Skip to main content
Search

# Code Generation

29 related articles

Subscribe to updates
  1. Stitch skills ship as a plugin marketplace: six coding agents can install design and build skills on demand

    The google-labs-code repository has released stitch-skills, a collection of agent skills and plugins for Google Stitch that follows the Agent Skills open standard. It organizes design, build and utility plugins as an addable marketplace for Codex, Antigravity, Gemini CLI, Claude Code and Cursor, while OpenCode requires manually copying skill folders and configuring the Stitch MCP server. The skills depend on a configured and running Stitch MCP server, and selective installs must include the dependencies between skills.

  2. arXiv

    Micius-Flash opens a 27B engineering model: 64.2% average on hardware coding, 15 points over a same-size baseline

    Micius-Flash is a 27-billion-parameter engineering model with native multimodal input and a 256K-token context window, specialized and evaluated primarily on semiconductor engineering, and released as an open-source model. A knowledge-tree-driven data flywheel organizes semiconductor knowledge, synthesizes data and validates quality, and multi-teacher on-policy distillation merges four expert capabilities, knowledge and reasoning, hardware coding, multimodal understanding and agentic tool use, into one student model. It averages 64.2% on 12 hardware coding benchmarks, 15.0 points above a same-size general-purpose baseline, and scores 62.4% on a semiconductor-specific agent benchmark, 17.2 points above that baseline.

  3. arXiv

    Reasoning models learn where to stop: LoCo cuts math completions by 41.7%–45.4% by training only the continue-or-conclude choice

    Large reasoning models often keep deliberating after enough information is available, raising decoding latency and memory cost. LoCo moves the compression decision to individual reasoning states: it finds candidate overthinking states in the model's own traces, freezes that prefix, regenerates shorter continuations whose answers verify as correct, and learns a continue-or-conclude preference with DPO. Across five math benchmarks, completion length drops 41.7%–45.4% while accuracy holds or improves; trained on math only, the same model cuts length by 23.4%–36.3% on three code benchmarks.

  4. Kimi K3.1 Pro ships as open weights: its coding agent scores 65.7% on Terminal-Bench 4.0, just behind Claude Opus 5.5

    Moonshot AI released Kimi K3.1 Pro, the first model in its K3.1 series and its most capable 3T-class open-weight model to date, with a 1-million-token context window, alongside a Highspeed variant delivering 2-3x the speed. It is a post-trained version of Kimi K3 built with substantially larger post-training compute, supported by optimizations across the RL framework, long-context inference, sandbox infrastructure and rollout scheduling. On Terminal-Bench 4.0, which tests coding agents on realistic terminal tasks, K3.1 Pro scores 65.7%, ahead of GPT-6 Astra at 58.2% and narrowly behind Claude Opus 5.5 at 66.4%.

  5. Memetic Computing

    A correct answer is not a correct derivation: GRPO gets a noise-aware group weight that keeps learning at high reward noise

    GRPO, a common reinforcement-learning recipe for reasoning models, judges a rollout only by whether its final answer is right, so a correct answer can hide unreliable reasoning, and a single corrupted reward can strongly distort group-relative advantages when a response group is highly imbalanced. S-GRPO minimizes the expected squared error between observed and latent advantages under a reward-noise model to obtain a closed-form, noise-aware group weight. Across several mathematical reasoning benchmarks and model scales it stabilizes training and keeps making progress under controlled reward noise where standard GRPO shows little meaningful learning.

  6. A quadruped stops reciting command lists: one plain-language goal becomes checkable code, and open models reach about 82% success

    Instead of drawing on a fixed command library, a quadruped exploring unknown space hands the user's high-level intention to a large language model that writes an executable program restricted to a robot function library, while a vision-language model supplies open-ended semantic labels and geometric and semantic thresholds trigger replanning. In Gazebo simulation, Qwen3.5 and Kimi-K2.5 reach about 82% overall success and GPT-5-nano about 46%; finite tasks hold up well, while open-ended tasks vary more.

  7. arXiv

    Long-horizon RL training stops collapsing mid-run: state-shift correction beats GRPO by 48% on math and 12% on code

    In long-horizon autoregressive reinforcement learning, local policy mismatch compounds along the trajectory into state distribution shift, destabilizing training and eventually collapsing it as the horizon grows. Trajectory-Aware Policy Optimization derives a state-distribution correction from the off-policy policy gradient and writes it into the objective, estimates it with a geometric mean to cut variance, and tightens the clipping boundary with the length of the preceding trajectory. Across eight competition-level math and code benchmarks it improves over GRPO by up to 48% on math and 12% on code on average, with training scores rising steadily while KL divergence stays controlled.

  8. arXiv

    Ask before you code: 341 repository-level tasks show clarification lifts requirement alignment by 133%

    For repository-level software development, researchers built ReqClarifyBench, pairing 341 tasks with vague user intents, human-validated requirement documents, and code evaluation artifacts, and ran multi-turn clarification with simulated users carrying three real-world personas. Across six LLMs, asking effectively did not mean writing better documents: the strongest questioner scored lower on requirement alignment, yet adding interactive clarification raised average alignment from 0.212 to 0.494, a 133.1% relative gain, with code-generation scores rising too.

  9. arXiv

    Cross-tokenizer on-policy distillation: strict alignment already covers 85.57–96.98% of student tokens, a student-selected top-16 subset retains at least 96% of the gain, while adding span supervision on mismatch groups lowers accuracy

    Across three heterogeneous teacher–student pairs on mathematical reasoning and code generation, this study finds that strict alignment already covers 85.57–96.98% of student tokens despite static vocabulary Jaccard overlap of only 39.49–64.87%, that the shared vocabulary retains nearly all predictive probability mass at strict positions, that restricting reverse KL to a student-selected top-16 subset of the shared vocabulary preserves at least 96% of the full-shared-vocabulary improvement, and that adding span log-probability MSE on mismatch groups achieves complete supervision coverage yet lowers accuracy in all 18 positive-weight settings.

  10. arXiv

    MAPE-K dual-AI framework lets a quadruped robot turn natural-language intentions into executable code, robust on finite tasks while open-ended exhaustive search stays cognitively limited

    The work introduces a MAPE-K-inspired dual-AI architecture in which an LLM translates high-level natural-language intentions into executable Python code constrained to a robotic function library, while a VLM provides open-ended semantic grounding through a distillation process, with geometric and semantic thresholds triggering replanning; benchmarked on quadruped exploration tasks in Gazebo across three frontier models, finite tasks show high success rates while open-ended exhaustive tasks degrade, with failures concentrated in semantic grounding, completeness checking, and termination control rather than low-level action execution.