Alternatives
Products that do what Pencil Puzzle Bench – LLM Benchmark for Multi-Step Verifiable Reasoning does
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…
- 1
- 2
- 3CB
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…
Feb 2026 · github.com
- 4MC
2018 · mazebattles.com
- 5A1
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…
2025 · github.com
- 6AT
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).…
2025 · github.com
- 7AO
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…
2024 · github.com
- 8LT
Measures the ability of various LLMs to navigate a fictional codebase via iterative directory tree expansion and observation. Each model's baseline ability is compared against combinations of various prompt engineering mods to quantify exactly how much they help or hinder the LLM. Interesting findings here: https://github.com/aiwebb/treenav-bench#interesting-findings
2024 · github.com
- 9
- 10AR
If you're interested in exploring what LLM-based agent systems these days actually do to solve certain benchmarks such as SWEBench or WebArena, we created a small leaderboard with our team, that allows to view a lot of public and OSS agent results including all the runtime traces (the step-by-step reasoning behind the scenes). Looking at traces is actually quite interesting, as they reveal a lot about the inner working and shortcomings of current agent system, e.g. see https://explorer.invariantlabs.ai/u/invariant/webarena--SteP... for an example trace.
2024 · explorer.invariantlabs.ai
- 11CA
Hi HN, I'd like to share a logic number puzzle game I created! It is somewhat reminiscent of Sudoku and Minesweeper, but with its own twist. The rules are simple but it can be quite challenging! Please enjoy :) The feedback on the game prototype was invaluable and I incorporated a lot into what I hope is a more polished version of the game. What's new: Introduced a daily puzzle challenge. Dark mode. Enhanced user experience with improved UI. I'd love to hear your thoughts on these updates. Feedback is always appreciated!
2023 · cubisum.com
- 12BA
I built CodeLens.AI - a tool that compares how 6 top LLMs (GPT-5, Claude Opus 4.1, Claude Sonnet 4.5, Grok 4, Gemini 2.5 Pro, o3) handle your actual code tasks. How it works: - Upload code + describe task (refactoring, security review, architecture, etc.) - All 6 models run in parallel (~2-5 min) - See side-by-side comparison with AI judge scores - Community votes on winners (blind voting) - Each evaluation gets reflected in the overall AI model leaderboard, showing us best ones Why I built this: Existing benchmarks (HumanEval, SWE-Bench) don't reflect real-world developer tasks. I wanted to…
Oct 2025 · codelens.ai
- 13IO
Hey folks, I’m the creator of WFGY — a semantic reasoning framework for LLMs. After open-sourcing it, I did a full technical and value audit — and realized this engine might be worth $8M–$17M based on AI module licensing norms. If embedded as part of a platform core, the valuation could exceed $30M. Too late to pull it back. So here it is — fully free, open-sourced under MIT. --- ### What does it solve? Current LLMs (even GPT-4+) lack *self-consistent reasoning*. They struggle with: - Fragmented logic across turns - No internal loopback or self-calibration - No modular thought units - Weak…
2025 · github.com
- 14GA
I wanted to learn more about RAG implementations, so I built something to solve the constant digging through manuals whenever we play a game. It's fairly simplistic, but actually has worked pretty well for some of these conflicts. Everythings Open Source on GitHub if you're curious (or have ideas), and I'd love to hear feedback from fellow boardgamers!
2024 · gamegame.ai
- 15WM
We wanted to test if a smaller model like GPT-4.1-mini could beat its bigger brother 4.1 at the game Tic-Tac-Toe using only context engineering. We put them in a 100-game tournament. For the smaller model, we gave it a few examples of winning moves from past games right before it made its own move. The results were clear. Without the examples, the smaller model struggled against GPT-4.1. With the examples, its effectiveness increased by nearly 200%, and it consistently won. It's a simple demonstration, but it shows that a smaller, faster model with good, timely examples can outperform a more…
2025 · github.com
- 16MD
We’re excited to share ML-Dev-Bench, a new open-source benchmark that tests AI agents on real-world ML development tasks. Unlike typical coding challenges or Kaggle-style competitions, our benchmark simulates end-to-end ML workflows including: - Dataset handling and preprocessing - Debugging model and code failures - Implementing new model architectures - Fine-tuning and improving existing models With 30 diverse tasks, ML-Dev-Bench evaluates agents across critical stages of ML development. To complement this, we built Calipers, a framework that provides systematic performance evaluation and…
2025 · github.com
- 17AD
Ever wish you could get the best arguments for both sides of a debate? I built an AI-powered debate platform that pits language models against each other on controversial topics. Each AI is randomly assigned a side (pro/con). You vote before and after to see if you were persuaded. Most content today presents lopsided arguments. They provide strong points for one side, weak ones for the other. This project aims to surface the strongest arguments from both sides, using LLMs to simulate a fair debate. With enough usage, I want to use it to benchmark LLMs. My hypothesis is that randomly…
2025 · bot-bicker.vercel.app
- 18IL
LLM Application development is extremely iterative, more so than any other types of development. This is because in addition to all the activities involved in regular application development, we also need to make the LLM Application accurate and reduce hallucination. To improve performance, we need to trial and error various combinations of LLM models, prompt templates (e.g., few-shot, chain-of-thought), prompt context with different RAG architecture, try different agent architecture, and more. There are thousands of permutations to try. We need to be able to easily experiment with these…
2024 · palico.ai
- 19TA
In this post, we document the results of some experiments comparing vanilla Graph RAG (just a single pass of text2cypher) vs. a router agent Graph RAG approach that can call vector search tools alongside text2cypher. The routing agent uses an LLM to decide which vector search tool to call, depending on the terms identified in the question, and it works quite well. The results show that recent frontier LLMs like `gpt-4.1` and the trusty workhorse `gemini-2.0-flash` produce great quality Cypher reliably and reproducibly, with some prompt engineering to ensure that the graph schema is formatted…
2025 · blog.kuzudb.com
- 20CL
As I was playing with the Outlines library (https://outlines-dev.github.io/outlines/), I discussed with my friend Maxime how funny it would be if we set up a way to pair LLMs in chess matches till one wins. The first time I tried it, it required substantial prompt engineering to get some of those LLMs to propose valid moves. Large language models can mostly stay focused and even play rather well; see https://news.ycombinator.com/item?id=37616170 for example. However small language models aren't as easy to convince. Some of those LLMs have seen very little…
2024 · huggingface.co
- 21IS
Hey HN! For that last 8 months I've been trying to make agents that can hack web applications to find vulnerabilities in them - An AI Security Tester. The system has 29 agents in total, a custom LLM Orchestration framework which works on the task-subtask architecture (old-school but works amazingly for my use case, and is pretty reliable) with custom agent calling mechanism. No Auo-Gen, Langchain and Crew AI - Everything custom built for pentesting. Each test runs in an isolated Kali linux environment (on AWS Fargate), where the agents have full access to the environment to undertake any…
2025
- 22RA
I built a local-first UI that adds two reasoning architectures on top of small models like Qwen, Llama and Mistral: a sequential Thinking Pipeline (Plan → Execute → Critique) and a parallel Agent Council where multiple expert models debate in parallel and a Judge synthesizes the best answer. No API keys, zero .env setup — just pip install multimind. Benchmark on GSM8K shows measurable accuracy gains vs. single-model inference.
Mar 2026 · github.com
- 23GB
Hey HN, We’re excited to share PySpur, an open-source tool that provides a graph-based interface for building, debugging, and evaluating LLM workflows. Why we built this: Before this, we built several LLM-powered applications that collectively served thousands of users. The biggest challenge we faced was ensuring reliability: making sure the workflows were robust enough to handle edge cases and deliver consistent results. In practice, achieving this reliability meant repeatedly: 1. Breaking down complex goals into simpler steps: Composing prompts, tool calls, parsing steps, and branching…
2024 · github.com
- 24R0
I built a tiny physics solver LLM that performs surprisingly well on easy-to-medium difficulty physics problems. Most LLMs today still struggle with physics QA (as PhyBench recently highlighted), so I wanted to see how far I could push a small model with careful data and minimal compute. Model: Qwen3-1.7B Supervised Finetuning: ~1500 curated examples spanning kinematics, EM, acoustics, and more RL Fine-tuning: GRPO, 1-shot RLVR style (single example, 70 steps) Total cost: ~$5 on H100 It started with a cold-start SFT (~3 epochs, loss to 0.3), then I ran RL with accuracy reward that climbed…
2025 · huggingface.co
Ranked by how close each launch is in meaning, then by votes. Refine with a description →