AI Model2026
A 3B-parameter causal LLM tuned for verifiable multi-step reasoning in math, coding and STEM using a Spectrum-to-Signal post-training pipeline (SFT, RL, offline self-distillation); not recommended for tool-calling/agent tasks.
Frames AI research as a trainable practice of reading, building, debugging, and fast feedback. The essay is most useful for researchers learning how to avoid hype-chasing, benchmark tunnel vision, and agent-induced blind spots.
Uses Parallel Looped Transformers (PLT) to make loop count a practical knob for code models, finding two loops give the best test-time gains. Trains 7B models on 18T tokens and attributes saturation beyond two loops to a gain–cost tradeoff from positional mismatch.
AI Agent Papers2026
Assesses whether coding agents can generate complete, playable games end-to-end inside the Godot engine. Implements an interaction-grounded evaluation (replayed demonstrations + rubric-guided multimodal judging) across 140 tasks and 15 game families; top agents score ~41%.
Proposes ZPPO, a distillation method that keeps the teacher inside prompts rather than injecting teacher gradients, using binary- and negative-candidate prompts plus a prompt replay buffer to recover learning signal on hard examples; shows gains for small Qwen3.5 students across 31 multimodal benchmarks.
Evaluates multimodal LLMs' ability to reconstruct past observations and act in controllable non-Markov games. Introduces RNG-Bench with two games (Matching Pairs, 3D Maze), three controllable difficulty axes, a head-to-head duel protocol, and a Memory Gap metric to separate forgetting from action errors.
AI Dataset2026
Benchmark for evaluating procedural skill evolution in LLM agents: isolates reusable skill bodies, role-specific work surfaces, and hidden oracle assets to measure whether skill refinements transfer across tasks, roles, and model backbones. Includes 382 workplace tasks, 22 skills, and a controlled evaluation protocol.
AI Dataset2026
Provides a small, manually annotated benchmark for evaluating vision–language models that convert robot and egocentric manipulation videos into timestamped subtask segments and concise action labels. Contains 100 episodes, 743 gold segments, and MP4 bytes embedded per row.
AI Model2026
35B Mixture-of-Experts agent model for long-horizon, multi-domain agent workflows; trained with a knowledge–action infrastructure that produces ~45K-token trajectories and supports native tool calling and function integration for research and deployment.
AI Dataset2026
Provides 2,170 reference-grounded evaluation samples across seven agent domains (MCP, Search, Terminal, SWE, Android, Web, OS) to score language world models on Format, Factuality, Consistency, Realism and Quality. Includes per-domain JSONL files, judge prompts and an evaluation script for reproducible scoring.
AI Dataset2026
Compares 30 frontier LLMs generating static SVG markup from 500 prompts using 1,355,161 human votes across three leaderboards (Preference, Coherence, Alignment); provides raw SVGs, 768×768 rasterized PNGs, and per-comparison human vote records under a CC-BY-4.0 prompt license.
AI Model2026
230M-parameter multilingual instruction-tuned text-only LLM for on-device agentic pipelines and data extraction; 32K context, 19T-token pretraining, optimized for fast CPU/edge inference (e.g., 213 tok/s on Galaxy S25 Ultra, 42 tok/s on Raspberry Pi 5); not for heavy reasoning or complex code generation.