Research Note
Research Note: Illusion of Thinking Experiment and Response Matrix
The current paper evaluates four puzzle families: Tower of Hanoi, Checker Jumping, River Crossing, and Blocks World. Complexity changes through disk count, checker count,
In this article
Research Note: Illusion of Thinking Experiment and Response Matrix
Original experiment
The current paper evaluates four puzzle families: Tower of Hanoi, Checker Jumping, River Crossing, and Blocks World. Complexity changes through disk count, checker count, actor-agent pairs, or block count while the basic rules remain stable.
Most matched comparisons use Claude 3.7 Sonnet with and without extended thinking and DeepSeek R1 versus DeepSeek V3. Final-accuracy experiments also include o3-mini. The paper reports 25 samples per model, puzzle instance, and complexity level, subject to response-format filtering. Claude and DeepSeek runs were allowed generation lengths up to 64,000 tokens.
The simulator checks whether a proposed move is valid and whether the goal state is reached. The central success measure is a completely correct solution, not whether the model knows a compact algorithm or makes many valid moves before one error.
Claim-to-evidence matrix
| Question | Original result | Later qualification | Defensible use |
|---|---|---|---|
| Do reported models fail at higher puzzle complexity? | Accuracy approaches zero in the reported configurations | Yes, though thresholds and causes depend on task and representation | Report a configuration-specific failure |
| Does thinking always help? | No. Non-thinking pairs can lead at low complexity | Not materially disputed as a universal rule | Reject "more thinking is always better" |
| Is the Tower of Hanoi collapse only a token-limit problem? | Authors say early errors occur before the limit | Lawsen shows compact code can restore performance; later replication still finds failures around eight disks | Separate algorithm generation from long sequential execution |
| Were all River Crossing cases valid tests? | Version 1 included larger fixed-capacity cases | Later comments identify unsolvable configurations; version 3 narrows focus | Never treat impossible cases as model failures |
| Do tools invalidate the paper? | Authors intentionally test model reasoning without tools | Agentic comments show tools can change end-to-end outcomes | Model-only and system evaluations answer different questions |
| Did the paper prove that AI cannot reason? | No operational definition or universal population supports that claim | Later papers also avoid a settled universal answer | Describe the tested behavior, not machine cognition |
Result, explanation, and implication
The result is what happened under the test. The explanation is why it happened. The implication is what another team should change. These must remain separate.
For Tower of Hanoi, a result can be "the model's enumerated move sequence became invalid." Possible explanations include state-tracking failure, accumulated per-move error, representation choice, early stopping, or some combination. Product implications depend on whether the product asks the model to enumerate moves, generate executable code, call a solver, verify state after each action, or hand the task to a person.
For River Crossing, an unsolvable instance is a task-construction failure. Removing it does not prove the model can solve every valid instance. It means that valid and invalid configurations must be analyzed separately.
Synthesis
The paper remains valuable because it makes complexity adjustable and inspects more than the final answer. The responses remain valuable because they test whether the result survives changes to representation, solvability, tools, and interaction.
The strongest conclusion is not that one side won. It is that evaluation configuration is part of the claim. A reader should be able to reconstruct the model, version, task generator, prompt, tool access, budget, sample count, filtering, grader, success rule, and date before applying the result.
Sources
Follow the evidence.
- machinelearning.apple.com: illusion of thinkingmachinelearning.apple.com
- NIST AI RMF Measure guidanceairc.nist.gov
- youtu.be: 2unoT550UWAyoutu.be
- arxiv.org: 2507arxiv.org
- openai.com: a practical guide to building ai agentsopenai.com
- arxiv.org: 2506arxiv.org
- anthropic.com: demystifying evals for ai agentsanthropic.com
- arxiv.org: 2506arxiv.org
- open.spotify.com: 23EBn3y6n59SKM2C6T2lDsopen.spotify.com
- arxiv.org: 2506arxiv.org
- arxiv.org: 2506arxiv.org
- daltonanderson.ghost.io: apples ai strategy the flawed illusion of thinkingdaltonanderson.ghost.io
- anthropic.com: building effective agentsanthropic.com