---
title: ThinkRetrieve Research Explorer
emoji: 🔎
colorFrom: blue
colorTo: indigo
sdk: static
app_file: index.html
pinned: false
tags:
  - arxiv:2608.10928
  - reasoning
  - retrieval-augmented-generation
  - test-time-scaling
---

# ThinkRetrieve Research Explorer

**Accepted to Findings of EMNLP 2026.**

[ThinkRetrieve: Retrieval-Augmented Reasoning Traces for Test-Time Scaling](https://arxiv.org/abs/2608.10928)
retrieves worked solutions inside an ongoing reasoning trace.

This interactive viewer shows **recorded outputs**, not live model inference.
Compare prompt retrieval, plain reasoning, extra thinking, and intermediate-state
retrieval across every case in the repository's small science-agent demonstration.

The six synthetic incidents and private tool routes were selected after exploratory
screening. They are not a held-out evaluation, a general agent accuracy estimate,
or a biological safety benchmark. This demonstration is separate from the paper's
math/science benchmark evaluation. A failing case is shown first.

Data source: [repository revision 7e245c1](https://github.com/itsvaibhav01/ThinkRetrieve/tree/7e245c10b05a5b1c83f283a659cd210a0e19c8fb/thinkretrieve/benchmarks/science_agent).

- [Paper and citation](https://itsvaibhav01.github.io/ThinkRetrieve/paper/)
- [Project and code](https://github.com/itsvaibhav01/ThinkRetrieve)
- [Experiment design and limitations](https://github.com/itsvaibhav01/ThinkRetrieve/blob/7e245c10b05a5b1c83f283a659cd210a0e19c8fb/thinkretrieve/benchmarks/science_agent/README.md)

Vaibhav Singh, Soumya Suvra Ghosal, Sarvesh Gharat, Soumyabrata Pal,
Ramasuri Narayanam, and Dinesh Manocha. ThinkRetrieve: Retrieval-Augmented
Reasoning Traces for Test-Time Scaling. 2026. DOI: 10.48550/arXiv.2608.10928.

Acceptance status is author-confirmed. The linked arXiv version remains the
available bibliographic record until the official proceedings record is available.