Apple's 'Illusion of Thinking' — what's left of it one year later?
In June 2025, Apple Machine Learning Research published the paper The Illusion of Thinking. The thesis: the then-new reasoning models such as o3-mini and DeepSeek-R1 do not actually reason — they collapse to zero accuracy as problem complexity rises, and their “reasoning effort” even decreases once tasks become too hard. The paper triggered one of the fiercest methodology debates of the AI year. Almost a year later, a sober look back is worthwhile: what held up, what was refuted, and what does it mean for practice?
Apple's core findings from 2025
- Three complexity regimes: at low complexity, standard LLMs beat reasoning models. At medium complexity, explicit “thinking” won. At high complexity, all models fell to zero.
- Inverse scaling effect: reasoning models put in less effort the harder the task became — even with token budget available.
- Algorithm execution fails: even when the correct solution algorithm was given explicitly in the prompt, models failed at executing it.
- Test bench: Tower of Hanoi, Checker Jumping, River Crossing, Blocks World — four puzzle classes with controllable complexity.
What was true before
The reasoning-model wave in early 2025 was marked by strong optimism: o1, then o3-mini, then DeepSeek-R1 hit unprecedented numbers on math and coding benchmarks. The common reading: longer “thinking” (chain-of-thought, self-reflection, tree search) qualitatively produces real step-by-step inference. Apple’s paper went against that grain: the gain, they argued, was not reasoning but better pattern matching within the training distribution.
What has been clarified since
1. The methodology critique was partly justified. A few weeks after publication, the follow-up paper The Illusion of the Illusion of Thinking appeared — and it landed. Apple had evaluated Tower of Hanoi at high disk counts without checking whether the correct solution even fit into the available token budget. At 15 disks, the optimal solution requires more than 30,000 move steps; the model simply cannot write them out. What Apple interpreted as “reasoning collapse” was, at least in part, a token-limit artifact.
2. The core finding survived nonetheless. Even after correcting the token-limit problem, the result holds: reasoning models generalize worse to structurally novel problems than their benchmark numbers suggest. Several follow-up studies from autumn 2025 and spring 2026 confirm that controlled complexity scaling — where the token budget grows with the problem — still drops accuracy markedly faster than on training-in-distribution tasks.
3. The industry response was test-time search. The industry’s answer to the critique was not “less reasoning” but “more test-time compute”. o3 (full version), Claude Opus 4.7 with extended thinking, and DeepSeek-R2 lean heavily on inference-time search strategies. That pushes the complexity ceiling higher — but does not remove it. Apple’s basic critique has thus been domesticated rather than refuted.
Why it matters
What can we say with confidence today? Three things:
First — benchmarks are unreliable proxies. AIME, MATH, GPQA and the like test problems that sit close to training data structurally. A model can score 95 percent there and drop to 20 percent on a structurally shifted puzzle of identical formal complexity. Anyone deploying models for productive reasoning tasks should run their own out-of-distribution tests. Apple’s main contribution is methodological: they showed cleanly how to measure such things.
Second — the “longer answer = more reasoning” confusion persists. In fact, with current reasoning models, answer length often correlates inversely with accuracy on hard tasks. Seeing a reasoning output of 8,000 tokens should not automatically be read as a quality signal — it is often a signal that the model is stuck.
Third — compute leverage shifts the problem. With test-time search, parallel rollouts, and external verification, today’s stack solves tasks that 2024 models with reasoning attachments failed at. That is real. But it is not the same as a qualitative jump in inference capacity per token. If you need token-efficient reasoning for your deployment, keep that in mind.
What you can do now
If you deploy reasoning models productively: define at least one out-of-distribution test set for your domain and measure with it. Vendor benchmarks are useful as marketing material, often less so as selection aid for your specific use case.
If you compare models: look not only at benchmark score but also at tokens consumed per solved task and answer length on failed tasks. The latter is a good indicator of whether the model is stuck in a loop or productively searching.
If you have not read the paper yet: it is still worth reading a year later — precisely because the discussion around it was so productive. Both sides of the debate (Apple’s original and the “Illusion of the Illusion” counter-paper) can be worked through in an afternoon and sharpen one’s view of benchmark results considerably.
Entdecke mehr
AI Workflows by Keyword: How We Make Recurring Routines Enforceable
A typed keyword triggers a fixed AI routine — and every single step must be committed before the next one appears. Why that's the actual trick.
GlossarRecall
Recall (hit rate, sensitivity) measures the share of all actually existing relevant cases that a system finds. Formula: relevant cases found divided by all relevant cases that truly exist. Requires a known ground truth.
LexikonMeasuring LLM Quality — Evals, Benchmarks, Judges
How to tell whether an LLM system actually works: three evaluation layers from benchmarks to CI evals, plus pitfalls like Goodhart and judge bias.