TL;DR
Models do best at multi-step search when they display four specific reasoning patterns: they verify information across multiple sources, assess how reliable those sources are, switch search strategy when the first one fails, and fix their own mistakes. How often these behaviors show up correlates directly with final answer accuracy — the pattern holds from the weakest models to the strongest.
The study's main finding flips the usual training logic: the reasoning process matters more than whether the final answer is right. A model trained on examples with correct reasoning but wrong answers, after further training, performs just as well as a model trained on correct answers. The catch is that models don't develop these behaviors spontaneously — even after reinforcement learning their frequency goes down, not up.
The fix — Behavior Priming: first train the model on trajectories that display all four behaviors (regardless of whether the answer is right), then apply standard RL. That yields +35% accuracy compared with direct RL, because it lays the groundwork for effective exploration (higher pass@k, more policy entropy) and for test-time scaling (longer trajectories).
Want this for your own case? Apply this study to your task: AI adapts its method to what you are working on and writes the prompt.
Four Behaviors for Effective Search
1. Information Verification Seeking information from several independent sources, cross-referencing, and explicitly citing evidence during reasoning.
2. Authority Evaluation Spotting contradictions between sources, analyzing source reliability (official documents vs forums vs news sites), and prioritizing the most authoritative information.
3. Adaptive Search Dynamically changing the search strategy based on previous results — different keywords, a different angle on the problem — when the first attempt didn't surface the needed information.
4. Error Recovery Recognizing mistakes in earlier steps (misinterpretation, an incorrect inference) and actively correcting them.
Example in Action
Task: You're checking out a startup called "SmartStock" before a job interview — the company promises a revolution in logistics, but you want to understand what's really going on.
Prompt:
I need to fact-check the startup "SmartStock" (EIN 12-3456789).
Follow these principles:
1. VERIFICATION: Find information from at least 3 different types of sources
(official website, news, reviews, business registration records)
2. AUTHORITY: If the sources contradict each other,
assess how reliable each one is and explain why you trust one more
3. ADAPTATION: If the first search doesn't turn up what you need,
try other keywords or another angle (instead of "SmartStock reviews",
search "SmartStock employees Glassdoor" or "warehouse automation 2024")
4. CORRECTION: If you drew a wrong conclusion at any step,
flag the mistake explicitly and fix it
Show me the whole search process, not just the final answer.Result: The model shows a multi-step process: first search → what it found → second search with different keywords → comparing sources → spotting contradictions (the website says "$5M revenue", the business database shows "$500k") → judging which source to trust → a final conclusion with reasoning.
Why It Works
LLM weakness: models tend to trust the first answer they find and don't verify it across sources. When they run into contradictions, they often pick at random rather than by source reliability. They don't go back to earlier steps to fix mistakes.
LLM strength: models handle a structured reasoning process well when the steps are spelled out: "first find, then compare, then assess reliability". They can simulate critical thinking if it's written into the instructions.
How the principles use that strength: the four behaviors turn chaotic searching into a systematic process. Instead of "find and answer", the model follows a clear algorithm: find → verify against other sources → assess reliability → adapt the strategy if needed → fix mistakes. That reduces cognitive load and raises quality.
The study's key insight: these behaviors don't appear spontaneously even after RL training. You have to instill them explicitly through examples, and then the model learns to apply them more effectively via RL.
Prompt Template
Task: {description of a task that requires finding information}
Follow the principles of effective search:
1. INFORMATION VERIFICATION:
- Find information from at least {N} different sources
- Compare them against each other
- State explicitly which source says what
2. AUTHORITY EVALUATION:
- If sources contradict each other, assess the reliability of each
- Prioritize: {types of authoritative sources for your task}
- Explain why you trust one source more than another
3. ADAPTIVE SEARCH:
- If the first search comes up empty, try other keywords
- Change the angle: {examples of alternative angles}
4. ERROR RECOVERY:
- Check your earlier conclusions
- If you find a mistake, flag it explicitly and fix it
Show the whole process, not just the final answer.How to fill it in:
{description of a task}— your specific search task{N}— the minimum number of sources (usually 3-5){types of authoritative sources}— for your field (official documents > expert blogs > forums){examples of alternative angles}— different ways to phrase the search
Limitations
⚠️ Trajectory length: The principles increase the number of steps (on average from 4 to 7 steps per task). For simple questions that's overkill — one search is enough.
⚠️ Tokens: Demonstrating all four behaviors on every task takes more tokens. For routine tasks that's inefficient.
⚠️ Subjective criteria: Authority Evaluation works well for factual information (laws, statistics, technical data) but worse for subjective judgments (which product is better, which approach is more correct).
How They Tested It
The team ran a strong model (Gemini 2.5 Flash) and a weak one (Qwen3-1.7B) on the same questions through a single agent framework. They took 200 pairs of trajectories where Gemini solved the task and Qwen didn't. The analysis prompt: "Why is one trajectory successful and the other not?" They collected the answers, extracted common patterns, deduplicated → and ended up with 4 behaviors.
Validation: they measured how often these behaviors occurred in trajectories from Gemini 2.5 Flash, DeepSeek R1, Llama 3.2-3B and Qwen3-1.7B on three benchmarks (GAIA, WebWalker, HLE). The ranking of models by performance matched the ranking by behavior frequency exactly — from weakest to strongest.
Then they built datasets from 20k Gemini trajectories:
- SFT (Random) — random trajectories
- SFT (Correct) — only those with the correct answer
- Behavior Prime — only those with the four behaviors (regardless of answer correctness)
- Behavior Prime (Incorrect) — with the behaviors but a wrong answer
- Behavior Prime (Correct) — with the behaviors and the correct answer
They trained Qwen3-1.7B and Llama 3.2-3B on each dataset (SFT), then ran every model through the same RL (GRPO, 300 steps).
The surprising result: the model trained on Behavior Prime (Incorrect) (behaviors present, answers wrong) performed after RL just as well as Behavior Prime (Correct) (behaviors and answers both right). Both were 35% better than direct RL without SFT.
The mechanics of why it works: the behaviors raise pass@k (diversity of solution paths) and policy entropy (it doesn't collapse into a single strategy). That provides the foundation for effective RL. Models without the behaviors start with low entropy and quickly plateau. With the behaviors, high entropy is preserved, so the model explores longer and finds better solutions.
Resources
Beneficial Reasoning Behaviors in Agentic Search and Effective Post-training to Obtain Them
Jiahe Jin, Abhijay Paladugu, Chenyan Xiong (Carnegie Mellon University)
The code is to be released as open source (still in progress at the time the preprint was published).