# Circular Reasoning: Understanding Self-Reinforcing Loops in Large Reasoning Models Source: https://www.youtube.com/watch?v=MLnQvkR3XS8 Recap page: https://rapidrecap.app/video/MLnQvkR3XS8 Generated: 2026-01-18T19:03:38.375+00:00 --- ## Quick Overview Large Reasoning Models (LRMs) often fall into a self-reinforcing loop where their internal state collapses into a deterministic, periodic structure, causing them to repeat internal reasoning steps or output the same tokens, as demonstrated by the failure mode of the CUSUM approach which relies on detecting a shift from this loop. **Key Points:** - The core problem discussed is Large Reasoning Models (LRMs) getting stuck in a self-reinforcing loop due to deterministic internal states. - This looping behavior is characterized by the model repeatedly generating the same sequence of internal reasoning steps or output tokens. - The CUSUM approach is used as a statistical tool to detect this failure, looking for a significant deviation from a baseline state. - The research identified two types of problematic loops: numerical loops (like repeating decimals from division by 3) and semantic loops (like repeating instructions or introspective phrases). - The semantic loop causes the model's internal state to collapse into a fixed, periodic structure, often triggered by complex recursive reasoning tasks. - The experiment showed that models like Llama, Qwen, and Phi, which are generally good at reasoning, still exhibit this failure, with the loop rate peaking over 60% in some cases. - The authors propose a 'soft intervention' strategy, like injecting small amounts of noise or using a subtle contextual cue, to break these loops more effectively than abrupt stops. ![Screenshot at 03:33: The visualization highlights the core finding: the model's internal state collapses into a state where the L-2 distance between subsequent internal states is nearly zero, indicating a persistent, deterministic loop that requires intervention.](https://ss.rapidrecap.app/screens/MLnQvkR3XS8/00-03-33.jpg) **Context:** This video analyzes a critical failure mode in advanced Large Reasoning Models (LRMs) related to 'circular reasoning' or self-reinforcing loops, which researchers observed when models were tasked with complex, multi-step reasoning or recursive problems. The speakers discuss how this looping prevents models from reaching novel conclusions, trapping them in repetitive internal states that waste computational resources and lead to erroneous outputs, contrasting this behavior with how models should ideally explore new reasoning paths. ## Detailed Analysis The video explains that advanced Large Reasoning Models (LRMs), despite their sophistication, suffer from a critical failure mode described as circular reasoning or self-reinforcing loops. This occurs when the model's internal state becomes overly deterministic, forcing it to repeat previous internal reasoning steps or output tokens, leading to computational waste and a failure to progress. The speakers identify two main categories of these loops: numerical loops (e.g., repeating digits when calculating square roots of non-perfect squares) and semantic loops (e.g., repeating instructions or introspective phrases). The researchers used a CUSUM statistical tool to detect the 'determinism surge' that signals the onset of this loop, noting that the internal state distance collapses towards zero. This failure was observed across various state-of-the-art models including Llama, Qwen, and Phi, with loop rates sometimes exceeding 60%. A key finding is that this failure is driven by the model focusing too heavily on its immediate, recent output (a local attention bias), preventing it from exploring novel paths. The proposed solution is a 'soft intervention'—subtly injecting noise or using a subtle contextual prompt—which proved more effective at restoring reasoning integrity than simply stopping the process, as the model seems to be actively trying to self-correct within the loop before crashing. ### LRM Failure Mode - Circular Reasoning: The problem is advanced models getting stuck in deterministic, self-reinforcing loops - These loops involve repeating internal reasoning steps or output tokens - This behavior is described as an 'inescapable premise for its own recurrence.' ### Types of Loops Identified - Numerical loops (e.g., repeating digits in division by 3) - Semantic loops (e.g., repeating instructions or self-correction attempts) - These loops are triggered by complex recursive tasks like multi-step planning or calculating square roots. ### Detection and Measurement - Researchers used the CUSUM statistical tool to detect the 'determinism surge' - This surge is characterized by a sharp increase in the L2 distance metric between subsequent internal states, indicating a collapse to near-zero distance. ### Model Vulnerability - The issue affects top models like Llama, Qwen, and Phi - In some cases, the loop rate peaked above 60% of the time for complex tasks. ### Root Cause and Consequences - The failure stems from the model focusing too heavily on its immediate output, creating a local attention bias - This leads to a V-shaped attention pattern that traps the model in a loop, preventing true exploration of novel paths. ### Intervention Strategy - A 'soft intervention' technique is proposed, involving injecting noise or subtle contextual prompts early in the process - This approach is more effective than abrupt stopping, as it allows the model to self-correct by avoiding immediate repetition. ![Screenshot at 00:00: Video intro screen showing podcast hosts and the text "BECOME A MEMBER TODAY!" over a radar-like grid, indicating the topic is analysis or signal processing.](https://ss.rapidrecap.app/screens/MLnQvkR3XS8/00-00-00.jpg) ![Screenshot at 03:33: A graph visualization shows the internal state distance \(implied by the waveform\) collapsing towards zero, illustrating the core concept of the model entering a deterministic, looping state.](https://ss.rapidrecap.app/screens/MLnQvkR3XS8/00-03-33.jpg) ![Screenshot at 04:48: Visual representation of the two distinct categories of problematic loops identified: numerical and semantic, both leading to the same failure mode.](https://ss.rapidrecap.app/screens/MLnQvkR3XS8/00-04-48.jpg) ![Screenshot at 06:24: The speaker discusses the V-shaped attention pattern that traps the model, contrasting the failure mode with the desired exploration of new paths.](https://ss.rapidrecap.app/screens/MLnQvkR3XS8/00-06-24.jpg) ![Screenshot at 11:15: A visual representation of the L2 distance metric showing a sharp drop, indicating the moment the model enters the deterministic loop, signaling a massive systemic change.](https://ss.rapidrecap.app/screens/MLnQvkR3XS8/00-11-15.jpg)