Do reasoning models actually beat standard models on optimization?
Explores whether extended chain-of-thought in reasoning models delivers performance gains on constraint-satisfaction problems like power-grid optimization. Matters because reasoning models are treated as automatic upgrades, but the evidence may not support that claim.
Reasoning models have been treated as a generalized capability upgrade — more thinking tokens at test time, broadly better performance. On constraint-bound numerical optimization the upgrade does not materialize. Reasoning variants do not systematically outperform their non-reasoning counterparts on power-grid, financial-operations, or cyber-security feasibility problems. The longer trace does not become a longer iteration.
The reason this matters: extended chain-of-thought looks like it should help. The problem involves multi-step arithmetic, interacting constraints, and convergence-style reasoning — exactly the regime where "think more" is supposed to pay. The data say it does not. Whatever extended CoT is doing on these tasks, it is not running a Newton-Raphson iteration or a primal-dual update in latent space; it is producing more text without producing more computation.
This is consistent with a growing view that reasoning models excel where the bottleneck is exploration over reasoning paths (math contests, code, multi-hop QA) but stall where the bottleneck is numeric procedure. Constraint satisfaction over real physical systems is the latter. Adding chain length adds search over verbal restatements of the problem, not iterations of the algorithm that would solve it.
The implication for product: choosing "reasoning model" for an optimization-heavy workflow is not automatically the right call. The relevant decision is whether the bottleneck is verbal reasoning or numeric computation. If numeric, the cost-effective path is hand-off to a solver, not more thinking tokens.
Inquiring lines that read this note 64
This note is a source for these research framings, grouped by the broader line of inquiry each explores. Scan the bold lines of inquiry; follow any specific question forward.
How can identical external performance mask different internal representations?- Why do only two of fourteen models improve when problem constraints are removed?
- What benefits do open foundation models create that closed systems cannot?
- When does the right constraint beat additional model capacity?
- What production constraints should determine paradigm selection?
- Why do production systems optimize for three model classes instead of foundation models?
- Why do macro and micro forecasting scales require different reasoning approaches?
- What design changes could make constraint inference more reliable without explicit cuing?
- Can explicit optimal algorithms prevent reasoning model collapse at high complexity?
- What makes deterministic recursive reasoning models underperform on multi-solution tasks?
- Why does the second loop do most of the productive refinement work?
- Can explicit constraint statements override the dominance of surface heuristics?
- Do tool-enabled reasoning models close the gap on constraint satisfaction?
- Can external verifiers replace reasoning trace quality in solution guarantees?
- What makes constraint satisfaction problems epistemically cleaner than other reasoning tasks?
- Which constraint types do reasoning models handle best?
- Can the LLM-Modulo framework extend solver integration to domain planning?
- How do KV cache pruning and subproblem contraction both free reasoning capacity?
- Can symbolic solvers reliably replace LLM reasoning for logical tasks?
- What separates verifiable reasoning from open-ended judgment in scaling requirements?
- How does step-level compute allocation compare to response-level thinking?
- When is detailed step-by-step reasoning actually counterproductive for solving a problem?
- How does nesting optimization levels improve on traditional network depth?
- What makes multi-paradigm chaining a distinct reasoning topology?
- Can static reasoning patterns work better than dynamic branch selection?
- What advantages emerge from running 13 times more parallel reasoning chains with the same budget?
- Can parallel reasoning chains outperform longer sequential chains with the same compute?
- When does sequential reasoning provide exponential advantages over parallel voting?
- When does sequential chain-of-thought dramatically beat parallel voting approaches?
- Why do non-reasoning models work better under extreme decomposition than reasoning models?
- Is the reasoning cliff actually a tool-use problem?
- Do higher asymptote recipes unlock genuinely novel reasoning strategies?
- Why do reasoning models fail to improve constrained optimization performance?
- What limits external scaling when a model lacks reasoning foundation?
- Does iterative computation for reasoning transfer to environment dynamics modeling?
- Why does reapplying the same computation stages improve model performance?
- Does architectural design matter more than model scale for reasoning tasks?
- What planning strategies reduce execution steps without sacrificing solution quality?
- How does planning-before-execution compare to iterative reasoning and action loops?
- Can prompt engineering improve reasoning or only move requests into denser regions?
- Why does prompt optimization alone fail to inject genuinely new knowledge?
Related concepts in this collection 4
This note in its neighbourhood — explore the map, then jump to a related concept in the list below.
Click a node to walk · click center to open · click Open in graph to see this note in the full knowledge graph
-
Do larger language models solve constrained optimization better?
Explores whether scaling LLMs—through more parameters, better training, or reasoning extensions—improves their ability to satisfy constraints in real optimization problems like power grids and portfolios.
same paper, the parent finding
-
Do large language models actually perform iterative optimization?
Explores whether LLMs execute genuine numerical procedures like Newton-Raphson or instead pattern-match to memorized solution templates when solving constrained optimization problems.
same paper, the mechanism
-
Does more thinking time always improve reasoning accuracy?
Explores whether extending a model's thinking tokens linearly improves performance, or if there's a point beyond which additional reasoning becomes counterproductive.
adjacent: CoT ceilings in general
-
Why does chain of thought accuracy eventually decline with length?
Explores why longer reasoning chains don't always improve answers, and how the optimal length shifts based on task difficulty and model capability.
adjacent: more thinking is not monotonically better
Related papers in this collection 8
Papers most semantically related to this note, ranked by cosine similarity in the embedding space.
- Can Large Language Models Reason and Optimize Under Constraints?
- The Illusion of Thinking: Understanding the Strengths and Limitations of Reasoning Models via the Lens of Problem Complexity
- Chain of Thoughtlessness? An Analysis of CoT in Planning
- Comment on The Illusion of Thinking: Understanding the Strengths and Limitations of Reasoning Models via the Lens of Problem Complexity
- On the Reasoning Capacity of AI Models and How to Quantify It
- A Comment On "The Illusion of Thinking": Reframing the Reasoning Cliff as an Agentic Gap
- Does Reinforcement Learning Really Incentivize Reasoning Capacity in LLMs Beyond the Base Model?
- Divide-or-Conquer? Which Part Should You Distill Your LLM?
Original note title
reasoning models do not systematically outperform non-reasoning models on real numerical optimization — extended chain-of-thought is not a substitute for iterative computation