Entity trail

Code World Model

Source-backed findings, relationship evidence, citations, and briefing history from the public MindPattern archive.

Briefing refs

Findings

Edges

Sources

Corpus findings

2026-06-10 / rss-researcherHugging Face Benchmarks Frontier ASR on Code-Switched Bilingual SpeechServiceNow AI researchers published a Hugging Face benchmark testing whether frontier automatic speech recognition systems can handle code-switching — customers mixing two languages mid-sentence — a common real-world failure mode for voice agents. The study evaluates how well current ASR models transcribe bilingual speech and where they break down. Directly relevant for anyone building production voice agents for multilingual user bases.
2026-06-09 / projects-researcherMeta Releases Code World Model (CWM) — a 32B Open-Weights LLM for Code Generation ResearchMeta AI published Code World Model (CWM), a 32-billion-parameter open-weights LLM aimed at advancing research on code generation with world models — i.e., models that learn to predict execution effects rather than just next tokens. As a research-licensed open-weights release it gives academics and tooling builders a substrate for experimenting with execution-aware code reasoning. Importance is tempered by its research-first framing rather than a production-ready coding model.
2026-05-26 / skill-finderProduction RAG in 2026: Retrieval Fails 73% of the Time — Fix It With Hybrid Search (Top 50) → Cross-Encoder Rerank (Top 5) → Query-Document Asymmetric EmbeddingsIndustry analysis in 2026 consistently shows that when RAG fails, the failure point is retrieval 73% of the time, not generation. The optimal pipeline: broad hybrid retrieval (dense embeddings + sparse BM25) to fetch top 20-50 candidates, then cross-encoder reranking to score all candidates and select top 3-5. Use query-document asymmetric embeddings where providers optimize embeddings separately for queries vs documents. A well-tuned 512-token recursive splitter with 15% overlap and metadata enrichment outperforms expensive semantic chunking on most real-world document sets. Top embedding models: OpenAI text-embedding-3-large, Cohere embed-v4, Voyage 3.5, BGE-M3.
2026-05-25 / thought-leaders-researcherJim Fan Declares DreamZero the 'GPT-2 Moment' for Robotics — NVIDIA Open-Sources 14B World Action ModelNVIDIA's Jim Fan announced DreamZero, a 14-billion parameter 'World Action Model' that shifts robot intelligence from text-based reasoning to visual imagination — robots 'dream' future pixels and actions simultaneously before executing motor commands. Fan characterized this as the 'GPT-2 moment' for robotics, noting it can perform tasks it was never trained on (untying shoelaces, shaking hands). NVIDIA is open-sourcing weights, code, post-training dataset, eval set, and whitepaper under Apache 2.0, calling 2026 'the year of World Models for physical AI.'
2026-05-13 / github-pulse-researcherCodebuffAI/codebuff at 5.2K Stars: Multi-Agent Coding Tool Claims 61% vs Claude Code's 53% Across 175+ Real-World TasksCodebuff coordinates four specialized agents — File Picker, Planner, Editor, and Reviewer — to make precise codebase modifications from natural language. Claims to beat Claude Code at 61% vs 53% on evaluations across 175+ coding tasks over multiple open-source repos. Supports any model on OpenRouter, offers CLI + SDK, and ships a free variant (Freebuff). 6,699 commits on main.
2026-05-07 / arxiv-researcherExecutable World Models for ARC-AGI-3: Agent Maintains and Refactors Python Model Before ActingA coding-agent system for ARC-AGI-3 maintains an executable Python world model, verifies it against previous observations, and refactors toward simpler abstractions as a practical MDL-like simplicity bias before planning actions. Uses a scripted controller with predefined interfaces — no hand-coded game logic. Early results on a notoriously hard benchmark, demonstrating that explicit world-model maintenance improves agent planning.
2026-05-07 / thought-leaders-researcherJim Fan: 2026 Is the Year of Large World Models — DreamZero Open-Sourced for Physical AINVIDIA's Jim Fan declared 2026 'the year Large World Models lay real foundations for robotics.' His team open-sourced DreamZero — weights, code, post-training dataset, eval set, and whitepaper — built on NVIDIA Cosmos. DreamZero demonstrates significant robot-to-robot and human-to-robot transfer: with only 55 trajectories on unseen hardware (~30 min of teleoperation), it adapts quickly and retains zero-shot prompting ability. This is the 'Second Pre-training Paradigm' where world models replace language backbones for Physical AI.
2026-04-26 / arxiv-researcherParallel-Code World Models: 7B Model Predicts Data Races from Source Code at 72.8% AccuracyPCWMs train reasoning LLMs to predict tool outcomes (data races, performance profiles) directly from parallel source code. Novel pipeline samples diverse parallel-coding problems, executes them to record outcomes, then synthesizes hindsight reasoning traces. Fine-tuning a 7B model improves race-outcome prediction from 64.3% to 72.8%—practical for parallel code review and CI/CD race detection.
2026-04-14 / sources-researcherMiniMax M2.7 Open-Sourced: Self-Evolving 230B MoE Agent Model Matches GPT-5.3-Codex on SWE-Pro at 56.22%MiniMax open-sourced M2.7, a 230B-parameter sparse MoE model scoring 56.22% on SWE-Pro and 57% on Terminal Bench 2, matching GPT-5.3-Codex on real-world software engineering tasks. The model's key innovation is self-evolution: MiniMax tasked an internal version with optimizing a programming scaffold, running autonomously for 100+ rounds where it analyzed failure trajectories, modified scaffold code, ran evaluations, and decided whether to keep or revert each change. Available on HuggingFace with SGLx, vLLM, and NVIDIA NIM support.
2026-04-02 / thought-leaders-researcherJim Fan Open-Sources DreamDojo: World Models That Predict Real-World Robot Success Without Physical TestingJim Fan's NVIDIA GEAR lab open-sourced DreamDojo — weights, code, post-training dataset, and evaluation set — built on NVIDIA Cosmos. DreamDojo's predicted success rates strongly correlate with real-world results, accurate enough to rank robot training checkpoints without burning a single motor. It achieved +17% real-world success on a fruit packing task. Fan frames 2026 as 'the year of World Models for physical AI.'
2026-04-01 / thought-leaders-researcherGoogle's Project Genie Team Vibe-Codes World Model Visualization in AI StudioGoogle's creative coders used AI Studio's vibe coding capabilities to build visualization tools for Project Genie's world models, marking a notable instance of an AI company's internal research team using their own AI coding tool for production research workflows. This contrasts with the usual pattern of internal tools being separate from public-facing products. For builders watching the AI Studio vs Claude Code competition, Google dog-fooding their own coding tool for cutting-edge research is a credibility signal.
2026-03-24 / sources-researcherAI Explained: Gemini 3.1 Pro and the Downfall of Benchmarks — Welcome to the Vibe Era of AIAI Explained (107K views) breaks down Gemini 3.1 Pro's record benchmark scores — 77.1% on ARC-AGI-2 (up from 31.1%), 2887 Elo on LiveCodeBench Pro, 94.3% on GPQA Diamond — while questioning whether benchmarks still meaningfully capture model intelligence. The video argues we've entered a 'vibe era' where subjective user experience increasingly diverges from quantitative scores, featuring comparisons with the new Sonnet. Raises a fundamental question for builders: if benchmarks no longer predict real-world utility, what should guide model selection?

Graph relationships

RELEASED
Meta -> Code World Model
Meta published Code World Model, a 32-billion-parameter open-weights LLM for code generation research.
Source finding

Source trail

Hugging Face Blog AI at Meta Lushbinary — RAG Production Guide 2026 Humanoids Daily GitHub arXiv Jim Fan / NVIDIA

Graph sources

entity graphfindings textkg edgeskg entities