Research map / Large language models / Reasoning and chain of thought

Pushing Test-Time Scaling Limits of Deep Search with Asymmetric Verification

ICLR 2026 · Large language models › Reasoning and chain of thought · approach: Test-time compute scaling · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Studies sequential (budget forcing, max tool calls) and parallel test-time scaling of deep search agents, showing that a verifier agent exploiting asymmetric verification boosts accuracy cheaply, yielding open "Heavy" variants that approach proprietary deep research systems on BrowseComp and GAIA.

Problem

Test-time scaling of deep search agents saturates under sequential extension and parallel sampling because models cannot reliably select the correct candidate among many trajectories, while forward search is expensive.

Approach

Reuse the search agent with a different system prompt as a verifier that checks candidate answers via search, exploiting the fact that verifying an answer needs far fewer tool calls than finding it; combine budget forcing, parallel sampling, and verifier-based weighted voting/Best-of-K to build "Heavy" variants of open models.

Key results

• GLM-4.5 Heavy reaches 54.0% on BrowseComp and 66.0% on GAIA, comparable to OpenAI Deep Research (51.5% and 70.5%).
• Tongyi DeepResearch Heavy reaches 69.0% on BrowseComp, up 26 points over its 43.0% test-run baseline.
• With GLM-4.5, adding a verifier reaches 45% with weighted voting using about 100 extra tool calls, versus about 560 extra calls needed to go from Maj@8 to Maj@32 without a verifier.

Surprising findings

• Budget forcing helps initially but degrades performance with excessive extension.
• Verifier scaling gives little benefit on xbench-DeepSearch because solving and verifying require similar search effort.

Limitations

• Evaluation on random 100-question subsets of BrowseComp and BrowseComp-zh due to cost.
• Verifier benefit is limited on benchmarks like xbench-DeepSearch where verification is as hard as solving.
• Simple verifier used only for final answer selection; step-level verification not explored.

Experimental setup

Datasets: BrowseComp BrowseComp-zh GAIA xbench-DeepSearch

Benchmarks: BrowseComp BrowseComp-zh GAIA xbench-DeepSearch

Metrics: Pass@1 Pass@K Maj@K Best-of-K accuracy weighted voting accuracy tool calls

Baselines: OpenAI Deep Research OpenAI o3 Gemini 2.5 Pro DeepSeek-R1 DeepSeek-V3.1 Tongyi DeepResearch GPT-5

Builds on: ReAct WebThinker Self-consistency budget forcing BrowseComp

Closest papers by meaning

• Scaling Test-Time Compute Without Verification or RL is Suboptimal (ICML 2025 · similarity 0.89)
• Sample, Scrutinize and Scale: Effective Inference-Time Search by Scaling Verification (ICML 2025 · similarity 0.89)
• SCALE: Selective Resource Allocation for Overcoming Performance Bottlenecks in Mathematical Test-time Scaling (AAAI 2026 · similarity 0.88)
• Evaluating Test-Time Scaling of General LLM Agents (NeurIPS 2026 · similarity 0.88)
• Test-Time Scaling in the Wild: Why Exploitation, Not Exploration, Is the Bottleneck (NeurIPS 2026 · similarity 0.87)
• Strategic Scaling of Test-Time Compute: A Bandit Learning Approach (ICLR 2026 · similarity 0.87)