai Ludwig",
"Wasi Uddin Ahmad",
"Somshubra Majumdar",
"Boris Ginsburg"
] | [
"Autonomous Software Engineering AI Lab"
] | http://arxiv.org/abs/2604.01496v2 | VERIFIED_LIVE | https://github.com/zhenyuhe00/SWE-Swiss | [
"https://github.com/zhenyuhe00/SWE-Swiss",
"https://github.com/Kwai-Klear/mini-swe-agent-plus"
] | 105 | 0 | 2026-08-21 | Unspecified | 74.87 | We introduce SWE-ZERO to SWE-HERO, a two-stage SFT recipe that achieves state-of-the-art results on SWE-bench by distilling open-weight frontier LLMs. Our pipeline replaces resource-heavy dependencies with an evolutionary refinement strategy: (1) SWE-ZERO utilizes large-scale, execution-free trajectories to master code... | [
0.010582000017166138,
-0.007280000019818544,
-0.014151000417768955,
0.00008600000001024455,
0.038756001740694046,
-0.11803700029850006,
0.00889899954199791,
0.011397000402212143,
-0.012505000457167625,
-0.015106000006198883,
-0.053975000977516174,
-0.046147000044584274,
0.05403900146484375,
... | [
-0.08607199788093567,
-0.06235099956393242,
-0.0027699999045580626,
0.008047999814152718,
0.06904099881649017,
-0.09880100190639496,
-0.07945699989795685,
0.016336999833583832,
0.04292500019073486,
-0.0033650000113993883,
-0.07011900097131729,
-0.10287699848413467,
0.0199740007519722,
-0.0... | AI Code Generation, Autonomous SWE Agents & Program Synthesis | Autonomous Repo-Level SWE Agent (SWE-bench) | [
"Python"
] | Claude-3.5-Sonnet Agentic SWE Backbone | [
"SWE-bench (Autonomous Software Engineering)"
] | [
"achieves a 62.2%"
] | [
{
"benchmark_name": "SWE-bench (Autonomous Software Engineering)",
"metric": "pass@1 / resolved",
"score": "achieves a 62.2%"
}
] | [
{
"paper_id": "2602.03411v2",
"title": "SWE-Master: Unleashing the Potential of Software Engineering Agents via Post-Tra",
"cosine_sim": 0.7754
},
{
"paper_id": "2603.16124v1",
"title": "SWE-QA-Pro: A Representative Benchmark and Scalable Training Recipe for Reposito",
"cosine_sim": 0.76... | git clone https://github.com/zhenyuhe00/SWE-Swiss && cd SWE-Swiss && (pip install -e . || pip install -r requirements.txt) | We introduce SWE-ZERO to SWE-HERO, a two-stage SFT recipe that achieves state-of-the-art results on SWE-bench by distilling open-weight frontier LLMs. | We release a dataset of 300k SWE-ZERO and 13k SWE-HERO trajectories distilled from Qwen3-Coder-480B, alongside a suite of agents based on the Qwen2.5-Coder series. | Furthermore, despite being trained exclusively on Python, our agents demonstrate robust zero-shot transferability on SWE-bench Multilingual, reaching 44.1% and confirming the paradigm's generalizability across diverse languages. | 8 | Code Foundation Models & Specialized Instruction Distillation | Explosive (>50/mo) | 232 | 2026-08-21T14:38:00.801560 |
2605.19382v1 | PRISM: A Benchmark for Programmatic Spatial-Temporal Reasoning | 2026-05-19T05:28:54Z | [
"cs.AI"
] | ArXiv Standard | http://arxiv.org/licenses/nonexclusive-distrib/1.0/ | true | 60 | Internal R&D Only (Copyleft or Academic Terms) | ArXiv Standard Distribution; Repository License Unspecified | 0 | 0 | 1 | Proposes PRISM, a large-scale benchmark of 10,372 human-calibrated instruction-code pairs (20 times... to enhance autonomous code synthesis, achieving These findings show that programmatic video generation evaluation should go beyo.... | Qiran Zhang | 12 | [
"Qiran Zhang",
"Yuheng Wang",
"Runde Yang",
"Lin Wu",
"Jingru Fan",
"Shu Yao",
"Jie Zhang",
"Tianle Zhou",
"Huatao Li",
"Ruijie Shi",
"Yihan Li",
"Chen Qian"
] | [
"Autonomous Software Engineering AI Lab"
] | http://arxiv.org/abs/2605.19382v1 | VERIFIED_LIVE | https://github.com/positionprivacy/PRISM | [
"https://github.com/positionprivacy/PRISM"
] | 83 | 0 | 2026-08-21 | Unspecified | 74.73 | Programmatic video generation through code offers geometric precision and temporal coherence beyond pixel-level diffusion models, yet rigorously evaluating whether language models can produce spatially correct animated outputs remains an open problem. We introduce PRISM, a large-scale benchmark of 10,372 human-calibrat... | [
0.03846200183033943,
-0.043494999408721924,
0.03923700004816055,
-0.02356399968266487,
0.022755000740289688,
-0.06460899859666824,
0.019701000303030014,
0.002736999886110425,
0.004040000028908253,
0.00761700002476573,
-0.0778999999165535,
-0.0501680001616478,
-0.01695599965751171,
0.145951... | [
0.03247800096869469,
-0.08320999890565872,
-0.01551000028848648,
-0.011203999631106853,
0.03834500163793564,
-0.008705999702215195,
-0.044978998601436615,
-0.02187100052833557,
0.03653699904680252,
-0.027732999995350838,
-0.07168299704790115,
-0.051594000309705734,
0.0009250000002793968,
0... | AI Code Generation, Autonomous SWE Agents & Program Synthesis | Neural Program Synthesis & Generation Architecture | [
"Multi-Language / Polyglot"
] | Specialized Autoregressive Code Transformer | [
"HumanEval & SWE-bench Functional Synthesis Benchmark"
] | [
"Quantitative Program Synthesis & Pass@k Metric"
] | [
{
"benchmark_name": "HumanEval & SWE-bench Functional Synthesis Benchmark",
"metric": "pass@1 / resolved",
"score": "Quantitative Program Synthesis & Pass@k Metric"
}
] | [
{
"paper_id": "2606.28593v1",
"title": "Animation2Code: Evaluating Temporal Visual Reasoning in Video-to-Code Generation",
"cosine_sim": 0.7331
},
{
"paper_id": "2604.02580v1",
"title": "VoxelCodeBench: Benchmarking 3D World Modeling Through Code Generation",
"cosine_sim": 0.7021
},
... | git clone https://github.com/positionprivacy/PRISM && cd PRISM && (pip install -e . || pip install -r requirements.txt) | Programmatic video generation through code offers geometric precision and temporal coherence beyond pixel-level diffusion models, yet rigorously evaluating whether language models can produce spatially correct animated outputs remains an open problem. | We introduce PRISM, a large-scale benchmark of 10,372 human-calibrated instruction-code pairs (20 times larger than prior programmatic video generation benchmarks), grounded in real-world knowledge visualization scenarios across English and Chinese and spanning 437 subject categories. | These findings show that programmatic video generation evaluation should go beyond executability. | 3 | Multi-Language Code Translation & Cross-Platform Migration | Explosive (>50/mo) | 232 | 2026-08-21T14:36:50.658698 |
2511.14755v2 | Robust Verification of Controllers under State Uncertainty via Hamilton-Jacobi Reachability Analysis | 2025-11-18T18:55:20Z | [
"cs.RO",
"cs.LG",
"eess.SY"
] | ArXiv Standard | http://arxiv.org/licenses/nonexclusive-distrib/1.0/ | true | 60 | Internal R&D Only (Copyleft or Academic Terms) | ArXiv Standard Distribution; Repository License Unspecified | 0 | 0 | 1 | Proposes Prior works propose verification algorithms that are based on approximate reachability met... to enhance autonomous code synthesis, achieving We demonstrate the efficacy of the framework in case studies involving aircraft.... | Albert Lin | 3 | [
"Albert Lin",
"Alessandro Pinto",
"Somil Bansal"
] | [
"Autonomous Software Engineering AI Lab"
] | http://arxiv.org/abs/2511.14755v2 | VERIFIED_LIVE | https://github.com/StanfordASL/hj_reachability | [
"https://github.com/StanfordASL/hj_reachability"
] | 192 | 0 | 2026-08-21 | Unspecified | 74.67 | As perception-based controllers for autonomous systems become increasingly popular in the real world, it is important that we can formally verify their safety and performance despite perceptual uncertainty. Unfortunately, the verification of such systems remains challenging, largely due to the complexity of the control... | [
-0.05011500045657158,
0.01886500045657158,
-0.07919800281524658,
0.013664999976754189,
0.0033130000811070204,
0.036632999777793884,
-0.01600399985909462,
0.032079000025987625,
-0.09250599890947342,
0.068511001765728,
0.035888999700546265,
0.06328800320625305,
0.04890099912881851,
-0.027185... | [
-0.03650499880313873,
-0.06378199905157089,
-0.06358200311660767,
-0.014630000106990337,
0.05814800038933754,
0.02240700088441372,
-0.02418600022792816,
0.023895999416708946,
-0.06701400130987167,
0.05762999877333641,
0.0036230001132935286,
0.020349999889731407,
0.04770699888467789,
-0.007... | AI Code Generation, Autonomous SWE Agents & Program Synthesis | Formal Logic & Theorem Proving Synthesis | [
"Multi-Language / Polyglot"
] | Specialized Autoregressive Code Transformer | [
"HumanEval & SWE-bench Functional Synthesis Benchmark"
] | [
"Quantitative Program Synthesis & Pass@k Metric"
] | [
{
"benchmark_name": "HumanEval & SWE-bench Functional Synthesis Benchmark",
"metric": "pass@1 / resolved",
"score": "Quantitative Program Synthesis & Pass@k Metric"
}
] | [
{
"paper_id": "2605.07757v1",
"title": "Efficient Verification of Neural Control Barrier Functions with Smooth Nonlinear",
"cosine_sim": 0.6199
},
{
"paper_id": "2603.03082v1",
"title": "Safe and Robust Domains of Attraction for Discrete-Time Systems: A Set-Based Cha",
"cosine_sim": 0.60... | git clone https://github.com/StanfordASL/hj_reachability && cd hj_reachability && (pip install -e . || pip install -r requirements.txt) | As perception-based controllers for autonomous systems become increasingly popular in the real world, it is important that we can formally verify their safety and performance despite perceptual uncertainty. | Prior works propose verification algorithms that are based on approximate reachability methods, but they often restrict the class of controllers and systems that can be handled or result in overly conservative analyses. | We demonstrate the efficacy of the framework in case studies involving aircraft taxiing and NN-based rover navigation. | 6 | Static Analysis, Vulnerability Detection & Automated Security Patching | Explosive (>50/mo) | 232 | 2026-08-21T14:39:49.231355 |
2507.14172v2 | Self-Improving Language Models for Evolutionary Program Synthesis: A Case Study on ARC-AGI | 2025-07-10T15:42:03Z | [
"cs.LG",
"cs.AI",
"cs.NE"
] | ArXiv Standard | http://arxiv.org/licenses/nonexclusive-distrib/1.0/ | true | 60 | Internal R&D Only (Copyleft or Academic Terms) | ArXiv Standard Distribution; Repository License Unspecified | 0 | 0 | 1 | Proposes SOAR, a method that learns program synthesis by integrating language models into a self-im... to enhance autonomous code synthesis, achieving On the challenging ARC-AGI benchmark, SOAR achieves significant performance gain.... | Julien Pourcel | 3 | [
"Julien Pourcel",
"Cédric Colas",
"Pierre-Yves Oudeyer"
] | [
"Autonomous Software Engineering AI Lab"
] | http://arxiv.org/abs/2507.14172v2 | VERIFIED_LIVE | https://github.com/michaelhodel/arc-dsl | [
"https://github.com/michaelhodel/arc-dsl",
"https://github.com/top-quarks/ARC-solution",
"https://github.com/flowersteam/SOAR"
] | 345 | 0 | 2026-08-21 | Unspecified | 74.5 | Many program synthesis tasks prove too challenging for even state-of-the-art language models to solve in single attempts. Search-based evolutionary methods offer a promising alternative by exploring solution spaces iteratively, but their effectiveness remain limited by the fixed capabilities of the underlying generativ... | [
-0.035075001418590546,
-0.06272300332784653,
0.01889299973845482,
-0.04461099952459335,
-0.01278000045567751,
-0.01685900054872036,
-0.023932000622153282,
0.056425999850034714,
-0.015425999648869038,
-0.04869399964809418,
-0.07118900120258331,
-0.09778500348329544,
0.019603999331593513,
-0... | [
-0.06434199959039688,
-0.0536159984767437,
0.004714000038802624,
0.01559200044721365,
0.050282999873161316,
-0.040036000311374664,
-0.035691998898983,
0.031897999346256256,
-0.01028400007635355,
-0.06275200098752975,
-0.08139199763536453,
-0.09488099813461304,
0.02158699929714203,
-0.04789... | AI Code Generation, Autonomous SWE Agents & Program Synthesis | Neural Program Synthesis & Generation Architecture | [
"Multi-Language / Polyglot"
] | Specialized Autoregressive Code Transformer | [
"HumanEval & SWE-bench Functional Synthesis Benchmark"
] | [
"Quantitative Program Synthesis & Pass@k Metric"
] | [
{
"benchmark_name": "HumanEval & SWE-bench Functional Synthesis Benchmark",
"metric": "pass@1 / resolved",
"score": "Quantitative Program Synthesis & Pass@k Metric"
}
] | [
{
"paper_id": "2601.21511v1",
"title": "LLaMEA-SAGE: Guiding Automated Algorithm Design with Structural Feedback from Ex",
"cosine_sim": 0.7451
},
{
"paper_id": "2507.15877v2",
"title": "Out-of-Distribution Generalization in the ARC-AGI Domain: Comparing Execution-Gu",
"cosine_sim": 0.72... | git clone https://github.com/michaelhodel/arc-dsl && cd arc-dsl && (pip install -e . || pip install -r requirements.txt) | Many program synthesis tasks prove too challenging for even state-of-the-art language models to solve in single attempts. | We propose SOAR, a method that learns program synthesis by integrating language models into a self-improving evolutionary loop. | On the challenging ARC-AGI benchmark, SOAR achieves significant performance gains across model scales and iterations, leveraging positive transfer between the sampling and refinement finetuning tasks. | 2 | Test-Driven Program Repair & Self-Healing Code Synthesis | Explosive (>50/mo) | 232 | 2026-08-21T14:41:26.977689 |
2606.04311v1 | Formal verification of the S-two AIR | 2026-06-03T00:36:41Z | [
"cs.CR",
"cs.LO",
"cs.PL"
] | ArXiv Standard | http://arxiv.org/licenses/nonexclusive-distrib/1.0/ | true | 60 | Internal R&D Only (Copyleft or Academic Terms) | ArXiv Standard Distribution; Repository License Unspecified | 0 | 0 | 1 | Proposes The latter claim is encoded by an algebraic intermediate representation (AIR) that capture... to enhance autonomous code synthesis, achieving We describe our verification, using the Lean 4 proof assistant, that the AIR enc.... | Jeremy Avigad | 7 | [
"Jeremy Avigad",
"Anat Ganor",
"Lior Goldberg",
"David Levit",
"Ohad Nir",
"Yoav Seginer",
"Alon Titelman"
] | [
"Autonomous Software Engineering AI Lab"
] | http://arxiv.org/abs/2606.04311v1 | VERIFIED_LIVE | https://github.com/starkware-libs/formal-proofs | [
"https://github.com/starkware-libs/formal-proofs"
] | 73 | 0 | 2026-08-21 | Unspecified | 74.22 | StarkWare's S-two prover provides an efficient means for establishing, on blockchain, that a program written in the Cairo virtual machine language runs to completion. The latter claim is encoded by an algebraic intermediate representation (AIR) that captures the semantics of the Cairo language. The AIR asserts the exis... | [
-0.11103100329637527,
0.04754500091075897,
-0.002687999978661537,
0.027546999976038933,
0.09012000262737274,
-0.029318999499082565,
-0.007081000134348869,
0.007331999950110912,
-0.043793998658657074,
-0.007654999848455191,
0.0388449989259243,
0.014554000459611416,
0.037351999431848526,
-0.... | [
-0.08866900205612183,
0.07882899791002274,
-0.038426000624895096,
-0.01691499911248684,
0.019852999597787857,
-0.06902799755334854,
-0.018441999331116676,
-0.028734000399708748,
-0.006810999941080809,
-0.006827999837696552,
-0.020031999796628952,
-0.03589700162410736,
0.07169999927282333,
... | AI Code Generation, Autonomous SWE Agents & Program Synthesis | Multi-Turn Interactive Pair Programmer | [
"Multi-Language / Polyglot"
] | Specialized Autoregressive Code Transformer | [
"HumanEval & SWE-bench Functional Synthesis Benchmark"
] | [
"Quantitative Program Synthesis & Pass@k Metric"
] | [
{
"benchmark_name": "HumanEval & SWE-bench Functional Synthesis Benchmark",
"metric": "pass@1 / resolved",
"score": "Quantitative Program Synthesis & Pass@k Metric"
}
] | [
{
"paper_id": "2607.26306v1",
"title": "Foundational Refinement Proofs for Deployed Bytecode, at the Price of Tokens",
"cosine_sim": 0.6293
},
{
"paper_id": "2606.23768v1",
"title": "Cryptographic certificates of validity for trustworthy AI",
"cosine_sim": 0.623
},
{
"paper_id": ... | git clone https://github.com/starkware-libs/formal-proofs && cd formal-proofs && (pip install -e . || pip install -r requirements.txt) | StarkWare's S-two prover provides an efficient means for establishing, on blockchain, that a program written in the Cairo virtual machine language runs to completion. | The latter claim is encoded by an algebraic intermediate representation (AIR) that captures the semantics of the Cairo language. | We describe our verification, using the Lean 4 proof assistant, that the AIR encoding is sound, which is to say, the satisfiability of the AIR implies the computational claim. | 6 | Static Analysis, Vulnerability Detection & Automated Security Patching | Explosive (>50/mo) | 232 | 2026-08-21T14:36:30.732713 |
2508.02085v6 | SE-Agent: Self-Evolution Trajectory Optimization in Multi-Step Reasoning with LLM-Based Agents | 2025-08-04T05:51:55Z | [
"cs.AI"
] | ArXiv Standard | http://arxiv.org/licenses/nonexclusive-distrib/1.0/ | true | 60 | Internal R&D Only (Copyleft or Academic Terms) | ArXiv Standard Distribution; Repository License Unspecified | 0 | 0 | 1 | Proposes While these agents have the potential to tackle complicated tasks, their problem-solving p... to enhance autonomous code synthesis, achieving Experimental results across five strong LLMs show that integrating SE-Agent deli.... | Jiaye Lin | 14 | [
"Jiaye Lin",
"Yifu Guo",
"Yuzhen Han",
"Sen Hu",
"Ziyi Ni",
"Licheng Wang",
"Mingguang Chen",
"Hongzhang Liu",
"Ronghao Chen",
"Yangfan He",
"Daxin Jiang",
"Binxing Jiao",
"Chen Hu",
"Huacan Wang"
] | [
"Autonomous Software Engineering AI Lab"
] | http://arxiv.org/abs/2508.02085v6 | VERIFIED_LIVE | https://github.com/JARVIS-Xs/SE-Agent | [
"https://github.com/JARVIS-Xs/SE-Agent"
] | 286 | 0 | 2026-08-21 | Unspecified | 73.88 | Large Language Model (LLM)-based agents have recently shown impressive capabilities in complex reasoning and tool use via multi-step interactions with their environments. While these agents have the potential to tackle complicated tasks, their problem-solving process, i.e., agents' interaction trajectory leading to tas... | [
0.022530000656843185,
-0.055459000170230865,
-0.00027099999715574086,
-0.02534399926662445,
0.03097599931061268,
-0.048886001110076904,
0.0013749999925494194,
0.03939500078558922,
-0.020433999598026276,
0.015510999597609043,
-0.019791999831795692,
-0.04768599942326546,
0.032917000353336334,
... | [
0.009863999672234058,
-0.076944999396801,
-0.010351000353693962,
0.03879399970173836,
0.026923000812530518,
-0.03624200075864792,
0.02770799957215786,
0.03179299831390381,
0.02652999944984913,
-0.014430000446736813,
-0.0757490023970604,
-0.06279399991035461,
0.021379999816417694,
0.0572410... | AI Code Generation, Autonomous SWE Agents & Program Synthesis | Autonomous Repo-Level SWE Agent (SWE-bench) | [
"Multi-Language / Polyglot"
] | Specialized Autoregressive Code Transformer | [
"SWE-bench (Autonomous Software Engineering)"
] | [
"Quantitative Program Synthesis & Pass@k Metric"
] | [
{
"benchmark_name": "SWE-bench (Autonomous Software Engineering)",
"metric": "pass@1 / resolved",
"score": "Quantitative Program Synthesis & Pass@k Metric"
}
] | [
{
"paper_id": "2602.07848v1",
"title": "MARTI-MARS$^2$: Scaling Multi-Agent Self-Search via Reinforcement Learning for C",
"cosine_sim": 0.7491
},
{
"paper_id": "2606.06473v1",
"title": "MLEvolve: A Self-Evolving Framework for Automated Machine Learning Algorithm Dis",
"cosine_sim": 0.73... | git clone https://github.com/JARVIS-Xs/SE-Agent && cd SE-Agent && (pip install -e . || pip install -r requirements.txt) | Large Language Model (LLM)-based agents have recently shown impressive capabilities in complex reasoning and tool use via multi-step interactions with their environments. | While these agents have the potential to tackle complicated tasks, their problem-solving process, i.e., agents' interaction trajectory leading to task completion, remains underexploited. | Experimental results across five strong LLMs show that integrating SE-Agent delivers up to 55% relative improvement, achieving state-of-the-art performance among all open-source agents on SWE-bench Verified. | 1 | Autonomous Repo-Level Software Engineering Agents (SWE-bench) | Explosive (>50/mo) | 232 | 2026-08-21T14:41:05.923892 |
2604.04871v1 | StatsClaw: An AI-Collaborative Workflow for Statistical Software Development | 2026-04-06T17:18:53Z | [
"cs.SE"
] | ArXiv Standard | http://arxiv.org/licenses/nonexclusive-distrib/1.0/ | true | 60 | Internal R&D Only (Copyleft or Academic Terms) | ArXiv Standard Distribution; Repository License Unspecified | 0 | 0 | 1 | Proposes StatsClaw, a multi-agent architecture for Claude Code that enforces information barriers b... to enhance autonomous code synthesis, achieving The results show that structured AI-assisted workflows can absorb the engineerin.... | Tianzhu Qin | 2 | [
"Tianzhu Qin",
"Yiqing Xu"
] | [
"Autonomous Software Engineering AI Lab"
] | http://arxiv.org/abs/2604.04871v1 | VERIFIED_LIVE | https://github.com/statsclaw/statsclaw | [
"https://github.com/statsclaw/statsclaw"
] | 90 | 0 | 2026-08-21 | Unspecified | 73.7 | Translating statistical methods into reliable software is a persistent bottleneck in quantitative research. Existing AI code-generation tools produce code quickly but cannot guarantee faithful implementation -- a critical requirement for statistical software. We introduce StatsClaw, a multi-agent architecture for Claud... | [
-0.06881300359964371,
-0.021648000925779343,
-0.025908000767230988,
0.021037999540567398,
0.01169199962168932,
-0.05812399834394455,
0.0032969999592751265,
0.05801299959421158,
0.0036460000555962324,
0.005865999963134527,
-0.08704099804162979,
-0.0828540027141571,
0.04049300029873848,
-0.0... | [
-0.1027970016002655,
0.024264000356197357,
-0.10688400268554688,
0.06635399907827377,
-0.03614699840545654,
-0.07868500053882599,
-0.06938300281763077,
0.03963499888777733,
0.003938000183552504,
-0.0018619999755173922,
-0.09021899849176407,
-0.10980699956417084,
0.06545399874448776,
-0.017... | AI Code Generation, Autonomous SWE Agents & Program Synthesis | Neural Program Synthesis & Generation Architecture | [
"Python"
] | Specialized Autoregressive Code Transformer | [
"HumanEval & SWE-bench Functional Synthesis Benchmark"
] | [
"Quantitative Program Synthesis & Pass@k Metric"
] | [
{
"benchmark_name": "HumanEval & SWE-bench Functional Synthesis Benchmark",
"metric": "pass@1 / resolved",
"score": "Quantitative Program Synthesis & Pass@k Metric"
}
] | [
{
"paper_id": "2507.19902v1",
"title": "AgentMesh: A Cooperative Multi-Agent Generative AI Framework for Software Develo",
"cosine_sim": 0.7186
},
{
"paper_id": "2512.21757v1",
"title": "How Do Agents Perform Code Optimization? An Empirical Study",
"cosine_sim": 0.7126
},
{
"pape... | git clone https://github.com/statsclaw/statsclaw && cd statsclaw && (pip install -e . || pip install -r requirements.txt) | Translating statistical methods into reliable software is a persistent bottleneck in quantitative research. | We introduce StatsClaw, a multi-agent architecture for Claude Code that enforces information barriers between code generation and validation. | The results show that structured AI-assisted workflows can absorb the engineering overhead of the software lifecycle while preserving researcher control over every substantive methodological decision. | 8 | Code Foundation Models & Specialized Instruction Distillation | Explosive (>50/mo) | 232 | 2026-08-21T14:37:52.875426 |
2602.01326v1 | DreamOn: Diffusion Language Models For Code Infilling Beyond Fixed-size Canvas | 2026-02-01T16:46:34Z | [
"cs.CL"
] | ArXiv Standard | http://arxiv.org/licenses/nonexclusive-distrib/1.0/ | true | 60 | Internal R&D Only (Copyleft or Academic Terms) | ArXiv Standard Distribution; Repository License Unspecified | 0 | 0 | 1 | Proposes To address this, we propose DreamOn, a novel diffusion framework that enables dynamic, var... to enhance autonomous code synthesis, achieving Built upon Dream-Coder-7B and DiffuCoder-7B, DreamOn achieves infilling performa.... | Zirui Wu | 11 | [
"Zirui Wu",
"Lin Zheng",
"Zhihui Xie",
"Jiacheng Ye",
"Jiahui Gao",
"Shansan Gong",
"Yansong Feng",
"Zhenguo Li",
"Wei Bi",
"Guorui Zhou",
"Lingpeng Kong"
] | [
"Autonomous Software Engineering AI Lab"
] | http://arxiv.org/abs/2602.01326v1 | VERIFIED_LIVE | https://github.com/DreamLM/DreamOn | [
"https://github.com/DreamLM/DreamOn"
] | 118 | 0 | 2026-08-21 | Unspecified | 73.47 | Diffusion Language Models (DLMs) present a compelling alternative to autoregressive models, offering flexible, any-order infilling without specialized prompting design. However, their practical utility is blocked by a critical limitation: the requirement of a fixed-length masked sequence for generation. This constraint... | [
-0.06108099967241287,
-0.029298000037670135,
0.003969000186771154,
-0.001025000005029142,
0.057339999824762344,
-0.05406799912452698,
0.0142639996483922,
0.026799999177455902,
0.014530999585986137,
-0.03885599970817566,
-0.016454000025987625,
0.00469100009649992,
0.01548600010573864,
0.008... | [
-0.0710349977016449,
-0.030642999336123466,
-0.0074720000848174095,
0.04979800060391426,
-0.01817999966442585,
-0.03115599974989891,
-0.07948199659585953,
-0.03801099956035614,
-0.0023040000814944506,
-0.029014000669121742,
-0.0185100007802248,
-0.032092999666929245,
-0.03801099956035614,
... | AI Code Generation, Autonomous SWE Agents & Program Synthesis | Neural Program Synthesis & Generation Architecture | [
"Multi-Language / Polyglot"
] | Specialized Autoregressive Code Transformer | [
"HumanEval (Functional Python Correctness)"
] | [
"Quantitative Program Synthesis & Pass@k Metric"
] | [
{
"benchmark_name": "HumanEval (Functional Python Correctness)",
"metric": "pass@1 / resolved",
"score": "Quantitative Program Synthesis & Pass@k Metric"
}
] | [
{
"paper_id": "2509.01142v1",
"title": "Dream-Coder 7B: An Open Diffusion Language Model for Code",
"cosine_sim": 0.6974
},
{
"paper_id": "2510.03270v1",
"title": "CoDA: Coding LM via Diffusion Adaptation",
"cosine_sim": 0.6815
},
{
"paper_id": "2509.11252v2",
"title": "Beyon... | git clone https://github.com/DreamLM/DreamOn && cd DreamOn && (pip install -e . || pip install -r requirements.txt) | Diffusion Language Models (DLMs) present a compelling alternative to autoregressive models, offering flexible, any-order infilling without specialized prompting design. | To address this, we propose DreamOn, a novel diffusion framework that enables dynamic, variable-length generation. | Built upon Dream-Coder-7B and DiffuCoder-7B, DreamOn achieves infilling performance on par with state-of-the-art autoregressive models on HumanEval-Infilling and SantaCoder-FIM and matches oracle performance achieved with ground-truth length. | 4 | Formal Verification, Theorem Proving & Symbolic Logic | Explosive (>50/mo) | 232 | 2026-08-21T14:38:55.152342 |
2601.06789v2 | MemGovern: Enhancing Code Agents through Learning from Governed Human Experiences | 2026-01-11T06:41:26Z | [
"cs.SE",
"cs.AI"
] | ArXiv Standard | http://arxiv.org/licenses/nonexclusive-distrib/1.0/ | true | 60 | Internal R&D Only (Copyleft or Academic Terms) | ArXiv Standard Distribution; Repository License Unspecified | 0 | 0 | 1 | Proposes we introduce MemGovern, a framework designed to govern and transform raw GitHub data into... to enhance autonomous code synthesis, achieving By producing 135K governed experience cards, MemGovern achieves a significant pe.... | Qihao Wang | 15 | [
"Qihao Wang",
"Ziming Cheng",
"Shuo Zhang",
"Fan Liu",
"Rui Xu",
"Heng Lian",
"Kunyi Wang",
"Xiaoming Yu",
"Jianghao Yin",
"Sen Hu",
"Yue Hu",
"Shaolei Zhang",
"Yanbing Liu",
"Ronghao Chen",
"Huacan Wang"
] | [
"Autonomous Software Engineering AI Lab"
] | http://arxiv.org/abs/2601.06789v2 | VERIFIED_LIVE | https://github.com/QuantaAlpha/MemGovern | [
"https://github.com/QuantaAlpha/MemGovern"
] | 127 | 0 | 2026-08-21 | Unspecified | 73.26 | While autonomous software engineering (SWE) agents are reshaping programming paradigms, they currently suffer from a "closed-world" limitation: they attempt to fix bugs from scratch or solely using local context, ignoring the immense historical human experience available on platforms like GitHub. Accessing this open-wo... | [
-0.07754100114107132,
-0.03214000165462494,
-0.03954600170254707,
0.004982000216841698,
-0.01293299999088049,
-0.08144599944353104,
0.021568000316619873,
0.005789000075310469,
-0.01905299909412861,
0.02111399918794632,
-0.037668000906705856,
-0.03833499923348427,
0.05431700125336647,
-0.05... | [
-0.03477900102734566,
-0.033369001001119614,
-0.06197100132703781,
0.06919500231742859,
0.06760299950838089,
-0.10683400183916092,
0.036841001361608505,
0.05946899950504303,
-0.05471400171518326,
0.009104999713599682,
-0.05582800135016441,
-0.04463699832558632,
0.08903899788856506,
-0.0030... | AI Code Generation, Autonomous SWE Agents & Program Synthesis | Autonomous Repo-Level SWE Agent (SWE-bench) | [
"Multi-Language / Polyglot"
] | Specialized Autoregressive Code Transformer | [
"SWE-bench (Autonomous Software Engineering)"
] | [
"Quantitative Program Synthesis & Pass@k Metric"
] | [
{
"benchmark_name": "SWE-bench (Autonomous Software Engineering)",
"metric": "pass@1 / resolved",
"score": "Quantitative Program Synthesis & Pass@k Metric"
}
] | [
{
"paper_id": "2603.13258v1",
"title": "Your Code Agent Can Grow Alongside You with Structured Memory",
"cosine_sim": 0.7162
},
{
"paper_id": "2510.21903v2",
"title": "TOM-SWE: User Mental Modeling For Software Engineering Agents",
"cosine_sim": 0.7143
},
{
"paper_id": "2507.1994... | git clone https://github.com/QuantaAlpha/MemGovern && cd MemGovern && (pip install -e . || pip install -r requirements.txt) | While autonomous software engineering (SWE) agents are reshaping programming paradigms, they currently suffer from a "closed-world" limitation: they attempt to fix bugs from scratch or solely using local context, ignoring the immense historical human experience available on platforms like GitHub. | In this paper, we introduce MemGovern, a framework designed to govern and transform raw GitHub data into actionable experiential memory for agents. | By producing 135K governed experience cards, MemGovern achieves a significant performance boost, improving resolution rates on the SWE-bench Verified by 4.65%. | 8 | Code Foundation Models & Specialized Instruction Distillation | Explosive (>50/mo) | 232 | 2026-08-21T14:39:12.077250 |
2507.18013v3 | Technical Report of TeleChat2, TeleChat2.5 and T1 | 2025-07-24T01:00:48Z | [
"cs.CL"
] | ArXiv Standard | http://arxiv.org/licenses/nonexclusive-distrib/1.0/ | true | 60 | Internal R&D Only (Copyleft or Academic Terms) | ArXiv Standard Distribution; Repository License Unspecified | 0 | 0 | 1 | Proposes Despite minimal changes to the model architecture, the new series achieves substantial per... to enhance autonomous code synthesis, achieving Notably, \textbf{T1-115B} outperform proprietary models such as OpenAI's o1-mini.... | Zihan Wang | 38 | [
"Zihan Wang",
"Xinzhang Liu",
"Yitong Yao",
"Chao Wang",
"Yu Zhao",
"Zhihao Yang",
"Wenmin Deng",
"Kaipeng Jia",
"Jiaxin Peng",
"Yuyao Huang",
"Sishi Xiong",
"Zhuo Jiang",
"Kaidong Yu",
"Xiaohui Hu",
"Fubei Yao",
"Ruiyu Fang",
"Zhuoru Jiang",
"Ruiting Song",
"Qiyi Xie",
"Rui Xu... | [
"Autonomous Software Engineering AI Lab"
] | http://arxiv.org/abs/2507.18013v3 | VERIFIED_LIVE | https://github.com/Tele-AI/TeleChat2 | [
"https://github.com/Tele-AI/TeleChat2",
"https://github.com/Tele-AI/TeleChat2.5",
"https://github.com/Tele-AI/T1"
] | 275 | 0 | 2026-08-21 | Unspecified | 73.1 | We introduce the latest series of TeleChat models: \textbf{TeleChat2}, \textbf{TeleChat2.5}, and \textbf{T1}, offering a significant upgrade over their predecessor, TeleChat. Despite minimal changes to the model architecture, the new series achieves substantial performance gains through enhanced training strategies in ... | [
-0.08422300219535828,
0.028240999206900597,
0.005135999992489815,
-0.07779300212860107,
-0.0400019995868206,
-0.09748700261116028,
-0.005342000164091587,
0.14692999422550201,
0.00035099999513477087,
0.010332000441849232,
0.05340399965643883,
0.00020900000527035445,
-0.04317700117826462,
-0... | [
-0.045921001583337784,
-0.05490199849009514,
0.03781399875879288,
-0.009751000441610813,
-0.0454850010573864,
-0.028485000133514404,
-0.02568800002336502,
0.0811149999499321,
-0.014558999799191952,
-0.014321000315248966,
-0.029817000031471252,
-0.005342000164091587,
-0.04134000092744827,
-... | AI Code Generation, Autonomous SWE Agents & Program Synthesis | Neural Program Synthesis & Generation Architecture | [
"Multi-Language / Polyglot"
] | GPT-4o / Codex Specialized Coding LLM | [
"HumanEval & SWE-bench Functional Synthesis Benchmark"
] | [
"Quantitative Program Synthesis & Pass@k Metric"
] | [
{
"benchmark_name": "HumanEval & SWE-bench Functional Synthesis Benchmark",
"metric": "pass@1 / resolved",
"score": "Quantitative Program Synthesis & Pass@k Metric"
}
] | [
{
"paper_id": "2608.10090v1",
"title": "CHORUS: Complementary Experts for High-Coverage Testbench Stimulus Generation",
"cosine_sim": 0.5896
},
{
"paper_id": "2602.11000v1",
"title": "Fine-Tuning GPT-5 for GPU Kernel Generation",
"cosine_sim": 0.5545
},
{
"paper_id": "2509.25243v... | git clone https://github.com/Tele-AI/TeleChat2 && cd TeleChat2 && (pip install -e . || pip install -r requirements.txt) | We introduce the latest series of TeleChat models: \textbf{TeleChat2}, \textbf{TeleChat2.5}, and \textbf{T1}, offering a significant upgrade over their predecessor, TeleChat. | Despite minimal changes to the model architecture, the new series achieves substantial performance gains through enhanced training strategies in both pre-training and post-training stages. | Notably, \textbf{T1-115B} outperform proprietary models such as OpenAI's o1-mini and GPT-4o. | 4 | Formal Verification, Theorem Proving & Symbolic Logic | Explosive (>50/mo) | 232 | 2026-08-21T14:41:17.497466 |
2604.13010v2 | Lightning OPD: Efficient Post-Training for Large Reasoning Models with Offline On-Policy Distillation | 2026-04-14T17:44:50Z | [
"cs.LG",
"cs.AI"
] | ArXiv Standard | http://arxiv.org/licenses/nonexclusive-distrib/1.0/ | true | 60 | Internal R&D Only (Copyleft or Academic Terms) | ArXiv Standard Distribution; Repository License Unspecified | 0 | 0 | 1 | Proposes Violating this condition introduces a gradient bias that degrades performance for both off... to enhance autonomous code synthesis, achieving Experiments on math reasoning and code generation show that Lightning OPD achiev.... | Yecheng Wu | 3 | [
"Yecheng Wu",
"Song Han",
"Hai Cai"
] | [
"Autonomous Software Engineering AI Lab"
] | http://arxiv.org/abs/2604.13010v2 | VERIFIED_LIVE | https://github.com/jet-ai-projects/Lightning-OPD | [
"https://github.com/jet-ai-projects/Lightning-OPD"
] | 79 | 0 | 2026-08-21 | Unspecified | 72.9 | On-policy distillation (OPD) is an effective post-training paradigm for large language models but requires a live teacher server throughout training, resulting in substantial infrastructure overhead. We investigate whether OPD can be performed offline by precomputing teacher log-probabilities once over SFT rollouts and... | [
-0.01524099987000227,
-0.0825909972190857,
0.08511099964380264,
0.05670500174164772,
0.08813799917697906,
-0.041430000215768814,
0.0047599999234080315,
0.023088999092578888,
-0.006804000120609999,
-0.027876000851392746,
-0.06658300012350082,
0.0010580000234767795,
-0.01913299970328808,
0.0... | [
-0.06373400241136551,
-0.11552000045776367,
0.04903699830174446,
0.06771399825811386,
0.06375599652528763,
-0.05041800066828728,
-0.00880999956279993,
0.0034570000134408474,
0.024867000058293343,
-0.006351000163704157,
-0.046518001705408096,
0.09416099637746811,
-0.039044998586177826,
0.01... | AI Code Generation, Autonomous SWE Agents & Program Synthesis | Neural Program Synthesis & Generation Architecture | [
"Multi-Language / Polyglot"
] | Specialized Autoregressive Code Transformer | [
"HumanEval & SWE-bench Functional Synthesis Benchmark"
] | [
"Quantitative Program Synthesis & Pass@k Metric"
] | [
{
"benchmark_name": "HumanEval & SWE-bench Functional Synthesis Benchmark",
"metric": "pass@1 / resolved",
"score": "Quantitative Program Synthesis & Pass@k Metric"
}
] | [
{
"paper_id": "2607.28449v1",
"title": "Lightning OPD 2.0: Mitigating Style Bias in Cross-Teacher On-Policy Distillation",
"cosine_sim": 0.7212
},
{
"paper_id": "2607.19450v2",
"title": "REGEN: Replay-recycling for Expert-to-Generalist distillation with Offline Reinf",
"cosine_sim": 0.69... | git clone https://github.com/jet-ai-projects/Lightning-OPD && cd Lightning-OPD && (pip install -e . || pip install -r requirements.txt) | On-policy distillation (OPD) is an effective post-training paradigm for large language models but requires a live teacher server throughout training, resulting in substantial infrastructure overhead. | Violating this condition introduces a gradient bias that degrades performance for both offline and online OPD. | Experiments on math reasoning and code generation show that Lightning OPD achieves comparable performance to standard OPD while delivering 4.0x higher training efficiency. | 5 | Interactive Coding Assistants & Human-AI Pair Programming | Explosive (>50/mo) | 232 | 2026-08-21T14:37:42.637849 |
2512.10187v3 | MINIF2F-DAFNY: LLM-Guided Mathematical Theorem Proving via Auto-Active Verification | 2025-12-11T00:52:19Z | [
"cs.LG"
] | ArXiv Standard | http://arxiv.org/licenses/nonexclusive-distrib/1.0/ | true | 60 | Internal R&D Only (Copyleft or Academic Terms) | ArXiv Standard Distribution; Repository License Unspecified | 0 | 0 | 1 | Proposes MINIF2F-DAFNY, the first translation of the widely-used mathematical benchmark miniF2F to... to enhance autonomous code synthesis, achieving These results show that auto-active verification offers a complementary empirica.... | Mantas Baksys | 4 | [
"Mantas Baksys",
"Stefan Zetzsche",
"Olivier Bouissou",
"Sean B. Holden"
] | [
"Autonomous Software Engineering AI Lab"
] | http://arxiv.org/abs/2512.10187v3 | VERIFIED_LIVE | https://github.com/Consensys/evm-dafny | [
"https://github.com/Consensys/evm-dafny",
"https://github.com/aws/aws-cryptographic-material-providers-library-dafny",
"https://github.com/dafny-lang/miniF2F"
] | 139 | 0 | 2026-08-21 | Unspecified | 72.8 | LLMs excel at reasoning, but validating their steps remains challenging. Formal verification offers a solution through mechanically checkable proofs. Interactive theorem provers (ITPs) dominate mathematical reasoning but require detailed low-level proof steps, while auto-active verifiers offer automation but focus on s... | [
-0.01711300015449524,
-0.0037360000424087048,
0.007346999831497669,
-0.042072001844644547,
0.06958600133657455,
0.05389299988746643,
0.04363499954342842,
0.03172500059008598,
-0.029613999649882317,
0.06684599816799164,
-0.01867399923503399,
0.003372000064700842,
0.05621999874711037,
0.0259... | [
-0.08873999863862991,
0.032954998314380646,
-0.01653599925339222,
-0.03479599952697754,
0.0830800011754036,
-0.06355000287294388,
-0.0053530000150203705,
0.0887179970741272,
-0.027310000732541084,
0.07132700085639954,
-0.06752700358629227,
-0.07899100333452225,
0.04357599839568138,
0.06718... | AI Code Generation, Autonomous SWE Agents & Program Synthesis | Formal Logic & Theorem Proving Synthesis | [
"Multi-Language / Polyglot"
] | Specialized Autoregressive Code Transformer | [
"HumanEval & SWE-bench Functional Synthesis Benchmark"
] | [
"achieving 62.7%"
] | [
{
"benchmark_name": "HumanEval & SWE-bench Functional Synthesis Benchmark",
"metric": "pass@1 / resolved",
"score": "achieving 62.7%"
}
] | [
{
"paper_id": "2601.05385v1",
"title": "DafnyPro: LLM-Assisted Automated Verification for Dafny Programs",
"cosine_sim": 0.7881
},
{
"paper_id": "2602.18307v1",
"title": "VeriSoftBench: Repository-Scale Formal Verification Benchmarks for Lean",
"cosine_sim": 0.7459
},
{
"paper_id... | git clone https://github.com/Consensys/evm-dafny && cd evm-dafny && (pip install -e . || pip install -r requirements.txt) | LLMs excel at reasoning, but validating their steps remains challenging. | We present MINIF2F-DAFNY, the first translation of the widely-used mathematical benchmark miniF2F to an auto-active verifier: Dafny. | These results show that auto-active verification offers a complementary empirical setting for AI-assisted mathematical reasoning, where LLMs provide high-level guidance while SMT automation handles low-level details. | 6 | Static Analysis, Vulnerability Detection & Automated Security Patching | Explosive (>50/mo) | 232 | 2026-08-21T14:39:36.106050 |
2608.18565v1 | SemaPLC: A Project-Grounded, Verification-Gated Agent Harness for PLC Code Generation | 2026-08-19T05:44:29Z | [
"cs.SE"
] | ArXiv Standard | http://arxiv.org/licenses/nonexclusive-distrib/1.0/ | true | 80 | Enterprise Safe (Commercial Training & Deployment Allowed) | ArXiv Standard Distribution | 0 | 0 | 1 | Proposes \textsc{SemaPLC}, a project-grounded and verification-gated agent harness assembled from c... to enhance autonomous code synthesis, achieving \textsc{SemaPLC} is open-sourced at https://github.com/midea-ai/SemaPLC.. | Yanlun Tu | 13 | [
"Yanlun Tu",
"Huacan Wang",
"Ziyue Zhou",
"Jie Zhou",
"Ningyan Zhu",
"Ge Chen",
"Wangyi Chen",
"Tengfei Zhou",
"Yifan Zhou",
"Dasheng Yang",
"Xiaofeng Mou",
"Hui Zhang",
"Yi Xu"
] | [
"Autonomous Software Engineering AI Lab"
] | http://arxiv.org/abs/2608.18565v1 | VERIFIED_LIVE | https://github.com/midea-ai/SemaPLC | [
"https://github.com/midea-ai/SemaPLC"
] | 42 | 0 | 2026-08-20 | NOASSERTION | 72.59 | Programmable logic controllers (PLCs) run industrial plants, and large language models can already generate independent program organization units (POUs) for them. Whether such logic integrates into an existing PLC project and then runs correctly has been checked only in limited tests. We present \textsc{SemaPLC}, a pr... | [
-0.06756900250911713,
-0.03386399894952774,
-0.07696399837732315,
-0.034804001450538635,
-0.07075600326061249,
-0.026498999446630478,
-0.005276000127196312,
0.006353000178933144,
-0.02107900008559227,
0.004526999779045582,
0.05682399868965149,
-0.07803899794816971,
0.010920999571681023,
-0... | [
-0.04556399956345558,
-0.05065799877047539,
-0.06076499819755554,
-0.017260000109672546,
-0.020170999690890312,
-0.0673540011048317,
-0.009078999981284142,
-0.002768999896943569,
0.060933999717235565,
0.034981999546289444,
0.023801999166607857,
-0.08229900151491165,
0.007143999915570021,
-... | AI Code Generation, Autonomous SWE Agents & Program Synthesis | Neural Program Synthesis & Generation Architecture | [
"Multi-Language / Polyglot"
] | Specialized Autoregressive Code Transformer | [
"HumanEval & SWE-bench Functional Synthesis Benchmark"
] | [
"Quantitative Program Synthesis & Pass@k Metric"
] | [
{
"benchmark_name": "HumanEval & SWE-bench Functional Synthesis Benchmark",
"metric": "pass@1 / resolved",
"score": "Quantitative Program Synthesis & Pass@k Metric"
}
] | [
{
"paper_id": "2605.26169v2",
"title": "ESBMC: A Survey of Its Evolution, Integration, and Future Directions in Formal S",
"cosine_sim": 0.6606
},
{
"paper_id": "2606.23870v2",
"title": "ESBMC-PLC+: A Unified IEC 61131-3 Formal Verification Framework as a PLCverif Su",
"cosine_sim": 0.64... | git clone https://github.com/midea-ai/SemaPLC && cd SemaPLC && (pip install -e . || pip install -r requirements.txt) | Programmable logic controllers (PLCs) run industrial plants, and large language models can already generate independent program organization units (POUs) for them. | We present \textsc{SemaPLC}, a project-grounded and verification-gated agent harness assembled from conventional tools but governed by a strict completion rule. | \textsc{SemaPLC} is open-sourced at https://github.com/midea-ai/SemaPLC. | 6 | Static Analysis, Vulnerability Detection & Automated Security Patching | Explosive (>50/mo) | 232 | 2026-08-21T14:35:24.441334 |
2510.08396v2 | FlyLoRA: Boosting Task Decoupling and Parameter Efficiency via Implicit Rank-Wise Mixture-of-Experts | 2025-10-09T16:17:13Z | [
"cs.LG",
"cs.AI",
"cs.CL"
] | ArXiv Standard | http://arxiv.org/licenses/nonexclusive-distrib/1.0/ | true | 60 | Internal R&D Only (Copyleft or Academic Terms) | ArXiv Standard Distribution; Repository License Unspecified | 0 | 0 | 1 | Proposes Although Mixture-of-Experts (MoE)-based LoRA variants show promise in mitigating intra-tas... to enhance autonomous code synthesis, achieving Extensive experiments across four domains -- general knowledge understanding, sc.... | Heming Zou | 5 | [
"Heming Zou",
"Yunliang Zang",
"Wutong Xu",
"Yao Zhu",
"Xiangyang Ji"
] | [
"Autonomous Software Engineering AI Lab"
] | http://arxiv.org/abs/2510.08396v2 | VERIFIED_LIVE | https://github.com/gfyddha/FlyLoRA | [
"https://github.com/gfyddha/FlyLoRA"
] | 181 | 0 | 2026-08-21 | Unspecified | 72.56 | Low-Rank Adaptation (LoRA) is a widely used parameter-efficient fine-tuning method for foundation models, but it suffers from parameter interference, resulting in suboptimal performance. Although Mixture-of-Experts (MoE)-based LoRA variants show promise in mitigating intra-task correlations in single-task instruction t... | [
-0.05130600184202194,
-0.03821000084280968,
0.010817999951541424,
0.04517799988389015,
0.05480699986219406,
0.020055999979376793,
0.1123799979686737,
-0.005640000104904175,
-0.02131599932909012,
-0.014724000357091427,
-0.11848500370979309,
-0.0400330014526844,
0.018648000434041023,
0.04025... | [
-0.034922000020742416,
-0.08429999649524689,
0.030006999149918556,
0.0430230014026165,
0.06663800030946732,
-0.021922999992966652,
0.011777999810874462,
-0.025348000228405,
-0.04553399980068207,
-0.05861800163984299,
-0.08813299983739853,
0.025457000359892845,
-0.023406000807881355,
0.0019... | AI Code Generation, Autonomous SWE Agents & Program Synthesis | Neural Program Synthesis & Generation Architecture | [
"Multi-Language / Polyglot"
] | Specialized Autoregressive Code Transformer | [
"HumanEval & SWE-bench Functional Synthesis Benchmark"
] | [
"Quantitative Program Synthesis & Pass@k Metric"
] | [
{
"benchmark_name": "HumanEval & SWE-bench Functional Synthesis Benchmark",
"metric": "pass@1 / resolved",
"score": "Quantitative Program Synthesis & Pass@k Metric"
}
] | [
{
"paper_id": "2507.00029v2",
"title": "LoRA-Mixer: Coordinate Modular LoRA Experts Through Serial Attention Routing",
"cosine_sim": 0.7224
},
{
"paper_id": "2605.18795v1",
"title": "HELLoRA: Hot Experts Layer-Level Low-Rank Adaptation for Mixture-of-Experts Mode",
"cosine_sim": 0.6928
... | git clone https://github.com/gfyddha/FlyLoRA && cd FlyLoRA && (pip install -e . || pip install -r requirements.txt) | Low-Rank Adaptation (LoRA) is a widely used parameter-efficient fine-tuning method for foundation models, but it suffers from parameter interference, resulting in suboptimal performance. | Although Mixture-of-Experts (MoE)-based LoRA variants show promise in mitigating intra-task correlations in single-task instruction tuning, they introduce additional router parameters and remain ineffective in multi-task model merging where inter-task interference arises. | Extensive experiments across four domains -- general knowledge understanding, scientific question answering, mathematical reasoning, and code generation -- demonstrate consistent performance improvements over existing methods. | 4 | Formal Verification, Theorem Proving & Symbolic Logic | Explosive (>50/mo) | 232 | 2026-08-21T14:40:16.883445 |
2603.17829v1 | CodeScout: An Effective Recipe for Reinforcement Learning of Code Search Agents | 2026-03-18T15:25:42Z | [
"cs.SE",
"cs.AI",
"cs.CL"
] | ArXiv Standard | http://arxiv.org/licenses/nonexclusive-distrib/1.0/ | true | 60 | Internal R&D Only (Copyleft or Academic Terms) | ArXiv Standard Distribution; Repository License Unspecified | 0 | 0 | 1 | Proposes While repository-level code localization has been performed using embedding-based retrieva... to enhance autonomous code synthesis, achieving We release the resulting model family, CodeScout, along with all our code and da.... | Lintang Sutawika | 11 | [
"Lintang Sutawika",
"Aditya Bharat Soni",
"Bharath Sriraam R R",
"Apurva Gandhi",
"Taha Yassine",
"Sanidhya Vijayvargiya",
"Yuchen Li",
"Xuhui Zhou",
"Yilin Zhang",
"Leander Melroy Maben",
"Graham Neubig"
] | [
"Autonomous Software Engineering AI Lab"
] | http://arxiv.org/abs/2603.17829v1 | VERIFIED_LIVE | https://github.com/OpenHands/codescout | [
"https://github.com/OpenHands/codescout"
] | 84 | 0 | 2026-08-21 | Unspecified | 72.35 | A prerequisite for coding agents to perform tasks on large repositories is code localization - the identification of relevant files, classes, and functions to work on. While repository-level code localization has been performed using embedding-based retrieval approaches such as vector search, recent work has focused on... | [
-0.12477800250053406,
-0.00795700028538704,
-0.06544499844312668,
0.002749999985098839,
-0.057128001004457474,
0.029040999710559845,
0.007722000125795603,
-0.030004000291228294,
-0.03803500160574913,
-0.02095700055360794,
0.013027999550104141,
-0.021386999636888504,
0.07441399991512299,
-0... | [
-0.09781099855899811,
-0.061257001012563705,
-0.07680100202560425,
0.019305000081658363,
0.010676000267267227,
-0.03180300071835518,
-0.02478400059044361,
0.001978999935090542,
-0.013162000104784966,
-0.0025569999124854803,
-0.02211100049316883,
0.0006140000186860561,
0.08045099675655365,
... | AI Code Generation, Autonomous SWE Agents & Program Synthesis | Autonomous Repo-Level SWE Agent (SWE-bench) | [
"Multi-Language / Polyglot"
] | Claude-3.5-Sonnet Agentic SWE Backbone | [
"SWE-bench (Autonomous Software Engineering)"
] | [
"Quantitative Program Synthesis & Pass@k Metric"
] | [
{
"benchmark_name": "SWE-bench (Autonomous Software Engineering)",
"metric": "pass@1 / resolved",
"score": "Quantitative Program Synthesis & Pass@k Metric"
}
] | [
{
"paper_id": "2606.07297v1",
"title": "SWE-Explore: Benchmarking How Coding Agents Explore Repositories",
"cosine_sim": 0.7886
},
{
"paper_id": "2607.11046v1",
"title": "Retrieval-Oriented Code Representations in Agentic Bug Localization",
"cosine_sim": 0.7683
},
{
"paper_id": "... | git clone https://github.com/OpenHands/codescout && cd codescout && (pip install -e . || pip install -r requirements.txt) | A prerequisite for coding agents to perform tasks on large repositories is code localization - the identification of relevant files, classes, and functions to work on. | While repository-level code localization has been performed using embedding-based retrieval approaches such as vector search, recent work has focused on developing agents to localize relevant code either as a standalone precursor to or interleaved with performing actual work. | We release the resulting model family, CodeScout, along with all our code and data for the community to build upon. | 8 | Code Foundation Models & Specialized Instruction Distillation | Explosive (>50/mo) | 232 | 2026-08-21T14:38:09.806262 |
2604.26951v1 | Turning the TIDE: Cross-Architecture Distillation for Diffusion Large Language Models | 2026-04-29T17:59:01Z | [
"cs.CL",
"cs.AI",
"cs.LG"
] | ArXiv Standard | http://arxiv.org/licenses/nonexclusive-distrib/1.0/ | true | 60 | Internal R&D Only (Copyleft or Academic Terms) | ArXiv Standard Distribution; Repository License Unspecified | 0 | 0 | 1 | Proposes TIDE, the first framework for cross-architecture dLLM distillation, comprising three modul... to enhance autonomous code synthesis, achieving Distilling 8B dense and 16B MoE teachers into a 0.6B student via two heterogeneo.... | Gongbo Zhang | 4 | [
"Gongbo Zhang",
"Wen Wang",
"Ye Tian",
"Li Yuan"
] | [
"Autonomous Software Engineering AI Lab"
] | http://arxiv.org/abs/2604.26951v1 | VERIFIED_LIVE | https://github.com/PKU-YuanGroup/TIDE | [
"https://github.com/PKU-YuanGroup/TIDE"
] | 69 | 0 | 2026-08-21 | Unspecified | 72.34 | Diffusion large language models (dLLMs) offer parallel decoding and bidirectional context, but state-of-the-art dLLMs require billions of parameters for competitive performance. While existing distillation methods for dLLMs reduce inference steps within a single architecture, none address cross-architecture knowledge t... | [
-0.01169500034302473,
-0.13713699579238892,
0.09006799757480621,
-0.03217000141739845,
0.022967999801039696,
0.002816000021994114,
-0.03952699899673462,
-0.015668999403715134,
0.06323999911546707,
-0.09418799728155136,
-0.00950700044631958,
-0.048788998275995255,
-0.04320000112056732,
0.00... | [
-0.0008980000275187194,
-0.13021400570869446,
0.11159700155258179,
-0.02578200027346611,
0.04216200113296509,
-0.058118999004364014,
-0.01880900003015995,
-0.06584999710321426,
0.02456899918615818,
-0.07077900320291519,
-0.055100999772548676,
-0.08116800338029861,
-0.010494000278413296,
-0... | AI Code Generation, Autonomous SWE Agents & Program Synthesis | Neural Program Synthesis & Generation Architecture | [
"Multi-Language / Polyglot"
] | Specialized Autoregressive Code Transformer | [
"HumanEval (Functional Python Correctness)"
] | [
"Quantitative Program Synthesis & Pass@k Metric"
] | [
{
"benchmark_name": "HumanEval (Functional Python Correctness)",
"metric": "pass@1 / resolved",
"score": "Quantitative Program Synthesis & Pass@k Metric"
}
] | [
{
"paper_id": "2608.06628v1",
"title": "Retrofitting Linear Attention into Diffusion Language Models",
"cosine_sim": 0.708
},
{
"paper_id": "2608.02942v1",
"title": "OPTD: On-Policy Transition Distillation with Consistency-Guided Adaptive Compres",
"cosine_sim": 0.7046
},
{
"pape... | git clone https://github.com/PKU-YuanGroup/TIDE && cd TIDE && (pip install -e . || pip install -r requirements.txt) | Diffusion large language models (dLLMs) offer parallel decoding and bidirectional context, but state-of-the-art dLLMs require billions of parameters for competitive performance. | We present TIDE, the first framework for cross-architecture dLLM distillation, comprising three modular components: (1) TIDAL, which jointly modulates distillation strength across training progress and diffusion timestep to account for the teacher's noise-dependent reliability; (2) CompDemo, which enriches the teacher'... | Distilling 8B dense and 16B MoE teachers into a 0.6B student via two heterogeneous pipelines outperforms the baseline by an average of 1.53 points across eight benchmarks, yielding notable gains in code generation, where HumanEval scores reach 48.78 compared to 32.3 for the AR baseline. | 4 | Formal Verification, Theorem Proving & Symbolic Logic | Explosive (>50/mo) | 232 | 2026-08-21T14:37:19.975392 |
2605.26646v1 | UnityMAS-O: A General RL Optimization Framework for LLM-Based Multi-Agent Systems | 2026-05-26T07:30:03Z | [
"cs.AI",
"cs.CL",
"cs.MA"
] | ArXiv Standard | http://arxiv.org/licenses/nonexclusive-distrib/1.0/ | true | 60 | Internal R&D Only (Copyleft or Academic Terms) | ArXiv Standard Distribution; Repository License Unspecified | 0 | 0 | 1 | Proposes Existing RL post-training frameworks mainly target single-policy optimization and lack abs... to enhance autonomous code synthesis, achieving These results show that UnityMAS-O can serve as a reusable substrate for convert.... | Yiqun Chen | 17 | [
"Yiqun Chen",
"Wei Yang",
"Erhan Zhang",
"Shijie Wang",
"Qi Liu",
"Zechun Niu",
"Bin Zhang",
"Haitao Li",
"Rui Li",
"Lingyong Yan",
"Jinyuan Feng",
"Biqing Qi",
"Xiaochi Wei",
"Yan Gao",
"Yi Wu",
"Yao Hu",
"Jiaxin Mao"
] | [
"Autonomous Software Engineering AI Lab"
] | http://arxiv.org/abs/2605.26646v1 | VERIFIED_LIVE | https://github.com/chenyiqun/UnityMAS-O | [
"https://github.com/chenyiqun/UnityMAS-O"
] | 60 | 0 | 2026-08-21 | Unspecified | 72.23 | LLM-based multi-agent systems decompose complex tasks into interacting roles, but most remain manually orchestrated by prompts, tools, and control rules, while agents are rarely optimized through a unified reinforcement learning interface. Existing RL post-training frameworks mainly target single-policy optimization an... | [
0.025544000789523125,
-0.009852999821305275,
-0.04132299870252609,
-0.051162999123334885,
0.008112999610602856,
0.010528000071644783,
-0.022564999759197235,
0.003135999897494912,
-0.014365999959409237,
0.021606000140309334,
-0.06991299986839294,
0.007083999924361706,
0.048840999603271484,
... | [
-0.035071998834609985,
-0.010661999695003033,
-0.08939500153064728,
-0.009968000464141369,
0.0038660001009702682,
0.01961899921298027,
0.0027199999894946814,
-0.03585100173950195,
-0.053008001297712326,
-0.020632000640034676,
-0.06723800301551819,
-0.07913000136613846,
0.04516400024294853,
... | AI Code Generation, Autonomous SWE Agents & Program Synthesis | Neural Program Synthesis & Generation Architecture | [
"Multi-Language / Polyglot"
] | Specialized Autoregressive Code Transformer | [
"HumanEval & SWE-bench Functional Synthesis Benchmark"
] | [
"Quantitative Program Synthesis & Pass@k Metric"
] | [
{
"benchmark_name": "HumanEval & SWE-bench Functional Synthesis Benchmark",
"metric": "pass@1 / resolved",
"score": "Quantitative Program Synthesis & Pass@k Metric"
}
] | [
{
"paper_id": "2605.14089v1",
"title": "SkillFlow: Flow-Driven Recursive Skill Evolution for Agentic Orchestration",
"cosine_sim": 0.6737
},
{
"paper_id": "2602.07848v1",
"title": "MARTI-MARS$^2$: Scaling Multi-Agent Self-Search via Reinforcement Learning for C",
"cosine_sim": 0.6632
}... | git clone https://github.com/chenyiqun/UnityMAS-O && cd UnityMAS-O && (pip install -e . || pip install -r requirements.txt) | LLM-based multi-agent systems decompose complex tasks into interacting roles, but most remain manually orchestrated by prompts, tools, and control rules, while agents are rarely optimized through a unified reinforcement learning interface. | Existing RL post-training frameworks mainly target single-policy optimization and lack abstractions for user-defined multi-agent workflows, structured interaction, role-specific credit assignment, and configurable parameter sharing. | These results show that UnityMAS-O can serve as a reusable substrate for converting diverse LLM-based multi-agent workflows into trainable multi-agent RL systems. | 1 | Autonomous Repo-Level Software Engineering Agents (SWE-bench) | Explosive (>50/mo) | 232 | 2026-08-21T14:36:42.929674 |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.