aceperformance of Large Reasoning Models on controllable puzzles of varying compositional complexity
Source article: The Illusion of Thinking
Recent generations of frontier language models have introduced Large Reasoning Models (LRMs) that generate detailed thinking processes before providing answers. While these models demonstrate improved performance on reasoning benchmarks, their fundamental capabilities, scaling properties, and limitations remain insufficiently understood. Current evaluations primarily focus on established mathematical and coding benchmarks, emphasizing final answer accuracy. However, this evaluation paradigm often suffers from da…
Contested: both sides are scored from claims and sources, not community votes.
Wetland History Research Paper - DPLA - 588a00bebde880221a3b28cf955e5433 by Kucich, Alex; Dewyngaert, Scarlett; Owens, Meghan. CC BY 4.0 · https://creativecommons.org/licenses/by/4.0
Published September 23, 2025, this peer-reviewed study systematically tested frontier Large Reasoning Models that generate detailed thinking processes before answering. Using controllable puzzle environments to vary compositional complexity, the authors analyzed final accuracy and internal reasoning traces and compared LRMs to standard LLMs under equivalent inference compute.
The work matters because it moves evaluation beyond final-answer accuracy on contaminated math and coding benchmarks to trace structure and scaling behavior. It remains uncertain how these puzzle-based collapse patterns and inconsistent algorithm use generalize to real-world scientific or applied reasoning workflows outside the controlled environments.
- Study uses controllable puzzle environments that allow precise manipulation of compositional complexity while maintaining consistent logical structures.
- Comparison under equivalent inference compute identifies three regimes: low-complexity where standard models outperform LRMs, medium-complexity where LRMs have advantage, high-complexity where both collapse.
- Analysis finds LRMs fail to use explicit algorithms and reason inconsistently across puzzles, with limitations in exact computation.
Large Reasoning Models generate detailed thinking traces before answering and demonstrate improved performance on reasoning benchmarks, with advantage over standard LLMs on medium-complexity controllable puzzles.
Frontier Large Reasoning Models face a complete accuracy collapse beyond certain puzzle complexities and exhibit a counterintuitive scaling limit where reasoning effort declines despite adequate token budget.
The rundown
The authors evaluated frontier LRMs using controllable puzzle environments that enable precise manipulation of compositional complexity while keeping logical structures consistent, allowing inspection of both final answers and internal reasoning traces.
Across puzzles they observed three regimes under equivalent inference compute: standard LLMs outperforming LRMs at low complexity, LRMs advantaged at medium complexity, and both collapsing at high complexity, with LRMs failing to use explicit algorithms and showing inconsistent reasoning.
Findings are bounded to controllable puzzle environments and reveal that LRMs have limitations in exact computation and inconsistent reasoning across puzzles.
Sources
- Peer-reviewedSuperIntelligence - Robotics - Safety & Alignment2025-09-23
How should this claim be treated?
The debate