Back Home

AI 研究

Seer Lets Small Research Agents Simulate Search Results Before Choosing What to Do Next

The authors report that a 4B agent’s accuracy on GAIA rose from 13.6% to 41.7%. But the system also requires a shared 9B auxiliary model, and hardware measurements have yet to support claims of real-world cost savings or speedups.

zh-Hant

Small research agents comparing several search directions can first predict what each path might find, then choose one to execute for real. Seer turns this idea into a research framework: candidate paths that are not selected do not need actual tool calls, at the cost of additional model inference. Researchers from Sichuan University and other institutions submitted the original paper on November 19. What can currently be confirmed is the authors’ experimental results; these do not establish that deployment on a typical computer would be cheaper or faster. Paper and submission record

Predictions help choose a path; only real responses count as evidence

Seer has three roles: an agent proposes candidate actions; a “tool-response simulator” predicts what a search or webpage visit might return; and a “value estimator” scores how likely the task is to succeed after adding that content. The two auxiliary models are first trained on real tool-interaction data, then reused across agents without retraining each one individually.

Evaluation simulates only one step ahead. After an action is selected, the system obtains the actual tool result and adds only the real response to the subsequent history. Simulated content influences path selection, but is not carried forward as citation evidence. Method and experimental setup

Higher answer accuracy comes with simulation error

GAIA is a question-answering benchmark that tests reasoning, web browsing, and tool-use capabilities. Original GAIA paper The Seer authors report that Qwen3.5-4B’s accuracy rose from 13.6% for the baseline agent to 41.7%, an increase of 28.1 percentage points; the latter compared 32 candidates at a time. The comparison used the same agent and tools, with up to 100 rounds per path, an input limit of 128k tokens, and a time limit of 9,000 seconds. These are author-reported results, not measurements or independent reproductions by this publication. Table 2 and evaluation setup

Simulation can also lead the system down the wrong path. In a comparison of four candidates, the 4B agent paired with an estimator achieved 38.8% accuracy on GAIA when it executed the candidates to obtain responses; accuracy fell to 29.1% when simulated responses were used for scoring. Expanding the candidate set can recover some of the loss, but simulated results cannot be treated as equivalent to real searches. Component comparison, Table 4

Local costs remain a key question before adoption

“Unselected branches do not search” does not mean the overall task will necessarily cost less: path selection and task length still affect the total number of calls, and simulation itself requires computation. The authors also note that their comparison between Seer and executing four complete paths did not match total compute. Cost definitions and comparison conditions

The auxiliary models also require resources: both are based on Qwen3.5-9B, share base weights, and use different LoRA fine-tuning adapters. The experiment’s webpage excerpts were additionally processed by a Qwen3.5-397B-A17B service, which all comparison methods shared. So 41.7% should not be read as the capability of a single 4B model running entirely offline. The authors did not measure peak memory, power consumption, or latency on actual devices. Training setup and deployment limitations

For readers running their own agents, the concrete tradeoff in this study is to spend more inference comparing search directions in order to reduce external call costs for options that are not selected. The reported figures support improvements in tool-assisted question answering; whether those translate into cost advantages on everyday devices remains to be determined through on-device measurements.

Sources

  1. Seer: Guiding Small Deep-Research Agents with Shared Decision Models
  2. Seer: Guiding Small Deep-Research Agents with Shared Decision Models
  3. GAIA: a benchmark for General AI Assistants