AI Evaluation · Search Agents

When queries get vague,
agents fall apart.

First surfaced in Tandemly Briefing — 2026-05-29.

Xiaohongshu researchers built a benchmark to test search agents on how people actually search: vaguely, over multiple turns, revealing what they want piece by piece. The best frontier model scored 30.3 out of 100. The gap between that number and a passing grade is the gap between benchmark performance and what users actually experience.

Core concept
Vibe search: a pattern where users progressively disclose intent across turns, rather than specifying it upfront. The agent must ask good questions, not just retrieve well, to close the gap.
scroll to explore

Benchmarks say agents search well.
Users say they don't.

There is a persistent mismatch between how search agents score on benchmarks and how well they perform for real users. VibeSearchBench is an attempt to close that gap by testing what the existing benchmarks skip.

Most AI search benchmarks share an assumption that rarely gets examined: the user already knows what they want. Queries are well-formed. The correct answer has a predefined structure. The agent's job is to retrieve the right content and produce output in the right format. That makes evaluation tractable. It also makes it artificial.

Real information-seeking doesn't work this way. A researcher starting a literature review might say "find me recent work on memory in transformer agents." Over the course of the search they realize they care specifically about retrieval-augmented approaches, they want results from the last year, and safety considerations matter too. None of that was in the original query. The full intent emerged through the search itself.

Existing benchmarks test whether an agent retrieves correctly once intent is clear. They don't test whether the agent can surface intent before retrieval. That second skill is what VibeSearchBench measures: proactive clarification across multiple turns, with no preset answer schema, and a user who reveals what they want only in response to good questions.

The question this paper asks

How well do frontier search agents perform when queries are vague, answers are schema-free, and the agent must earn information from a user through dialogue? The answer, it turns out, is not well. The best score is 30.3 Triplet F1.

200 tasks, two tracks,
no fixed answer shape.

The benchmark replaces single-turn fixed-schema queries with persona-driven conversations evaluated against knowledge graphs. Every design choice resists the shortcuts that inflate scores on simpler benchmarks.

Standard benchmark design
Query is fully specified. One turn. The answer has a fixed schema: a document ID, a ranked list, a structured JSON response. Agents compete on retrieval accuracy against a known target format.
VibeSearchBench design
Query is a vague starting point. Multiple turns required. Output is a knowledge graph: a directed graph of entities and relationships with no preset structure. Agents compete on whether they asked the right questions to build the right map.
1
User persona and progressive disclosure simulator
Each task pairs an agent with a simulated user who has a specific background and a latent information need. The simulator, run by an LLM, reveals preferences and constraints only in response to the agent's questions. Information is unlocked through clarification, not offered upfront. A good opening question unlocks more of the ground truth than a fast retrieval hop does.
2
200 bilingual tasks across 20 domains
VibeSearch-Pro (100 tasks) covers professional research: literature reviews, market analysis, technical due diligence. VibeSearch-Daily (100 tasks) covers everyday needs: shopping, travel, lifestyle decisions with evolving preferences. Tasks span Chinese and English, surfacing language-switching coherence failures that monolingual benchmarks miss entirely.
3
Knowledge graph evaluation with Triplet F1
The ground truth for each task is a directed graph encoding the target information as (entity, relation, entity) triplets. The agent's output graph is matched against ground truth using LLM-as-judge for node alignment and relation semantic equivalence, so synonymous phrasings aren't penalized. Triplet F1 measures precision and recall over matched triplets: how many did the agent find, and how many of what it found were correct?
Why schema-free outputs close a real gap

Fixed-schema benchmarks let agents score well by matching format without recovering content. A knowledge graph has no preset shape: the agent must discover what the correct output structure is by learning what the user actually needs. That's a different skill from retrieval, and it's the one that fails when queries are vague.

30.3 Triplet F1.
Plenty of room to improve.

The best frontier model captured roughly a third of the target information. Three patterns explain where the rest was lost.

Total tasks
200
100 pro + 100 daily
Domains covered
20
bilingual (zh + en)
Best frontier score
30.3
Triplet F1 (Claude + OpenClaw)
What 30.3 actually means

A score of 100 would mean perfect recovery of every entity and relationship in the ground-truth knowledge graph. At 30.3, the best-performing system captured roughly a third of the target triplets. The rest were missed: entities not identified, relationships not surfaced, or information the agent never thought to ask about. A single-turn retrieval benchmark would likely show a very different picture for the same models on the same underlying topics. The gap between the two scores is the cost of intent ambiguity.

Finding 1: Clarification behavior is the binding constraint

The agents that scored highest were the ones that asked better questions in the early turns. Retrieval quality had less influence on the final Triplet F1 than the agent's ability to surface the user's latent intent before retrieving anything. Good clarifying questions unlock information the retriever can then find. Poor or absent questions leave entire branches of the knowledge graph unexplored.

Finding 2: Professional tasks are substantially harder

VibeSearch-Pro tasks (literature reviews, market analysis, technical due diligence) scored lower than VibeSearch-Daily tasks across all models. Domain-specific professional searches require the agent to understand the shape of knowledge in unfamiliar fields and to ask questions that would make sense to a domain expert. Everyday search tasks involve evolving preferences, but the underlying knowledge structures are less opaque.

Finding 3: Bilingual tasks expose coherence failures

Chinese and English tasks tested whether agents could switch languages mid-task without losing context or continuity. Several models showed accuracy drops when the conversation crossed language boundaries in ways that monolingual benchmarks would never catch. The gap between Chinese-language and English-language subtask performance was not uniform across model families, suggesting that language-switching coherence is a distinct capability, not a side effect of general multilingual fluency.

Scope and limitations

200 tasks, manually curated, with LLM-as-judge evaluation for triplet matching. The judge's semantic equivalence decisions introduce some variance, particularly for domain-specific relation types. The benchmark is new; published results should be treated as baselines, not as ceilings. The authors note that Triplet F1 captures whether information was recovered but not how efficiently the agent's clarification strategy reached it. A high-scoring agent that asked 30 questions to recover 30 triplets is different from one that asked 5 targeted questions to reach the same result.

What this means
for search agent builders.

The gap between 30.3 and a passing score is an engineering target. Here is where to look first.

1
Evaluate on clarification behavior, not just retrieval accuracy
If your search agent eval only measures whether the final answer is correct, you are testing the end of the pipeline while ignoring where most errors enter. Add a multi-turn clarification phase to your evals. Measure whether the agent asks questions that would help a real user. Retrieval accuracy on fully specified queries tells you performance under ideal conditions. VibeSearchBench-style eval tells you performance under realistic ones.
2
Add schema-free outputs to at least one eval track
Fixed-schema evals reward format-matching. Schema-free evals reward content recovery. A knowledge-graph output forces the agent to discover the right output structure rather than matching a template. If your agent only gets tested against fixed schemas, you won't know how it performs when users haven't handed it the answer format upfront. That's most real product use.
3
Test professional and everyday search tasks separately
VibeSearch-Pro and VibeSearch-Daily produced different score profiles across models. If your product targets professional use cases (research, analysis, due diligence), don't assume performance on general-purpose search benchmarks transfers. The knowledge structures are more opaque, and clarification requirements are more domain-specific. A separate professional eval track will surface this gap.
4
For multilingual products: add cross-language coherence tests
Bilingual task performance exposed coherence failures that monolingual benchmarks miss. If your product serves users who switch languages or who search in one language for content in another, add cross-language coherence tests. The failure mode isn't mistranslation. It's loss of context when language switches mid-conversation, and it's invisible to standard monolingual benchmarks.
5
Treat single-turn retrieval scores as upper bounds for intent-incomplete tasks
The same models that perform well on single-turn retrieval benchmarks scored 30.3 on VibeSearchBench. Single-turn scores represent performance under ideal conditions: full intent, clear schema. Most real user queries don't meet those conditions. Measure both, and use the delta between them to understand how much your agent depends on query quality rather than clarification skill.

Where to go
from here.

The benchmark is public. The implementation is open. The baselines are low.

1
Read the paper
Xiaohongshu Inc. (2026). VibeSearchBench: Benchmarking Long-horizon Proactive Search in the Wild. arXiv:2605.27882.
2
Explore the benchmark repository
The full benchmark, task definitions, user simulator, and evaluation code are at github.com/VibeBench/VibeSearchBench. Two agent implementations are provided: a GeneralAgent (OpenAI-compatible, with search, visit, and Python execution tools) and an OpenClaw agent.
3
Start with VibeSearch-Pro
Begin with the professional subset. It is harder, the knowledge structures are more opaque, and the score gap between Pro and Daily will tell you how much of your agent's performance depends on familiar domain structure versus genuine clarification skill. The gap is diagnostic before it is a benchmark number.
4
Compare your VibeSearchBench score against a single-turn benchmark
Run the same model on a single-turn retrieval benchmark and on VibeSearchBench. The difference between the two scores is the cost of intent ambiguity in your specific deployment context. That gap is what a better clarification strategy needs to close.
5
For related work on search agent failure modes
The Tandemly synthesis of "Is Grep All You Need?" covers a complementary finding: the agent harness layer moves retrieval accuracy more than the retrieval method itself. Together the two papers suggest that both the scaffold and the clarification behavior matter more than retrieval method alone. See: grep-vs-vectors-agentic-search.