Reasoning Models and RL Research: Test-Time Compute — 2026-10-29

Top developments

OpenAI models generate novel scientific proofs

On October 26, 2026, OpenAI announced that its internal models produced three research results in mathematics, computer science, and physics, including a disproof of the Exponential Time Hypothesis. This demonstrates the practical application of advanced reasoning capabilities and test-time compute, moving beyond benchmark scores to genuine discovery in specialized scientific domains.

1 sources

unite.ai

unite.ai

Artificial Analysis v5 widens proprietary-open gap

The release of Artificial Analysis Intelligence Index v5 has shifted 60% of its composite score to private test sets, widening the gap between proprietary and open-weight AI models from 3 points to 12. This overhaul emphasizes agentic AI benchmarks, highlighting that proprietary models maintain a significant edge in complex, multi-step reasoning tasks where test-time scaling is most effective. Costs among the top models vary 13-fold, underscoring the economic trade-offs of high-performance inference.

Kimi K3.1 Pro launches with 3T-class capabilities

Kimi released K3.1 Pro on October 29, 2026, a 3T-class open-weight model featuring a 1M-token context window. The model claims major advances in coding, 3D modeling, and financial research, positioning itself as a direct competitor to closed models at a fraction of the cost. This release underscores the trend of open-weight models catching up in reasoning capabilities, though proprietary models still lead in complex agentic tasks.

DeepSeek highlights inference compute dominance

In a recent YC Paper Club discussion, DeepSeek's published data revealed that inference compute is 12 times higher than training compute, challenging the common assumption that training is the primary bottleneck. This finding reinforces the importance of test-time compute scaling and efficient inference architectures for reasoning models, as the cost and resource allocation for "thinking" at inference time significantly outweighs training expenses.

Local view

According to Chinese media outlet The Paper (thepaper.cn), DeepSeek's internal data showing that inference compute is 12 times higher than training compute has sparked a debate on the true cost drivers in the reasoning model era. The discussion, led by YC Paper Club, suggests that optimizing for test-time efficiency is becoming more critical than training scale for many applications.

Additionally, Zhihu reports that domestic models like Kimi-K3.1 Pro and MiniMax-V3.1-Flash are updating their agent models this week, reflecting the intense local competition in open-weight reasoning capabilities.

Context & numbers

• Cost Variance: Among the top reasoning models, costs vary 13-fold, with some frontier reasoning tiers running up to $150 per million input tokens and $600 per million output tokens.
• Benchmark Gap: The proprietary vs. open-weight model gap in agentic benchmarks has widened to 12 points in the new Artificial Analysis Index.
• Compute Ratio: DeepSeek data indicates inference compute is 12x training compute.

On the radar

• Model-Specific Reasoning Bias: New research highlights that reasoning models exhibit model-specific biases, and sparse attention mechanisms are becoming cheaper with shared pools, impacting system reliability.
• RL Framework Updates: Open-source RL frameworks like OpenRLHF and verl continue to evolve, with verl releasing verl-vla v0.1.0 for unified VLA post-training, supporting human-in-the-loop data collection and reinforcement learning across simulators and real robots.