§ Coding

Introduces Direct Pass@k optimization for pruning candidate reasoning solutions during test-time inference, balancing solution quality and diversity.

Framework for coding agents to dynamically select skill granularity (primitive/macro/hybrid) at runtime, enabling fallback recovery strategies.

Multi-token prediction and alternative tokenization enable language models to learn shortest-path graph search at scale, overcoming prior scaling failures.

Training-free method for coordinated parallel token commitment in few-step diffusion language models balancing reliability and joint consistency.

ProductWebBench: benchmark for frontend coding agents measuring both task completion and website continuity across 400 real-repo tasks.

Multi-directional Perturbation-Rewrite Search framework uses LLMs to automate heuristic design for NP-hard combinatorial optimization with factorized transformations.

CAKE system automates AscendC kernel generation for Ascend NPUs via adaptive multi-DSL routing and agent-guided evolution, addressing data scarcity gap.

SWE-DIVE benchmarks coding agents via multi-turn interaction with dynamic user simulation beyond single-prompt SWE-bench evaluation.

Code Comprehensor uses dependency callgraphs and hierarchical documentation for large repository code understanding by LLMs.

EPIC reformulates PPO-style RL for LLM policy training via exact-KL prefix importance-ratio clipping, addressing policy-context mismatch in long-response generation.

HiddenLens, a six-layer BERT-LLM pipeline for detecting implicit assumptions in software requirements to reduce costly late-stage resolution.

FrontierCS 2 benchmark of 48 open-ended CS R&D tasks measures LLM agent contributions in working research environments.

Cloudflare deploys OpenAI Daybreak for vulnerability discovery and Codex for code generation to accelerate development and security workflows.

Adaptive multi-token prediction uses online control to optimize LLM decoding horizon based on verification feedback, balancing latency and compute.

RTLRepairDataset and RTLRepairBench identify measurement confounds in HDL repair benchmarks using simulation and formal equivalence.

LLM-driven evolutionary program search optimizes semantic index design for generative retrieval, automating end-to-end SID component tuning.

Binary Discrete Diffusion enables partial-uncertainty representation in discrete generative models for reasoning and constrained coding tasks.

LLMs hallucinate in cross-language code translation via semantic conflicts, target-language violations, and inconsistent dependencies; existing mitigations insufficient.

VISH framework selects compact hotword subsets for speech-LLM contextual ASR via frame-level CTC posterior verification; training-free and scalable via shared-prefix Viterbi search.

Token-level framework detects and localizes AI-generated code within files using multi-granularity analysis, outperforming file-level compression methods.

UvA-MT uses LLM sampling, refinement, and pairwise voting for WMT26 machine translation across 23 language pairs without task-specific training.

ProVe reverses speculative decoding: large model proposes token blocks, smaller model verifies in parallel, improving LLM inference speed.

Agentic workflow guides LLM coding agents to implement 6G physical-layer algorithms on embedded baseband silicon with vector assembly optimization.

Simon Willison reflects on how coding agents may disrupt traditional programmer employment by automating logic-puzzle work that defined the profession's appeal and compensation.

TokenGate enables learnable token pruning in LLMs for faster, memory-efficient long-context prefill using a differentiable top-K approximation.

OpenAI improves ChatGPT performance on low-resource languages via new evals for SQLite, COBOL, and software engineering tasks.

PatternBench isolates LLM code optimization from compiler optimizations; measures compiler-transcendent semantic reasoning in code generation.

Utility-Verified Abstraction teaches code agents to query cloud models while masking private code via role/type abstraction instead of identifier removal.

System enables LoRA fine-tuning on phone NPUs by computing input gradients using existing inference-held weights, avoiding device upload.

Online error localization framework for streaming ASR enables targeted speaker clarification by identifying low-confidence spans before propagation.

Empirical study of 1,500 LLM refactoring attempts across 15 Java repos shows multi-agent approaches improve dependency-aware code generation over single-prompt baselines.

Anthropic's Claude shifts code review from PR gate to continuous in-editor assistance, catching routine errors before human review.

FuTransformer CUDA kernel for Hopper GPUs uses FP8 prefill, INT4 decode, and direct KV-cache placement for Llama inference.

LLMs trained on experimental feedback learn to propose model improvements without expensive RL, enabling recursive self-improvement for AI research automation.

MGE-PO improves RLHF credit assignment for math reasoning by pairing correctness with generation entropy to reward low-entropy successes and penalize high-entropy failures.

Simon Willison implements SPARQL 1.1 query support as a SQLite extension in Rust, demonstrating Claude Code's dynamic workflow capabilities.

Benefit Street Partners deployed Codex for code review and financial document analysis to accelerate loan valuation comparisons.

3DCodeVerse dataset and framework enable coding agents to generate, verify, and scale 3D executable code for spatial reasoning tasks.

LeanGraphSearch improves formal theorem proving by retrieving relevant lemmas via graph-based multi-hop dependency analysis in Lean 4.

Diffusion language models enable few-shot infilling for multi-region code/text edits by learning from demonstrations without causal constraints.

Scaffolding in coding agents improves performance on ARC-AGI-3 and Terminal-Bench even with stronger models; environment-agnostic design persists.

Poison Valleys diagnostic traces path-dependent failures in LLM code repair agents via hypothesis/implementation tracking on 1,844 requests across PersistQueue and Dijkstra tasks.

OPRD uses multi-scope reminder supervision and outcome routing for long-horizon coding agents; combines autoregressive self-imitation with reminder-free student rollouts.

TITAN framework uses causal intervention and sparse feature decomposition to identify and optimize translation circuits in multilingual LLMs for low-resource language pairs.

PrismaCoder open-source multimodal model for code-to-visual generation using decomposed rubric synthesis to improve data quality and training pipelines.

LAGO meta-agent framework treats specialized agents as reusable options; resolves 58.6% of SWE-bench tasks with frozen Qwen3.5-35B.

ReproBAIT benchmark evaluates coding agents' ability to reconstruct biological AI tools from limited descriptions without internet access.

Semi-supervised graph learning for code-to-architecture mapping using dependency graphs and self-training to recover modular decomposition from source files.

Wenyan Thinking rewrites chain-of-thought reasoning in Classical Chinese to reduce token budget for Qwen3-4B-Thinking-2507 inference.

Block deploys OpenAI models for threat investigation and patch generation, reducing security response cycles to minutes.