https://ruisun33.github.io/The Thinking Notes 2026-07-07T01:12:46+00:00 R S https://ruisun33.github.io/ Jekyll © 2026 R S /assets/img/favicons/favicon.ico /assets/img/favicons/favicon-96x96.png Trace-Bench2025-02-25T00:00:00+00:00 2026-02-27T06:41:41+00:00 https://ruisun33.github.io/posts/trace-bench/ R S A benchmark and RL training substrate for evaluating causal diagnostic reasoning in AI agents TL;DR Ad-Agentic-Bench is a benchmark of 6,000 synthetic episodes that evaluates an AI agent’s ability to perform root cause analysis through iterative exploration using Python tools with SQL support. Each episode embeds a known causal intervention inside realistic, high-dimensional time-series data... The Geometry of Learnable Reward2025-02-25T00:00:00+00:00 2026-02-27T06:34:21+00:00 https://ruisun33.github.io/posts/geometry-of-learnable-reward/ R S A Framework for Understanding RL with Verifiable Signals 1. The Fundamental Object: Mapping Complexity RL with verifiable rewards asks a model to learn a function: [f: \mathcal{S} \rightarrow \mathcal{A}] where $\mathcal{S}$ is the space of problem states and $\mathcal{A}$ is the space of verifiable answers. But this framing is deceptively simple. The real object being learned is not this... Dynamic Context Management2025-02-24T00:00:00+00:00 2026-02-27T06:41:41+00:00 https://ruisun33.github.io/posts/dynamic-context-control/ R S Adaptive context management for LLMs via Primal-Dual Thompson Sampling TL;DR LLMs have finite context windows, and current methods for deciding what to keep (recency truncation, summarization) are static heuristics that ignore task structure and system constraints. We reframe context management as a constrained resource allocation problem and propose Primal-Dual Thompson Sampling (PD-TS) — a... Ad Arena2025-02-23T00:00:00+00:00 2026-02-27T06:41:41+00:00 https://ruisun33.github.io/posts/ad-arena/ R S A simulation environment and LLM benchmark for search advertising bidding strategies TL;DR Ad Arena provides a simulation environment for search advertising auctions, extended into a public LLM benchmark with a web leaderboard. It pits language models against baseline bidding strategies in realistic scenarios, measuring not just final performance but how quickly agents learn to optimize over...