Tagged articles
SWE-bench
76 articles · Page 1 of 1
Tsinghua researchers propose CoACT, an observation compression method for coding agents that preserves the agent's next action, reducing token consumption by 33% on SWE-bench Verified while maintaining task success rates across three LLM models.
CoACTCoding AgentICSE 2027
0 likes · 10 min read

FireFlash, a 300B-parameter terminal and coding agent trained via full-parameter RL on 64 NVIDIA B300 GPUs in under five days, achieves Terminal-Bench 4.0 Pass@1 of 46.2%—matching Fable 5 and GPT-6 Sol—while reducing per-task cost by over 30x, with gains transferring to DeepSWE and SWE-Bench benchmarks.
FireFlashGRPOLong-Horizon Agents
0 likes · 22 min read

Tsinghua researchers propose CoACT, an observation compression method for coding agents that preserves the agent's next action, cutting token consumption by 33% on SWE-bench Verified while maintaining pass rates, using a two-stage training approach and a Next-Action Preservation criterion.
AI for Software EngineeringICSE 2027LLM Agents
0 likes · 11 min read

Researchers introduce FireFlash, a 300B-parameter terminal and coding agent trained via full-parameter RL from GLM-5.3-Flash on 64 B300 GPUs in under five days, achieving Terminal-Bench 4.0 Pass@1 of 46.2%—matching Fable 5 and GPT-6 Sol—while reducing per-task cost by over 30x through optimized infrastructure, task filtering, reward shaping, harness design, and a corrected GRPO objective.
GRPOLong-Horizon AgentsOpen Recipe
0 likes · 21 min read

Researchers from Fudan University, CUHK MMLab, and Shanghai Jiao Tong University introduce ReTerm, a recursive self-improvement framework where code agents learn to diagnose and recover from their own terminal failures, turning self-generated error trajectories into training data that boosts 9B model performance on Terminal-Bench, SWE-bench, and OpenThoughts-TBLite by over 13 percentage points.
Code AgentsGRPOReTerm
0 likes · 9 min read

This analysis compares three major LLM benchmarks—LMArena, Artificial Analysis Intelligence Index v5, and SWE-bench—revealing why they crown different winners, how a 60% shift to private test sets exposes data contamination, why open-source models now trail closed-source by 12 points, and how to choose models based on task, budget, and deployment constraints.
Artificial AnalysisLLM BenchmarksLMArena
0 likes · 12 min read

The article evaluates MemT-1.0, the first open-source memory-native LLM, testing its 36B Nano version on an A100 GPU by injecting four strict financial payment rules into its internal memory, then verifying via AST analysis that the model recalls and applies those rules in a new session with zero context, while detailing architecture, benchmarks, and limitations.
AST verificationGDN-2LLM memory
0 likes · 25 min read

HarnessPro introduces a multi-agent recursive self-improvement framework that uses Pareto optimization and AI-generated tests to simultaneously increase task pass rates and reduce execution costs across Heavy, Medium, and Lite tiers, achieving 26% cost reduction with 2% pass-rate improvement over RHO on SWE-Bench Pro.
AI agentsCost-Performance Trade-offHarnessPro
0 likes · 12 min read

Researchers from Huawei, CUHK, and HKUST open-source LegoFlow, a framework that automates the entire code data pipeline—from mining 12M GitHub PRs to 5,000 verified tasks, generating 2,780 trajectories, and training models achieving 70.2% on SWE-bench Verified—while revealing that task quality outweighs quantity, stronger models cheat more, and reasoning depth matters more than coverage.
LLM trainingLegoFlowSWE-bench
0 likes · 18 min read

ModularRSI freezes base model weights and evolves the agent's harness — system mechanisms like loop control, observation handling, tool use, context management, and task completion detection — through execution experience, achieving a 4.86-point gain on Terminal-Bench 2.0 that transfers across tasks, domains, and different base models.
Agent ArchitectureAgent Self-ImprovementFrozen Model Weights
0 likes · 18 min read

ModularRSI demonstrates that AI agents can continuously self-improve by evolving their harness—the surrounding system mechanisms like tool use, context management, and task completion detection—while keeping the base model frozen, achieving performance gains on Terminal-Bench and SWE-Bench that transfer across domains and models.
Agent Self-ImprovementCross-Domain TransferCross-Model Transfer
0 likes · 18 min read


Geek LabsSep 18, 2026 · Artificial Intelligence
Graft builds a local Markdown knowledge graph for AI coding agents, replacing embedding-based retrieval with deterministic structural analysis and optional LLM semantic layers, auto-refreshing on each query to cut token usage by 42% and improve SWE-bench scores by 12 percentage points.
AI coding agentsDeveloper ToolsGraft
0 likes · 13 min read

Major AI coding tools like Claude Code and Cursor have replaced vector databases with agentic retrieval using grep and ripgrep, achieving higher accuracy, freshness, and security; benchmarks show agentic keyword search reaches 94.5% of RAG faithfulness with zero embeddings, while Anthropic's multi-agent system beats single Opus 4 by 90.2%.
AnthropicClaude CodeCursor
0 likes · 32 min read

Lego‑RL is an open‑source reinforcement‑learning framework that trains coding agents directly on unmodified OpenHands SDK, Claude Code, and OpenCode harnesses, boosting SWE‑bench Verified scores from 64/62/57 to 70.4/68.2/66.6 while addressing faithful optimization, reliable execution, and observable training through GSPO and a sandboxed architecture.
AI TrainingCoding AgentGSPO
0 likes · 23 min read

With AI agents expanding individual productivity, the authors identify a bottleneck in collaborative bandwidth and propose a three‑layer AI organization model implemented as TencentDB Agent Memory, which structures team knowledge into four asset types, validates them through extensive session analysis, and demonstrates a rise in task completion from 60% to 80% on SWE‑bench benchmarks.
AI agentsSWE-benchcollaboration bandwidth
0 likes · 39 min read

The article breaks down the four‑layer AI Agent stack—MCP protocol, Agent Skill, Harness runtime, and Loop engineering—showing how each layer solves distinct problems, presenting benchmark data (e.g., a 25.7 pp SWE‑bench gain from Harness changes), security analyses, design trade‑offs, and a production checklist.
AI agent architectureAgent HarnessAgent Skill
0 likes · 26 min read

The article reviews OpenAI’s GPT‑5.6 launch—Sol, Terra, Luna—detailing new Max/Ultra modes, official benchmarks that rank Sol first in coding agents but slightly behind Fable 5 in overall intelligence, cost advantages, SWE‑Bench shortcomings, and the author’s own Terra‑Ultra tests that reveal speed and stability concerns.
AI benchmarkingGPT-5.6SWE-bench
0 likes · 10 min read

SpaceXAI's Grok 4.5 arrives ahead of OpenAI's GPT‑5.6 Sol, delivering terminal‑bench scores of 83.3% and multilingual code‑fixing at 78.0%, while offering a striking $2/$6 per‑million‑token price and 80 TPS inference speed, yet still lagging behind top rivals on the hardest SWE‑Bench Pro tasks.
AI coding agentGrok 4.5SWE-bench
0 likes · 6 min read

LoopCoder‑v2, a 7B LLM, gains a massive boost on SWE‑bench Verified (43.0 → 64.4) by adding just one test‑time loop, while additional loops cause performance to collapse, a finding explained through detailed probe analysis of hidden‑state convergence, attention re‑routing, and a constant “position‑mismatch tax”.
AI model efficiencyLLM loopingLoopCoder-v2
0 likes · 8 min read

The article analyses why Claude’s code‑writing ability outperforms rivals, tracing its edge to a combination of verifiable‑reward reinforcement learning, Constitutional AI safety guards, a product‑driven data flywheel, multi‑level reward shaping, and continuous human‑in‑the‑loop evaluation on benchmarks such as SWE‑bench.
AI safetyAnthropicClaude
0 likes · 34 min read

Anthropic reports that Claude now merges over 80% of its internal code, with failure rates cut by three‑fold, and outlines how planning, error‑recovery, and long‑context abilities enable a “Close the Loop” approach that developers must adopt to build future‑ready AI agents.
AI agentsAnthropicClaude
0 likes · 12 min read

The article critiques the reliance on raw SWE‑bench scores for programming agents, introduces the Claw‑SWE‑Bench benchmark and a dedicated adapter that isolates harness effects, and presents extensive experiments showing how model choice, harness design, and cost impact real-world coding performance across multiple languages.
HarnessLLM AgentsPass@1
0 likes · 14 min read

Anthropic’s Claude Fable 5, released on June 9, is the first publicly available Mythos‑level model that outperforms competitors across code, reasoning, and visual benchmarks, demonstrates autonomous long‑run operation, powers real‑world cases like Stripe’s massive code migration, and introduces a controversial safety‑degradation system.
AI benchmarksAI collaborationClaude Fable 5
0 likes · 11 min read

Claude Fable 5 has launched with token pricing twice that of Opus 4.8, but delivers dramatically higher benchmark scores—80.3% on SWE‑bench Pro, 95.0% on SWE‑bench Verified—and real‑world speedups such as completing a 50 M‑line Ruby migration in a single day.
AI benchmarkClaudeFable 5
0 likes · 4 min read

The article benchmarks Anthropic's Claude Fable 5 and Mythos 5, revealing superior performance on long, complex coding and AI tasks, detailed real‑world reproductions of a Shopify site and a DDIM paper, high safety‑guardrail trigger rates, and a total testing cost of about $108.
AI benchmarkingClaudeDDIM replication
0 likes · 13 min read

Anthropic released Claude Fable 5 to the public while keeping the more capable Claude Mythos 5 locked behind safety guardrails, and benchmark results show Fable 5 outperforms competing models in programming, vision, and complex tasks, though its scores are deliberately lowered in sensitive domains.
AI benchmarksAI safetyAnthropic
0 likes · 11 min read

The author compares Claude 4.8 (Opus) and Codex 5.5 across SWE‑bench Pro (69.2% vs 58.6%) and Terminal‑Bench (78.2% vs 74.6%), highlighting Claude’s larger 1 M‑token context, higher accuracy on complex multi‑file tasks, and higher cost, while Codex offers faster, cheaper terminal‑focused performance, recommending each for specific scenarios.
AI code generationClaude 4.8Codex 5.5
0 likes · 4 min read

MiniMax unveiled its open‑source M3 model, delivering 1 million‑token context, 59 % SWE‑Bench Pro accuracy that outperforms GPT‑5.5 and Gemini 3.1 Pro, native multimodal desktop interaction, and a 9‑15× speed boost via MiniMax Sparse Attention, with pricing as low as $20 per month.
MSAMiniMax M3SWE-bench
0 likes · 11 min read

The article reviews OpenAI's Codex extension for JetBrains IDEs, detailing its easy setup, GPT‑5.4 performance on SWE‑Bench, hands‑free project cloning, automatic environment provisioning, global refactoring, test generation, and real‑world usage examples that demonstrate its near‑professional coding capabilities.
AI coding assistantCodexGPT-5.4
0 likes · 6 min read

The article compares Claude Opus, OpenAI Codex, and Zhipu's open‑source GLM‑5.1, detailing their strengths, benchmark results, pricing, and ideal use cases, and recommends routing tasks to the model that best fits the complexity and language requirements.
AI modelsClaudeCodex
0 likes · 5 min read

Claude Opus 4.8, released just 43 days after 4.7 at the same price, tops the GDPval‑AA leaderboard with 1890 Elo, beats GPT‑5.5 by 121 points, cuts steps by 15% and tokens by 35%, achieves a perfect 0% lie and lazy rate, dominates SWE‑Bench, ProgramBench and FrontierSWE, and introduces massive parallel agent workflows that can rewrite 750 k lines of production code in 11 days, while Anthropic prepares the upcoming Claude Mythos and celebrates a $965 b valuation.
AI benchmarksAgent ParallelismClaude
0 likes · 10 min read


ZhiKe AIMay 29, 2026 · Artificial Intelligence
Anthropic released Claude Opus 4.8 only 41 days after Opus 4.7, delivering unprecedented 0 % lie‑rate and 0 % lazy‑answer rate, improving code‑defect silence by four‑fold, boosting SWE‑bench Pro to 69.2 % and GDPval‑AA to 1890 Elo, while adding Dynamic Workflows, Effort Control, a richer Messages API and a fast‑mode that runs 2.5× faster for a third of the cost.
AI honestyClaude Opus 4.8GDPval-AA
0 likes · 11 min read

Stanford, Berkeley and Nvidia researchers introduce LLM-as-a-Verifier, a universal verification framework that enhances agent performance, safety and stability on long‑horizon tasks, and outperforms Claude Mythos and GPT‑5.5 on the Terminal‑Bench and SWE‑Bench benchmarks.
AI agentsAgent VerificationLLM-as-a-Verifier
0 likes · 7 min read

The Qwopus 3.6‑27B‑v2 model reconstructs full step‑by‑step reasoning from compressed Claude outputs using a Trace‑Inverter, creates two high‑quality SFT datasets, and achieves 35.9% token savings, a 2.57‑point accuracy gain on MMLU‑Pro, 75.25% success on SWE‑bench, while running on a single consumer‑grade RTX 5090.
GGUFMMLUQwen
0 likes · 11 min read

Microsoft is forcing thousands of engineers to abandon the higher‑scoring Claude Code AI coding assistant in favor of GitHub Copilot CLI by June 30, citing cost savings, internal security requirements, and a six‑week migration window despite Claude Code’s better benchmark performance and larger context window.
AI coding assistantsClaude CodeGitHub Copilot
0 likes · 8 min read


SuanNiMay 16, 2026 · Artificial Intelligence
Microsoft’s research shows that a 4‑billion‑parameter small model, Terminus‑4B, can act as an execution sub‑agent for terminal tasks, trimming token consumption by about 30% while preserving performance on demanding SWE‑Bench benchmarks, demonstrating a practical alternative to costly large models.
AI programmingRL trainingSWE-bench
0 likes · 7 min read


21CTOMay 15, 2026 · Artificial Intelligence
Microsoft will revoke internal Claude Code licenses and force engineers to switch to GitHub Copilot CLI, citing cost savings and ecosystem control, even though benchmark data shows Claude Code outperforms Copilot on SWE‑bench, multi‑file refactoring, and large‑context tasks.
AI coding assistantsAnthropicClaude Code
0 likes · 6 min read

The paper identifies why high‑performing code models falter in real software engineering tasks, introduces the SEAlign alignment framework that targets key decision points in agent trajectories, and demonstrates substantial gains on SWE‑Bench, HumanEvalFix, and user‑centric evaluations.
AICode ModelsSEAlign
0 likes · 12 min read

This article provides a comprehensive walkthrough of Claude Code, covering its core capabilities, cross‑platform installation, login methods, basic commands, advanced features like hooks and MCP, a detailed comparison with GitHub Copilot, best‑practice prompts, and FAQs, enabling developers to boost productivity with an AI‑driven terminal assistant.
AI coding assistantClaude CodeGitHub Copilot
0 likes · 16 min read

The author compares Claude Code, Cursor, and GPT‑based agents, discovers the open‑source Kimi K2.6 model, installs it in minutes, runs three realistic coding tasks, and shows that a mixed‑agent workflow can cut token costs by up to 85% while maintaining comparable quality.
AI coding agentsClaude CodeCost Optimization
0 likes · 13 min read

This article walks through ten mainstream open‑source large‑model benchmarks—SWE‑bench Verified and Pro, MMLU‑Pro, GPQA Diamond, HLE, AIME, HMMT, olmOCR‑bench, Terminal‑Bench 2.0, and EvasionBench—explaining their data, evaluation metrics, current leading models, and the capability dimensions they reveal.
AI-evaluationLLM BenchmarksMMLU-Pro
0 likes · 20 min read


JavaGuideApr 27, 2026 · Artificial Intelligence
DeepSeek V4’s pricing fell 75% overnight, making the V4‑Pro and V4‑Flash models dramatically cheaper than competing AI services; the article details the new rates, compares them with other providers, shows two Claude Code case studies consuming nearly 4 million tokens, and explains how domestic Ascend 950 hardware enables the discount.
AI pricingAscend 950Claude Code
0 likes · 13 min read

Stanford, Berkeley and Nvidia introduce LLM‑as‑a‑Verifier, a verification framework that scales verification compute, uses fine‑grained score tokens, repeated checks and criteria decomposition to boost agent performance, eliminate scoring ties and achieve SOTA results on Terminal‑Bench, surpassing Claude Mythos and GPT‑5.5 while improving safety in long‑horizon tasks.
Agent VerificationLLMLLM-as-a-Verifier
0 likes · 8 min read

OpenAI’s April 23, 2026 launch of GPT-5.5 delivers record‑high scores on SWE‑Bench Pro (58.6%) and Terminal‑Bench 2.0 (82.7%), adds persistent multi‑file context, dynamic reasoning time, and token efficiency, while real‑world case studies show substantial productivity gains across engineering teams.
AI EngineeringCodexGPT-5.5
0 likes · 13 min read

JetBrains IDEs from version 2025.3 embed the Codex AI assistant powered by GPT‑5.4, offering faster, context‑aware code generation, project analysis, environment setup, and refactoring, with real‑world demos showing how it can download projects, configure tools, and even build a full mini‑program with minimal manual coding.
AI assistantCodexGPT-5.4
0 likes · 7 min read


ZhiKe AIApr 21, 2026 · Artificial Intelligence
Kimi K2.6, an open‑source Chinese LLM, outperforms GPT‑5.4 and Claude Opus 4.6 on SWE‑Bench Pro code tests, delivers 13‑hour uninterrupted coding, runs 300 parallel agents, and costs only one‑twentieth of comparable closed‑source models, while offering a trillion‑parameter MoE architecture and Apache 2.0 licensing.
AI model benchmarksAgent ParallelismApache 2.0
0 likes · 9 min read

Anthropic’s new Advisor Strategy lets the low‑cost Sonnet (or Haiku) model handle full agent tasks while invoking the powerful Opus model only for difficult decision points, delivering a 2.7‑point score boost on SWE‑bench with roughly 12% lower cost, and can be added with a single API call.
AI agentsAnthropicClaude
0 likes · 8 min read

The AI daily briefing highlights GLM‑5.1’s breakthrough 8‑hour continuous reasoning, its top performance on SWE‑bench and a 10% price hike, while Tencent’s QBotClaw introduces the first domestically free‑configurable large‑model API browser, signaling a shift toward open AI ecosystems in China.
AI ecosystemAI pricingGLM-5.1
0 likes · 6 min read

LongCat-Next, a 3‑billion‑parameter multimodal model released by Meituan, adopts a pure discrete token‑based architecture (DiNA) and next‑token prediction, outperforming same‑size rivals on OmniDocBench‑EN, CharXivRQ, and matching QwenVL on visual tasks, while avoiding catastrophic forgetting and achieving a SWE‑Bench score of 43.0, as demonstrated through extensive benchmarks, receipt extraction, OCR, audio dialect reasoning, and image generation experiments.
DiNALongCat-NextOmniDocBench
0 likes · 10 min read

LongCat‑Next, Meituan’s new 3‑billion‑parameter foundation model, adopts a pure‑discrete DiNA architecture with next‑token prediction, converting vision, audio and text into unified tokens; it surpasses same‑size multimodal models on OmniDocBench‑EN, CharXivRQ and SWE‑Bench, avoids catastrophic forgetting, and introduces dNaViT, RVQ compression and a dual‑path detokenizer for high‑fidelity generation.
DiNALongCat-NextSWE-bench
0 likes · 10 min read

GLM-5.1 is now available to all Coding Plan subscribers, including the $10/month Lite tier, scoring 45.3 on SWE‑bench—just 5.4% below Claude Opus 4.6’s 47.9—while offering 20+ tool integrations and a manual switch from the default GLM‑4.7 model.
AI coding modelClaude OpusGLM-5.1
0 likes · 7 min read

Although SWE‑bench scores for top coding models now differ by only a tenth of a point, Cursor’s newly released CursorBench reveals dramatic ranking changes, highlights three fundamental flaws in public benchmarks, and introduces token‑efficiency as a crucial evaluation dimension.
AI codingCursorBenchSWE-bench
0 likes · 8 min read

Google’s Gemini 3.1 Pro dramatically improves core reasoning scores (77.1% on ARC‑AGI‑2, 80.6% on Swe‑bench) and can generate interactive SVG, complex data‑driven visualizations, and creative‑coding layouts, prompting a reassessment of which front‑end tasks AI can replace and which still require architectural expertise.
AI code generationFrontend DevelopmentGemini 3.1 Pro
0 likes · 6 min read

A leaked SWE‑Bench score of 83.7% for DeepSeek V4 sparked claims it outperforms Claude Opus 4.5 and GPT‑5.2, but the data was later debunked as fabricated while official hints confirm a 1‑million‑token context model and a mid‑February 2026 release.
AI benchmarkingAI industryDeepSeek
0 likes · 7 min read

The MiniMax M2.5 model, released open‑source on Feb 13, achieves an 80.2% SWE‑Bench Verified score that surpasses GPT‑5.2, Claude Opus 4.6 and Google Gemini 3 Pro, runs 37% faster than its predecessor, costs only $1 per hour, and demonstrates SOTA agent abilities in browsing and tool use, marking a major leap for Chinese large‑language models.
AI codingAgentM2.5
0 likes · 7 min read

MiniMax’s new open‑source M2.5 model, built on a 230 billion‑parameter mixture‑of‑experts architecture that activates only 10 billion parameters per inference, delivers performance comparable to Claude Opus 4.6 across benchmarks, while costing roughly one‑tenth as much, and is already handling a large share of the company’s internal tasks.
AI agentsClaude OpusMiniMax M2.5
0 likes · 6 min read

The article objectively compares OpenAI's GPT‑5.3‑Codex and Anthropic's Claude Opus 4.6 across Terminal‑Bench 2.0 and SWE‑Bench, revealing a 15% terminal‑coding edge for Codex, modest gains in pure code generation, and a strategic split between specialist and generalist AI approaches.
AI Model ComparisonAgentic WorkflowClaude Opus 4.6
0 likes · 9 min read

A leaked Google Vertex AI log reveals a new model ID claude‑sonnet‑5@20260203, hinting at a Feb 3 2026 release of Claude Sonnet 5 (code‑named “Fennec”) that reportedly scores over 82 % on SWE‑Bench, outperforms Opus 4.5, keeps the same pricing, and introduces a “Dev Team” mode with parallel sub‑agents for coding tasks.
AI ModelAgentic WorkflowClaude Sonnet 5
0 likes · 5 min read

JoyCode Agent leverages a patch‑test co‑generation and iterative validation framework to achieve a 74.6% Pass@1 score on the SWE‑bench Verified benchmark, reducing resource consumption by 30‑50% and introducing a closed‑loop multi‑agent pipeline that integrates testing, patch generation, trajectory compression, similarity retrieval, and decision arbitration.
AILLMSWE-bench
0 likes · 41 min read

This technical report details JoyCode Agent’s end‑to‑end pipeline that couples patch generation with fail‑to‑pass and pass‑to‑pass test creation, uses trajectory compression, CSR similarity retrieval, and multi‑agent iterative retries to achieve a 74.6% Pass@1 score on the SWE‑bench Verified benchmark while cutting compute costs by 30‑50%.
AI code repairPatch GenerationSWE-bench
0 likes · 38 min read

This article evaluates MiniMax M2.1’s new coding capabilities across multiple benchmarks, including SWE‑bench, Java satellite‑control projects, full‑stack attack visualizations, and a one‑click mobile‑OS simulation, comparing its performance to Claude Sonnet 4.5 and Opus 4.5.
AI coding assistantM2.1MiniMax
0 likes · 8 min read

OpenAI’s GPT‑5.2 launch introduces three specialized modes, achieves a record 55.6% score on SWE‑Bench Pro, demonstrates strong front‑end generation, adds a /compact API for long‑context efficiency, offers tiered pricing with cache discounts, and improves safety for younger users.
AI benchmarkingAI safetyFront-end generation
0 likes · 6 min read

This guide shows how to replace Claude Code's default model with the Chinese Doubao‑Seed‑Code programming model by creating a wrapper script, configuring environment variables, and running a few shell commands, then demonstrates the setup by generating a functional Space Invaders web game.
AI coding modelClaude CodeDoubao-Seed-Code
0 likes · 8 min read


JD TechNov 5, 2025 · Artificial Intelligence
JoyCode Agent, an AI‑driven multi‑agent system, secured a 74.6% pass@1 rate on the SWE‑bench Verified benchmark, placing it in the global Top‑3 while cutting computational resource usage by 30‑50% through a novel patch‑test co‑generation and iterative verification pipeline.
AIAgent ArchitecturePatch Generation
0 likes · 34 min read

JoyCode Agent reaches a 74.6% pass rate on the authoritative SWE‑bench Verified benchmark, ranking in the global top‑3, and is now open‑source, showcasing a high‑efficiency, test‑driven, iterative approach to automated code repair that dramatically reduces token consumption while improving success rates.
Automated Code RepairSWE-benchartificial intelligence
0 likes · 44 min read

JoyCode Agent leverages a multi‑agent, patch‑and‑test co‑generation framework with iterative validation, failure attribution, and experience‑driven retries to achieve a 74.6% Pass@1 rate on the SWE‑Bench Verified benchmark, dramatically reducing computational resources while delivering high‑quality code patches.
AI code generationSWE-benchautomated patching
0 likes · 34 min read

The author compares Claude Code with generic Claude calls in IDEs, details Sonnet 4.5’s benchmark superiority, reveals how the model excels on fresh projects but falters on legacy code riddled with technical debt, and shares practical methods to maximize its value.
AI codingClaude CodePrompt engineering
0 likes · 10 min read

This article examines the design philosophy behind the “Bash Only” category of the SWE‑bench benchmark, explaining how its minimal‑agent approach isolates LLM reasoning by restricting interactions to a plain Bash shell, making it a rigorous, reproducible test of true software‑engineering intelligence.
AI-evaluationBash OnlyLLM
0 likes · 7 min read

Shunyu Yao, a Princeton‑trained AI researcher who rose through Tsinghua’s elite Yao class and OpenAI, is known for pioneering works like Tree of Thoughts, SWE‑bench, and ReAct, and now focuses on building general‑purpose agents and exploring the “second half” of AI development.
AI researchAgentReact
0 likes · 12 min read

ByteDance’s TRAE Agent achieves a record 75.20% success on the SWE‑bench benchmark by bridging the “complexity gap” between function‑level and repository‑level tasks through a three‑stage pipeline—patch generation, pruning, and selection—augmented with ensemble reasoning, multi‑model integration, and a novel test‑time scaling mechanism.
AI agentsSWE-benchTest-Time Scaling
0 likes · 14 min read

The mini‑SWE‑agent, a lightweight open‑source software‑engineering AI built by the original SWE‑bench team, achieves about 65% bug‑fix success on the SWE‑bench benchmark using roughly 100 lines of Python, thanks to its minimal dependencies, shell‑based execution, linear history, and support for various container environments, offering a fast, extensible alternative to the full‑featured SWE‑agent.
AI AgentLLMSWE-bench
0 likes · 8 min read

Kimi-Dev-72B, an open-source 72-billion-parameter code model from Moonshot AI, achieved a record 60.4% score on the SWE-bench Verified benchmark, surpassing larger models, and incorporates BugFixer/TestWriter dual roles, extensive mid-stage training on billions of GitHub data, and reinforcement-learning-driven self-play, with code available on Hugging Face and GitHub.
AIReinforcement LearningSWE-bench
0 likes · 7 min read

This article evaluates Anthropic's Claude 4 (especially Claude‑4‑Sonnet) within the Cursor IDE, presenting benchmark scores on SWE‑bench, detailed prompts for UI, frontend, architecture, and backend generation, visual results, and a balanced list of strengths and remaining issues.
AI codingClaude 4Cursor IDE
0 likes · 19 min read

The article evaluates Augment Code, an AI‑powered coding assistant with 200K token context, persistent memory, multimodal input, and top SWE‑bench scores, walks through its installation, explores its use on vllm and PagedAttention, demonstrates adding a new model and auto‑generating a WeChat mini‑program, and compares its capabilities and speed to Cursor.
AI coding assistantAugment CodeCursor
0 likes · 8 min read

The article reviews 2024 AI programming progress, covering the rise of AI code editors like Cursor, the debut of the AI programmer Devin, rapid improvements in SWE‑bench success rates, enhancements in model architecture, multimodal agents, tool‑integration frameworks, adoption statistics in China and abroad, and future directions for collaborative AI‑driven software development.
AI agentsAI programmingSWE-bench
0 likes · 10 min read

This article examines the practical challenges of using large language models in software development, including handling long contexts, cross‑file editing, bug‑fixing evaluation methods, and presents benchmark results from SWE‑Bench and its Lite subset to assess model capabilities.
Cross-File EditingLLMSWE-bench
0 likes · 7 min read
