Introducing BackSearch: A Time-Travel Web Search Tool for LLM Backtesting
rosstaylor90 · x · 2026-07-24
GenReasoning has released BackSearch, a search tool designed specifically for Large Language Models (LLMs) and agents.
- The Problem: When backtesting models on historical data (e.g., financial time series), default web search tools leak future information, compromising evaluation accuracy.
- Core Feature: It offers the exact same shape as standard web tools but allows users to set an asof date to simulate a snapshot of the internet at a specific point in time.
- Use Cases: Ideal for forecasting, quantitative finance, Reinforcement Learning (RL) environment simulation, and freezing web search for benchmark reproducibility.
- Caveat: Since LLMs can memorize future data through pre-training, users need to align the BackSearch date with the language model's cutoff.
The team has initially released a narrow index focused on the news domain for 2026, with plans to expand based on user feedback.
Related event: BackSearch: Time-Travel Web Search Tool to Prevent Data Leakage(2 posts)→
More from coding & agent
- Claude Code vs Codex: the harness may matter more than the model — al_kRicha · 2026-07-24
- Which model should power each agent task in production? — Logical-Silver-272 · 2026-07-24
- Offloop says its 4-person multi-agent harness beat Claude Code and Codex on GDPval — rohanpaul_ai · 2026-07-24
- Graph engineering, in one line: deterministic multi-agent loops become DAGs — vincent_koc · 2026-07-24
- BackSearch’s development was led by @ranko3000 with feedback from several collaborators — rosstaylor90 · 2026-07-24
- Default Codex CLI with GPT-5.5 scores 92.3% on XBOW, but the paper says the harness matters — evilsocket · 2026-07-24