The Illusion of Thinking; Apple's Latest Paper Exposes LLM "Reasoning" Limits

Just finished diving into Apple’s insightful new paper, “The Illusion of Thinking”, and it’s given me a lot to chew on regarding the current state of AI “reasoning.” 🧠

They took a clever approach: instead of relying solely on standard math benchmarks (which can have data contamination issues), Apple researchers tested leading “Large Reasoning Models” (LRMs) like OpenAI’s o1/o3-mini, DeepSeek-R1, and Claude 3.7 Sonnet Thinking using controllable puzzles – think Tower of Hanoi, River Crossing, and Blocks World. They systematically cranked up the complexity to see how these models truly “think.”

What they found is quite revealing:

1️⃣ The “Complexity Cliff”: As puzzles got harder, the models didn’t just stumble – their accuracy completely collapsed to ZERO beyond a certain point. Not a gradual decline, but a full stop.

2️⃣ The “Effort Paradox”: Here’s a shocker – when problems became too complex, the models actually reduced their “reasoning effort” (spent fewer tokens thinking), despite having ample token budget. It’s like they give up rather than dig deeper.

3️⃣ Three Distinct Performance Regimes Emerged:

• Low Complexity: Standard LLMs often performed better and more efficiently. The LRMs tended to “overthink.”
• Medium Complexity: This is where LRMs showed a clear advantage, their extended “thinking” (e.g., Chain-of-Thought) paid off.
• High Complexity: Both model types ultimately failed. The extra “thinking” in LRMs only delayed the inevitable collapse.

4️⃣ Trouble with Explicit Instructions: Even when provided with the exact algorithm to solve a puzzle, the models still hit the same performance wall. This suggests they aren’t truly executing logical steps but are still fundamentally predicting sequences.

5️⃣ Spotty “Reasoning” Prowess: A model might ace a complex Tower of Hanoi (requiring many correct moves) yet fail a simpler River Crossing puzzle that needs far fewer. This points more towards learned patterns for specific scenarios rather than generalized reasoning ability.

My Main Takeaway

Apple’s research strongly suggests that what many call AI “reasoning” today might be a very advanced “Illusion of Thinking.” These models are master pattern-matchers and sequence completers, but the kind of robust, generalizable, logical problem-solving we associate with human intelligence is still a frontier.

A fascinating (and perhaps sobering) perspective for those of us actively working in the AGI space. It seems true, adaptable reasoning remains a significant challenge.

What are your thoughts?

---

#AI #LLMs #ArtificialIntelligence #MachineLearning #Reasoning #AppleResearch #TechInnovation #DeepLearning #AGI #Apple #Walmart #WalmartGlobalTech #GenAI #GenerativeAI #LLM