Source item

Interesting paper out of Apple research on the reasoning limitations of LLM models, especially for complex problems. The paper is short and consumable. 🧵https://t.co/x5Ypj4VZFa Here is a ChatGPT summary of the paper: The Illusion of Thinking: Understanding the Strengths and Limitations of Reasoning Models via the Lens of Problem Complexity (Apple ML Research) 🧠 What the paper is about Apple introduces Large Reasoning Models (LRMs)—language models that generate internal “thinking” as a way to solve problems. (https://t.co/M9tLVW573U) It notes that typical benchmarks in math and coding can be contaminated (i.e. models have seen solutions before), and hide what’s happening inside. 🔍 How they tested Designed four controlled puzzle environments—Tower of Hanoi, Checker Jumping, River Crossing, Blocks World—with precisely adjustable complexity but consistent logic. (https://t.co/jIsLb2oqa2) This let them inspect not just answers, but the entire reasoning trace. 📉 Main findings Three performance regimes: Low complexity: Standard LLMs outperform the “thinking” models (they “overthink” and get tangled). Medium complexity: LRMs gain an edge—thinking helps. High complexity: Both collapse completely. (https://t.co/iY0SNiZjnT) Reasoning collapse paradox: As problems get harder, reasoning effort (token usage) rises then falls, even with tokens to spare—LRMs simply give up. (https://t.co/jIsLb2oqa2) Explicit algorithm failure: Giving the full Tower of Hanoi algorithm to LRMs didn’t help—they still failed on harder instances. 🧩 Analysis of reasoning behavior In easy puzzles, correct answers appear early; LRMs then continue generating token-by-token, often looping into incorrect reasoning. In mid-level tasks, LRMs explore and eventually converge—this is the “sweet spot.” In complex tasks, they fail early and stop thinking, not because they ran out of tokens, but because they hit a threshold and abandoned the process. 🤔 Implications Illusion vs reality: LRMs mimic thinking—they produce “reasoning” text—but underlyingly it’s advanced pattern matching, not true algorithmic reasoning. (https://t.co/vHMEdQ9YEU) Benchmark caution: Chain-of-Thought and similar methods can help—but they hide scaling failure when problems get hard. Need for new architectures: These failures suggest a structural boundary—more tokens alone won’t fix compositional reasoning breakdowns. Engineers take note: For real-world applications, test your models across complexity regimes to identify where they stop “thinking.” 🍵 Summary Strengths: Pulse-checks show LRMs work well from moderate complexity; Chain‑of‑Thought has value there. Weaknesses: In easy cases they “overthink,” in hard cases they collapse—sometimes even when fed algorithms. Conclusion: This is a reality check—LRMs don’t truly reason, and we’re seeing fundamental computational limits. The so-called “illusion of thinking” is real. Would be interesting to pair this with insights from Estermann & Wattenhofer (2025), who also observed that reasoning effort rises then stagnates beyond complexity thresholds (https://t.co/bzbF7tgwro, https://t.co/iY0SNiZjnT, https://t.co/76K5tKrkHh). But Apple’s controlled puzzles nail down the phenomenon cleanly. On one hand, this may be breakthrough research & challenges the common Silicon Valley wisdom that these models will scale ad infinitum. If they uncovered a fundamental blocker - its important. Would love to hear the counter analysis from one of the large model providers? With that said, I don't think its a good thing that Apple researchers tilt to the skeptical. How can you innovate if the core thinkers/leaders in your organization are non-believers? Might explain why Apple may be failing to lead in this important area.