Training on deliberately vague questions makes a 14B agent search deeper

We taught models to hold a conversation and solve equations a long time ago, but out in the real world they stumble on search and fact-checking. A single query is rarely enough: you have to follow leads, refine them, cross-check. The InfoAgent team built exactly that kind of web detective — an LLM agent that searches long and deliberately, reads pages, backtracks and keeps going. The core idea is to block the easy shortcuts and force the system to build a chain of steps the way an experienced analyst would.

How it works under the hood

The architecture is ReAct-style: the model alternates reasoning with calls to two tools — search and browse. Search returns a list of URLs with short but carefully assembled snippets; browse returns a long excerpt from a chosen page. The full history of steps and retrieved fragments goes into the context, so the agent leans on its own reasoning trail rather than on the model's memory alone.

Where to get tasks that break simple strategies

This is the interesting part. The authors build entity trees from Wikipedia and then deliberately blur the key details: names are replaced with descriptions, exact figures and dates with ranges, specific wordings with paraphrases. Subtrees are sampled from those trees, and the questions ask not about the object itself but about one of its attributes — the name of the team where an athlete made his debut, say — so you have to unwind the whole branch. They also discard questions that turn out too easy, automatically: if a strong model with web search gets one in a single click, that example doesn't make the cut.

Why a plain wiki retriever isn't enough

The team didn't treat commercial APIs as a black box and built its own search and browse infrastructure. The search flow works like this: take results from a search engine, crawl the pages, split them into fragments, filter through BM25, embeddings and a reranker, and only then ask an LLM to assemble a short, on-topic snippet. Browse uses a similar pipeline, minus the snippet generation and with a long fragment instead. This improves recall, controllability and speed, and it also buys stability during reinforcement learning, where the query volume is enormous.

Training: long trajectories as the cold start, RL as the amplifier

The model is Qwen3-14B, first fine-tuned on generated labeled trajectories (SFT) and then strengthened with RL (GRPO). One detail matters: the SFT trajectories are deliberately long — around 20 tool calls on average, often far more. That instills the right habit of digging deep instead of lunging at the first match. RL then teaches the model to pick tools better, refine queries and carry a solution through to the end.

What it does in practice

On hard benchmarks the agent holds its own in the sub-15B class, and in places it overtakes larger open models.

• BrowseComp: 15.3 (best result among <15B; above WebSailor-72B at 12.0)
• BrowseComp-ZH: 29.2 (strong cross-lingual generalization despite English-only training)
• Xbench‑DS: 40.4
• WebWalkerQA: 52.7 (above some 32B models)
• SimpleQA: 90.4 (close to large open systems with search)

Another visible effect: moving from SFT to RL lifts quality sharply and makes the agent's tool use smarter. Without SFT it repeats itself, issues few queries and runs out of steam fast. With SFT it starts using a wider variety of steps and sees a search through to an answer.

Caveats and limits worth knowing about

• Tool quality is critical. Swap the custom search for a simple wiki retriever and accuracy drops, then hits a ceiling fast.
• Very long trajectories cut both ways: they raise quality on hard tasks but burn context and time. The practical option is a mix of long and short examples.
• The authors tried adding a process reward for finding target entities along the trajectory. The training curve goes up, but final accuracy barely moves: the bonus lands on the wrong step, and the tasks are already solved reasonably well by a binary reward.
• In the current version RL ran at a 16k–32k context. Widening the context and going beyond Wikipedia would make the agent more useful still — and harder to train.

The takeaway

InfoAgent shows how to combine data, environment and training so that an LLM stops making facts up and becomes a patient researcher. Owning a transparent search stack buys control and scalability. And synthetic tasks built on blurred facts teach the agent to verify rather than guess. For products that need reliable facts, source checking and reproducibility, this is a step in the right direction.