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

LLMs Improving LLMs: Agentic Discovery for Test-Time Scaling

NeurIPS 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.

Proposes AutoTTS, an environment-driven framework where an LLM explorer (Claude Code) discovers test-time scaling controllers via code search over offline replayed reasoning trajectories, using a single-scalar beta parameterization and execution-trace feedback.

Problem

Test-time scaling strategies are hand-crafted heuristics for branching, probing, pruning and stopping; the computation-allocation space is under-explored and online evaluation of candidate strategies is expensive.

Approach

Formulate width-depth TTS as controller synthesis over pre-collected reasoning trajectories and probe signals (offline replay MDP), so candidate controllers are evaluated cheaply without LLM calls; constrain each controller to one scalar beta mapped monotonically to all hyperparameters, and feed full execution traces back to the explorer LLM.

Key results

• On held-out AIME25/HMMT25 averaged over four Qwen3 models, AutoTTS (beta=0.5) matches SC@64 accuracy (45.3 vs 45.2) while cutting tokens by about 69.5%.
• AutoTTS (beta=1.0) reaches the highest held-out accuracy in 5 of 8 cases versus handcrafted baselines.
• Full five-round discovery costs $39.9 and 160 minutes thanks to offline replay.

Surprising findings

• Removing beta parameterization cut tokens drastically (575.5K to 93.3K) but lowered held-out accuracy, indicating overfitting to the search set via extreme thresholds.
• On Qwen3-8B the discovered controller at beta=1.0 is essentially tied with SC@64 on average (62.7 vs 62.8).

Limitations

• Current instantiation only covers width-depth control with a fixed action set; richer control structures are untested.
• Discovery relies on a frontier coding agent (Claude Code); open-source agents are not evaluated.
• Evaluation is limited to math benchmarks plus one GPQA setting and one non-Qwen model.

Experimental setup

Datasets: AIME24 AIME25 HMMT25 GPQA-Diamond

Benchmarks: AIME24 AIME25 HMMT25 GPQA-Diamond

Metrics: accuracy total tokens search cost (USD) wall-clock time

Baselines: SC@64 ASC ESC Parallel-Probe

Builds on: Parallel-Probe Meta-Harness AlphaEvolve ADAS

Closest papers by meaning

• AgentTTS: Large Language Model Agent for Test-time Compute-optimal Scaling Strategy in Complex Tasks (NeurIPS 2025 · similarity 0.92)
• Evaluating Test-Time Scaling of General LLM Agents (NeurIPS 2026 · similarity 0.90)
• Every Rollout Counts: Optimal Resource Allocation for Efficient Test-Time Scaling (NeurIPS 2025 · similarity 0.90)
• ReasoningBank: Scaling Agent Self-Evolving with Reasoning Memory (ICLR 2026 · similarity 0.89)
• ExComm: Exploration-Stage Communication for Error-Resilient Agentic Test-Time Scaling (NeurIPS 2026 · similarity 0.89)
• Test-Time Adaptation for LLM Agents via Environment Interaction (ICLR 2026 · similarity 0.89)