LLM reasoning models may not actually reason ????

LLM reasoning models may not actually reason ???? Aka breaking down apples new paper that went viral - let's get into it! <3 Introduction The Illusion of Thinking: Understanding the Strengths and Limitations of Reasoning Models via the Lens of Problem Complexity Parshin Shojaee, Iman Mirzadeh, Keivan Alizadeh Maxwell Horton, Sany Beggi, Mahshad Poorgholami Apple Abstract Recent generations of frontier language models have introduced Large Reasoning Models (LLMs) that claim to demonstrate improved performance on reasoning benchmarks, their fundamental capabilities, scalability, and limitations remain poorly understood. This paper investigates the extent to which LLMs can demonstrate improved performance on reasoning benchmarks, emphasizing final answer accuracy. However, we also report on the limitations of these models, and the emergence of a reasoning effect paradox. We demonstrate that the reasoning effect paradox is a fundamental limitation of current LLMs, and that current approaches may be encountering fundamental barriers to generalizable reasoning capabilities. -Apple researchers systematically test reasoning models like deepseek r1, Claude sonnet 3.7 thinking w/ controllable puzzles (Reasoning models = LLM that generates long "thinking" traces supposedly leading to better reasoning than standard LLMs) -study uses tower of Hanoi, checker jumping, river crossing, blocks world; puzzles where complexity can be precisely controlled by parameters Why puzzles vs. math benchmarks?? -math benchmarks suffer from data contamination and don't allow complex control -solution = use four classic puzzles where difficulty scales predictably -tower of Hanoi = exponential, checkers = quadratic, river crossing + blocks world = linear ish Experimental Set Up -models tested: Claude 3.7 sonnet (thinking), deepseek r1, o3-mini variants provided with the same token budget, temp 1.0 (unclear why they picked this it's sort of odd to me), 25 samples per instance -custom extraction pipeline: regex for moved + tokenizer for position tracking -simulators validate each move and track first failure point Standard LLMs sometimes beat "thinking" models (!?) Figure 3: Illustration of the four puzzle environments. Columns show the progression from initial state (top) through intermediate state (middle) to target state (bottom) for puzzle: Tower of Hanoi, Checkers Jumping, River Crossing (transporting entities across a river), and Blocks World (black reconfiguration). 3. Puzzle Environments Our setup uses four controllable puzzle environments, each designed to test specific reasoning capabilities. The environments are: Tower of Hanoi, Checkers Jumping, River Crossing, and Blocks World. Each environment has a unique complexity profile, ranging from exponential to linear. The Tower of Hanoi is an exponential problem, requiring a minimum of 2^n-1 moves to solve, where n is the number of disks. Checkers Jumping is a quadratic problem, requiring a minimum of n^2 moves to solve. River Crossing and Blocks World are linear problems, requiring a minimum of n moves to solve. The environments are designed to be precisely controllable by parameters, allowing us to systematically vary the difficulty of each puzzle. This allows us to test the LLMs' ability to generalize their reasoning capabilities to novel and complex problems. -compared thinking vs. non thinking model pairs w/ equal compute -low complexity: standard LLMs more accurate and token efficient (?!) -medium complexity: LRMs show advantage from thinking -high complexity: both collapse to 0% accuracy (complete failure) Reasoning effect paradox Figure 4. Accuracy of thinking models (Claude 3.7 Sonnet, Deepseek-R1) versus their non- thinking counterparts across all puzzle environments and problem complexities. 4 Experiments & Results 4.1 Experimental Setup Most of our experiments are conducted on reasoning models and their non-thinking counterparts. We use Claude 3.7 Sonnet (thinking and non-thinking), Deepseek-R1 (thinking and non-thinking), and o3-mini (thinking and non-thinking). We focus on the Tower of Hanoi, Checkers Jumping, River Crossing, and Blocks World environments. We also report on the limitations of these models, and the emergence of a reasoning effect paradox. We demonstrate that the reasoning effect paradox is a fundamental limitation of current LLMs, and that current approaches may be encountering fundamental barriers to generalizable reasoning capabilities. 4.2 Three Regimes of Complexity Motivated by the observations in Fig. 2, we systematically investigate the impact of problem complexity on reasoning performance. We identify three regimes of complexity: low, medium, and high. In the low complexity regime, standard LLMs outperform "thinking" models. This is likely because the "thinking" models introduce additional overhead and complexity that is not necessary for simple problems. In the medium complexity regime, "thinking" models show an advantage. This is likely because the "thinking" models are better able to handle the increased complexity and explore the problem space more effectively. In the high complexity regime, both "thinking" and standard LLMs collapse to 0% accuracy. This is likely because the problems are too complex for current LLMs to solve, regardless of whether they employ "thinking" or not. This suggests that there may be fundamental limitations to the reasoning capabilities of current LLMs, and that new approaches may be needed to overcome these limitations. -measured thinking tokens vs. problem complexity; initially scales up with harder problems -but near collapse point, models REDUCE thinking despite available budget -models "give up" before hitting limits; fundamental scaling limitation discovered A look inside the models' thought processes Figure 6. Accuracy and thinking tokens vs. problem complexity for reasoning models across puzzle environments and experimental setups. The graphs show that accuracy increases with thinking tokens up to a certain point, after which it begins to decline. This suggests that there is an optimal amount of thinking for each problem, and that too much thinking can be detrimental. The graphs also show that the models' performance collapses to 0% accuracy at high complexity, regardless of the amount of thinking. This suggests that there may be fundamental limitations to the reasoning capabilities of current LLMs, and that new approaches may be needed to overcome these limitations. -extracted immediate solutions from thinking traces for more insight in the literature, lead to the waste of compute. As problems become moderately more complex, -low complexity: find solution early then overthink with wrong paths -medium: correct solutions emerge after extensive exploration -high: never find correct solutions throughout entire trace Additional testing with Tower of Hanoi Figure 8 (a) & (b) Despite providing the solution algorithm in the prompt, execution failure occurs at similar points, highlighting reasoning model limitations in logical step execution. (c) & (d) The reasoning models fail to generalize to new environments, suggesting that their reasoning is not generalizable. The Tower of Hanoi accuracy is compared to earlier errors in the River Crossing scenario. 5 Conclusion Our research rigorously examines frontier Large Reasoning Models (LLMs) through the lens of problem complexity using controllable puzzle environments. Our findings reveal fundamental limitations in current LLMs' reasoning capabilities. Specifically, we found that LLMs struggle to develop generalizable reasoning capabilities beyond certain complexity thresholds. We identified three regimes of complexity: low, medium, and high. In the low complexity regime, standard LLMs outperform "thinking" models. In the medium complexity regime, "thinking" models show an advantage. In the high complexity regime, both "thinking" and standard LLMs collapse to 0% accuracy. We also found that LLMs exhibit a "reasoning effect paradox," where they reduce thinking tokens as problem complexity increases, despite available budget. This suggests that current approaches may be encountering fundamental barriers to generalizable reasoning capabilities, and that new approaches may be needed to overcome these limitations. Most notably, we observed their limitations in performing exact computation, for example, -provided explicit Tower of Hanoi algorithm in prompt -performance did NOT improve; same collapse point (!!) -models fail at logical verification and step execution, not just strategy -can't follow given instructions = not true reasoning Cross puzzle inconsistencies Figure 6. Accuracy and thinking tokens vs. problem complexity for reasoning models across puzzle environments and experimental setups. The graphs show that accuracy increases with thinking tokens up to a certain point, after which it begins to decline. This suggests that there is an optimal amount of thinking for each problem, and that too much thinking can be detrimental. The graphs also show that the models' performance collapses to 0% accuracy at high complexity, regardless of the amount of thinking. This suggests that there may be fundamental limitations to the reasoning capabilities of current LLMs, and that new approaches may be needed to overcome these limitations. -tower of Hanoi = models achieve >50% accuracy w/ 8 disks; river crossing = complete fail w 0% on just 3 pairs -tower of Hanoi actually repeats recursive pattern over and over vs. river crossing requires reasoning abt dynamic constraints -may suggest models can follow memorized patterns but fail novel constraint logic Tbh idk how I feel abt this paper; I honestly don't think it's that amazing and their conclusions are a bit off imo I think pattern matching is a form of cognition and computational processing humans do to reason I'm also curious to test my own architecture on their puzzles! So stay tuned for that <33 Okie lmk ur thoughts and have an amazing day ~ bye bye <333