Pencil Puzzle Bench – LLM Benchmark for Multi-Step Verifiable Reasoning
I've been working on applying LLMs to long-context, verifiable problems over the past year, and today I'm releasing a benchmark of 62,000 pencil puzzles across 94 types (sudoku, nonori, slitherlink, etc.). The benchmark also allows for intermediate checks /rule breaks for all varieties at any step. I tested 51 models against a subset (300 puzzles) in two modes: single-shot (output the full solution) and agentic (iterate with verifier feedback). Some results: - Best model (GPT 5.2@xhigh) solves 56%. (~ half the puzzles are unsolved by any model) - Agentic solves average 29 turns. The…
What it does
In the maker’s words, at launch
I've been working on applying LLMs to long-context, verifiable problems over the past year, and today I'm releasing a benchmark of 62,000 pencil puzzles across 94 types (sudoku, nonori, slitherlink, etc.). The benchmark also allows for intermediate checks /rule breaks for all varieties at any step. I tested 51 models against a subset (300 puzzles) in two modes: single-shot (output the full solution) and agentic (iterate with verifier feedback). Some results: - Best model (GPT 5.2@xhigh) solves 56%. (~ half the puzzles are unsolved by any model) - Agentic solves average 29 turns. The longest attempt took ~1,200 turns over 14 hours. - Cost per success varies wildly (cheapest: $0.00033 — Grok 4.1 Fast Reasoning, most expensive: $238.16 — Claude Sonnet 4.6 (1M context)) - Reasoning depth (eg. @medium, @high, @xhigh) dramatically improves capability (up to repeated infrastructure failure for @xhigh) - Stark difference between US closed models (3 at >33%) and Chinese open models (top: 6%) Made the website to show off the dataset + play every puzzle, and even every replay AI agent solves step-by-step (fun to watch how it gets to solutions). Also here's the paper: https://arxiv.org/abs/2603.02119 I didn't test human ability to solve, but it seems these puzzles are pretty difficult. I'd be curious how HN audience fares on the puzzles.
Does the same job
all alternatives →
- CBCheddar-bench – unsupervised benchmark for coding agentsFeb 2026 · github.com · ▲9
I built a small benchmark to test CLI coding agents on blind bug detection. A challenger agent injects bugs and writes ground truth (`bugs.json`). A different reviewer agent audits the repo without seeing ground truth, and an LLM matcher scores bug-to-finding assignments. Current run: 50 repos, 150 challenges, 450 reviews, 2,603 injected bugs. Weighted detection: Claude 58.05%, Codex 37.84%, Gemini 27.81%. LLM-judge benchmarks are easy to get wrong, so I’d really appreciate critical feedback on benchmark fairness, scoring/matching methodology, and obvious failure modes I’m missing. Full…
- MCMazeBattles.com – Race to solve generated mazes2018 · mazebattles.com · ▲63
- A1A 100-Line LLM Framework2025 · github.com · ▲9
I've seen a lot of comments about how complex frameworks like LangChain can be. Over the holidays, I wanted to see how minimal an LLM framework could get if we stripped away everything non-essential. The result is an LLM framework in just 100 lines of code. These 100 lines capture what I see as the core abstraction of most LLM frameworks: a nested directed graph that breaks down tasks into multiple LLM steps, with branching and recursion to enable agent-like decision-making. From there, you can layer on more advanced features like agents, RAG, task decomposition, and more. I’ve intentionally…
- ATA tiny reasoning layer that steadies LLM outputs (MIT; +22.4% accuracy)2025 · github.com · ▲6
We kept shipping “simple” LLM features that were fluent-but-wrong. After too many postmortems we wrote down the failure patterns and added a small reasoning layer in front of the model. It’s model-agnostic, sits beside your existing stack, and you can implement it from a single PDF (MIT). What’s inside the PDF A problem map of 16 failure modes we kept hitting in real systems (OCR/layout drift, table-to-question mismatches, embedding≠meaning, pre-deploy collapse, etc.). Four lightweight gates you can add today: Knowledge-boundary canaries (empty/adversarial/known-fact probes).…
- AOAuto-optimizing deterministic LLM outputs using knowledge graphs2024 · github.com · ▲7
Hi, We are building an open-source framework for loading and structuring LLM context to create accurate and explainable LLM answers using knowledge graphs and vector stores. We built the tool with four main concepts in mind: 1. Loader -> uses dlt in the backend to load and structure the data 2. Cognify step -> creates a graph with summaries, labels and factoids that are interconnected across the documents and stored as a representation in the vector store 3. Optimizer -> Uses DSPy to optimize LLM queries, and we plan to extend it to most of the knobs we can turn, like chunking etc. 4. Search…
More ai this month
the category →
I trained a 125M-parameter transformer to autocomplete piano performances in real time (~108 notes/sec on an iPhone 15). The idea is basically GitHub Copilot or Tabnine, except instead of prompting it with code, you prompt it by playing a few notes on a MIDI piano. The model then continues what you played, entirely on-device. The app is free if anyone wants to try it. Happy to answer questions about the model, training, Core ML, or the many things that didn't work.
AI · 17d ago · simedw.com
Astute▲585Automate your B2B brand going viral, with new media creators
AI · 18d ago · company-app.joinastute.com
Hey HN, Henry from Cactus here! We previously released Cactus Needle, a 14MB agentic LLM for tool call, device use, and structured extraction for phones, wearables, smart homes, small robots and microcontrollers. We got really great feedback here, and have now incorporated the suggestions to release Needle 2. The whole model is a single 14MB binary that runs a full session in 28MB of RAM; 45m parameters at 2bit compression. Needle hits 500 tokens/sec decode speed on a Raspberry Pi 5, sits between 400-1,500 tokens/sec on VR devices like Meta Quest 3S and Apple Vision Pro, and ranges…
AI · 26d ago · cactuscompute.com
Launched alongside, March 2026
the whole month →
Switch from ChatGPT to Claude with import memory feature
AI · Mar 2026 · claude.com
