Test-Time Scaling: BrowseConf and Confidence-Guided Reasoning

🌏 中文版

The previous articles covered training, architecture, evaluation. This one covers a new dimension: test-time compute allocation.

How much time should a deep research agent spend on a task? Fixed 10 rounds? 20? Or depends on the situation?

BrowseConf (ACL Findings 2026) answers: let the agent tell you.

Core Problem: Test-Time Resource Allocation

Existing deep research agents face a dilemma:

• Too much: every task runs the maximum rounds, wasting resources
• Too little: simple tasks also run many rounds, inefficient
• Don't know how much: no mechanism to decide when "enough"

The traditional approach is a fixed cap (e.g., max 30 rounds). But research finds:

• Some tasks need only 5 rounds
• Some need 50 rounds
• The agent itself should know when it's done

BrowseConf: Confidence-Guided Test-Time Scaling

Core Insight

BrowseConf's finding is straightforward:

When the model is accurate on a task, its confidence is high; when accuracy is low, confidence is low.

Moreover, when the model "doesn't know," it says "I'm not sure"—rather than guessing.

This means: confidence is an effective signal for deciding when to stop, whether to rethink, and whether to continue searching.

Mechanism

agent completes reasoning → outputs answer + confidence score
→ judges confidence:
  → high confidence → accept answer, finish
  → low confidence → rethink, continue searching
→ repeats until confidence threshold or round limit

Key Results

Core conclusion: low-confidence answers are essentially random—the model knows it doesn't know, and should continue researching rather than answering forcefully.

Relationship to Test-Time Scaling

BrowseConf belongs to the broader Test-Time Scaling (TTS) category:

BrowseConf's unique contribution: using confidence as a signal for dynamic resource allocation rather than fixed allocation.

Practical Implications

For Developers

1. Cost optimization: high-confidence tasks use minimal tokens, reducing API costs
2. Quality control: low-confidence tasks automatically flagged, alerting users answers may be unreliable
3. Adaptability: the same model handles everything from simple to extremely difficult tasks

For Users

• Get not just answers, but confidence labels on answers
• For questions the agent "doesn't know," it continues researching rather than answering randomly

For Deep Research Systems

BrowseConf's paradigm integrates into any deep research pipeline:

• Inserted during reasoning as a stopping condition
• Combined with IterResearch's "strategic forgetting" (compress at high confidence, retain at low)
• Combined with STC's evaluation framework (confidence calibration is part of evaluation)

Series Connections

Key Takeaways

1. Expressing uncertainty is more valuable than the answer itself—knowing "I don't know" is the start of intelligence
2. Dynamic resource allocation beats fixed caps—let the agent decide when it's done
3. Confidence is a universal language—can cross tasks, models, and scenarios

References

• BrowseConf: Confidence-Guided Test-Time Scaling for Web Agents — Ou et al., ACL Findings 2026.
• Inference-Time Scaling of Verification: Self-Evolving Deep Research Agents — Related work, rubric-guided verification.
• GenCluster: Scaling Test-Time Compute to Achieve IOI Gold Medal — Parallel TTS approach.
• evaluation-stc-challenges — Previous article: evaluation challenges.