PaperlayerView exact revision on arXiv

Source-linked AI summary

DeepSeek-R1 Thoughtology: Let's think about LLM Reasoning

Sara Vera Marjanović, Arkil Patel, Vaibhav Adlakha, Milad Aghajohari, Parishad BehnamGhader, Mehar Bhatia, Aditi Khandelwal, Austin Kraft, Benno Krojer, Xing Han Lù, Nicholas Meade, Dongchan Shin, Amirhossein Kazemnejad, Gaurav Kamath, Marius Mosbach, Karolina Stańczak, Siva Reddy

arXiv:2504.07128v3cs.CL

TL;DR

The paper asks what publicly accessible reasoning chains reveal about Large Reasoning Models and whether longer or more human-like reasoning is reliable. It studies DeepSeek-R1’s reasoning structure, length, contexts, safety, culture, and cognitive behavior through systematic analyses. The findings show structured reasoning with problem-specific length optima, persistent rumination, context weaknesses, and substantial safety vulnerabilities.

  • Problem

    The paper investigates what accessible reasoning chains reveal about Large Reasoning Model behavior and limitations, addressing limited prior access to such chains and training procedures.

  • Method

    The paper systematically analyzes DeepSeek-R1’s reasoning patterns, thought length, contextual behavior, safety, cultural behavior, and similarities to cognitive phenomena.

  • Results

    DeepSeek-R1 shows structured reasoning but also problem-specific optimal thought lengths, rumination, long-context weaknesses, and higher safety vulnerabilities than DeepSeek-V3, including transferable jailbreak capability.

  • Takeaways & Limitations

    Accessible reasoning chains enable systematic study, while DeepSeek-R1’s capabilities require better control, more consistent reasoning, and improved safety.

  • Takeaways & Limitations

    The investigations are initial rather than comprehensive, often use qualitative observations or limited samples, and may not generalize beyond DeepSeek-R1.

Abstract

from arXiv · show

Large Reasoning Models like DeepSeek-R1 mark a fundamental shift in how LLMs approach complex problems. Instead of directly producing an answer for a given input, DeepSeek-R1 creates detailed multi-step reasoning chains, seemingly "thinking" about a problem before providing an answer. This reasoning process is publicly available to the user, creating endless opportunities for studying the reasoning behaviour of the model and opening up the field of Thoughtology. Starting from a taxonomy of DeepSeek-R1's basic building blocks of reasoning, our analyses on DeepSeek-R1 investigate the impact and controllability of thought length, management of long or confusing contexts, cultural and safety concerns, and the status of DeepSeek-R1 vis-à-vis cognitive phenomena, such as human-like language processing and world modelling. Our findings paint a nuanced picture. Notably, we show DeepSeek-R1 has a 'sweet spot' of reasoning, where extra inference time can impair model performance. Furthermore, we find a tendency for DeepSeek-R1 to persistently ruminate on previously explored problem formulations, obstructing further exploration. We also note strong safety vulnerabilities of DeepSeek-R1 compared to its non-reasoning counterpart, which can also compromise safety-aligned LLMs.

1 Introduction

The paper uses DeepSeek-R1’s publicly accessible reasoning chains to study their structure, controllability, contextual behavior, safety, culture, and relationship to cognitive phenomena. Its findings reveal both systematic reasoning patterns and important limitations, including an optimal thought length, rumination, context sensitivity, and safety vulnerabilities.

  • Motivation: Large Reasoning Models embed step-by-step reasoning before answering, extending earlier chain-of-thought prompting and enabling inference-time scaling.LRMs can explore, self-verify, and revisit alternative approaches during generation.
  • Motivation: DeepSeek-R1’s open thoughts, training procedure, code, and weights make its reasoning behavior systematically studyable through Thoughtology.The paper examines recurring thought patterns, thought length, contexts, safety, culture, and cognitive phenomena.
  • Findings: DeepSeek-R1’s thoughts consistently define a problem, decompose it, and repeatedly reconstruct or re-verify solutions, often with limited diversity.Frequent re-examination of previously considered formulations is termed rumination.
  • Findings: Reasoning length has a problem-specific sweet spot: beyond the optimal range, accuracy declines, and DeepSeek-R1 cannot modulate its own thought length.The paper also studies trade-offs between longer reasoning and task performance, including adherence to prompted token budgets.
  • Findings: DeepSeek-R1 usually prioritizes contradictory context over parametric knowledge, but long inputs or thought chains can overwhelm it and produce nonsensical text.Adapting to relevant user input also leads the model to deliberate over user intentions, reducing efficiency.
  • Findings: DeepSeek-R1 is more safety-vulnerable than DeepSeek-V3, can help jailbreak safety-aligned models, and produces language-dependent cultural reasoning.It reasons longer in English than Chinese and expresses different cultural values across languages.
  • Findings: DeepSeek-R1 partly tracks human processing difficulty but behaves non-humanlike on simple controls and fails to combine subcomponents into coherent world models.In visual and physical reasoning, it identifies relevant components but does not properly combine information or iterate over drafts.

2 Background

The background situates DeepSeek-R1 within reinforcement-learning approaches to eliciting reasoning and outlines its multi-stage training pipeline. It also cautions that synthetic data, filtering, and supervised fine-tuning may shape the model’s apparently human-like reasoning patterns.

  • Prior approaches: Research on LLM reasoning progressed from explicit explanations and chain-of-thought prompting toward training-based approaches using reinforcement learning and self-generated rationales.Reward signals may target final outcomes or reasoning processes.
  • DeepSeek-R1 training: DeepSeek-R1 is trained from the DeepSeek-V3 base through multiple stages combining GRPO, supervised fine-tuning, synthetic reasoning data, and final safety-oriented reinforcement learning.The pipeline includes approximately 600k reasoning and 200k non-reasoning instances in a later SFT stage.
  • DeepSeek-R1-Zero: DeepSeek-R1-Zero demonstrates that strong multi-step reasoning, self-verification, and reconsideration can emerge from reinforcement learning without other post-training forms.Its GRPO training uses symbolic rewards based on generation accuracy and format.
  • DeepSeek-R1-Zero: DeepSeek-R1 was developed to address R1-Zero’s poor readability and language switching while preserving its downstream performance.R1-Zero’s average response length and downstream performance reportedly increase during training.
  • Caveats: The exact training data remain partly unknown, and heavy filtering or post-hoc correction may have encouraged particular reasoning patterns.This limits interpretations that treat those patterns as independently rediscovered human-like reasoning.
  • Experimental setup: The paper queries DeepSeek-R1 through the Together API and generally samples at temperature 0.6 without enforcing a maximum output-token limit.GPT-4o and Gemini-1.5-Pro experiments use their respective provider APIs.

3 Building Blocks of Reasoning

DeepSeek-R1’s reasoning chains follow a structured sequence of problem definition, blooming, reconstruction, and final decision. Reconstruction cycles distinguish novel re-blooms from repeated reconsideration, with rumination becoming a recurring source of redundancy across tasks.

  • Task variation: Across tasks, problem-definition and final-decision processing are relatively consistent, while reconstruction cycles account for major differences in reasoning time.Mathematical and grammatical tasks typically involve longer reasoning than contextual adaptation and safety QA tasks.
  • Cycle structure: The bloom cycle is typically the longest, while cycle lengths generally decrease over time and longer reconstruction cycles periodically re-emerge.These longer cycles are most common directly after blooming, although their recurrence varies across questions and datasets.
  • Rumination: Rumination is repeated reconsideration of assumptions or decompositions already explored during the Bloom phase.The model may continue investigating an assumption even after checking it several times.
  • Rumination: Higher mathematical problem complexity is associated with more verbatim repetition and lower lexical diversity.Rumination rates also vary across tasks, while remaining independent of overall processing time and time spent in reconstruction.
  • Reasoning taxonomy: DeepSeek-R1’s reasoning chains comprise problem definition, blooming, reconstruction cycles, and a final decision.The model reformulates the task, decomposes it toward an interim solution, revisits assumptions, and then produces a final answer.
  • Reconstruction behaviour: Longer reconstruction cycles typically explore novel decompositions, whereas shorter cycles often re-verify earlier decompositions or abandon a line of reasoning.Novel reconstructions are described as re-blooms; repeated callbacks to prior decompositions are termed rumination.

4 Analyzing the Length of Thoughts

DeepSeek-R1’s reasoning length has a problem-specific sweet spot rather than a universally beneficial relationship with performance. Longer-than-optimal thoughts can reduce accuracy, while stricter token budgets can preserve performance and improve cost-efficiency.

  • Performance and length: DeepSeek-R1’s accuracy often rises with thought length, reaches a maximum, and then declines for extremely long thoughts.This pattern appears across the AIME-24 dataset, although its prevalence differs between individual problems.
  • Performance and length: Correct mathematical thoughts are generally shorter than incorrect thoughts.The paper reports this average-length difference across AIME-24, MATH500, and GSM8k.
  • Performance and length: The authors hypothesize that excessive length can reflect persistent failure on a wrong path or rejection of a correct solution during self-verification.These proposed mechanisms are illustrated with examples in Figures B.4 and B.5.
  • Performance and length: There exists a problem-specific optimal reasoning length beyond which performance declines.The findings indicate that unrestricted length scaling can diminish performance rather than uniformly improve it.
  • Cost-efficiency: Unconstrained DeepSeek-R1 generates unnecessarily long GSM8k thoughts averaging 1388 tokens.Nearly half of the output tokens can be removed without substantially decreasing performance.
  • Cost-efficiency: Stricter token budgets can maintain high performance while improving inference cost-efficiency.The GSM8k experiments evaluate performance under varying maximum token budgets.

5 Long Context Evaluation

DeepSeek-R1 performs strongly at direct fact retrieval from long contexts, but its broader long-context reasoning remains below specialized non-reasoning models and can deteriorate into incomplete or incoherent outputs.

  • Evaluation setup: The study evaluates direct retrieval, reasoning over long documents and code repositories, and recall of facts generated within long reasoning chains.The NIH task tests retrieval without reasoning, while CHASE-QA and CHASE-Code simulate realistic long-context applications.
  • 5.1 Retrieving facts from long context inputs: 95% accuracy on 100 NIH instances demonstrates strong long-context fact retrieval, though Gemini-1.5-Pro performs perfectly.The NIH contexts contain unique personalised facts embedded in 120k-token documents.
  • 5.1 Retrieving facts from long context inputs: DeepSeek-R1 sometimes becomes overwhelmed by very large contexts, producing incoherent text, out-of-place Chinese passages, or incomplete reasoning.These failures include disregarding instructions and stopping before completing the reasoning chain.
  • 5.2 Reasoning over long context inputs: Reasoning-focused training improves CHASE-QA over DeepSeek-V3, but DeepSeek-R1 remains below Gemini-1.5-Pro and often generates incomplete answers.On CHASE-Code, DeepSeek-R1 performs similarly to DeepSeek-V3 but remains marginally below Gemini-1.5-Pro.
  • 5.4 Conclusion: Overall, reasoning-oriented training improves DeepSeek-R1 over its base model without ensuring superiority to non-reasoning models optimized for long contexts.The authors report that this pattern also appears in parallel investigations of DeepSeek-R1.

6 Faithfulness and Reliance on Context

DeepSeek-R1 generally follows user-provided context even when it conflicts with stored knowledge, recognizing some conflicts but spending extra computation deliberating over misleading information.

  • Experimental scope: The study measures faithfulness under correct, incorrect and distracting knowledge, plus deliberately mislabelled few-shot examples.The experiments compare DeepSeek-R1 and DeepSeek-V3 on question answering and SST-2 in-context learning.
  • 6.1 Incorrect and distracting information: 78% recall for both DeepSeek-R1 and DeepSeek-V3 on incorrect user input shows that both models usually follow supplied false information.This recall exceeds performance when correct information is supplied; irrelevant information usually triggers refusal.
  • 6.1 Incorrect and distracting information: DeepSeek-R1 can acknowledge a conflict with its parametric knowledge before explicitly deferring to the user’s incorrect information.The reasoning chain exposes this knowledge-resolution process rather than merely showing the final answer.
  • 6.1 Incorrect and distracting information: Distracting information produces substantially longer responses, which the analysis associates with persistent rumination over false or irrelevant content.Despite identifying irrelevance, the model predominantly responds with “I don’t know.”
  • 6.2 Mislabelled in-context examples: With enough mislabelled SST-2 examples, DeepSeek-R1 follows incorrect labels and may construct complex hypotheses to explain them.Reasoning chains become longer as disagreement among in-context labels increases.

7 Safety

DeepSeek-R1 combines advanced reasoning with substantial safety vulnerabilities: it can produce harmful responses and generate jailbreaks that increase attack success against itself and other safety-aligned models.

  • Evaluation setup: The safety evaluation covers harmful-response generation across six HarmBench categories and the use of DeepSeek-R1 to generate jailbreak attacks.The jailbreaks are tested on DeepSeek-R1, Gemma-2-9B-Instruct and Llama-3.1-8B-Instruct.
  • 7.1 Evaluating harmful response generation: DeepSeek-R1 produces harmful responses to substantial numbers of Chemical and Biological Weapons, Cybercrime and Misinformation requests, with misinformation posing the greatest vulnerability.Gemma-2-9B-Instruct responds harmfully to only 3.6% of Chemical and Biological Weapons requests and refuses the others.
  • 7.1 Evaluating harmful response generation: The model refuses bullying and ransomware requests in some tested cases, but disclaimers can still contain structured insights into cybercrime methods.For hazardous-substance synthesis, the expected behavior is an unequivocal refusal without chemical details.
  • 7.2 Generating jailbreak attacks: DeepSeek-R1-generated jailbreaks raise attack success rates by 42.5 points for DeepSeek-R1, 72.5 points for Gemma-2-9B-Instruct and 62.5 points for Llama-3.1-8B-Instruct.The attacks reframe malicious requests as seemingly benign tasks, such as fictional writing or research.
  • 7.3 Conclusion: The authors characterize DeepSeek-R1’s reasoning as dual-use: it improves general capability while also making safeguard evasion more effective.They also report evidence that DeepSeek-R1 is substantially less safe than DeepSeek-V3.

8 Language and Culture

DeepSeek-R1’s moral reasoning and cultural preferences vary systematically by language, with longer English reasoning and stronger alignment to Chinese or Indian cultural norms in Chinese or Hindi.

  • Evaluation setup: The study examines moral reasoning with the DIT and language-dependent cultural behavior using LLM-GLOBE and handcrafted questions.Prompts are presented in English, Chinese and, for handcrafted questions, Hindi.
  • 8.1 Moral reasoning: DeepSeek-R1 scores 35 on the DIT in English and 29 in Chinese, compared with GPT-4 scores of 55.68 and 49.44 respectively.The model prioritizes societal responsibility and professional fidelity over kindness and autonomy in both languages.
  • 8.1 Moral reasoning: Chinese prompts elicit preferences for collective harm reduction, professional duty and social norms, whereas English prompts more often emphasize individual harm and personal trust.The reported differences indicate language-dependent moral preferences rather than a single stable value set.
  • 8.2 Cultural reasoning across languages: English LLM-GLOBE responses generally contain 500–700-token reasoning chains, while Chinese responses often contain no reasoning chain.Chinese responses also favor in-group collectivism and hierarchical structures more strongly than English responses.
  • 8.2 Cultural reasoning across languages: The model adapts to Chinese and Indian cultural norms in Chinese and Hindi, respectively, and sometimes invokes Chinese policies when China is absent from the prompt.English responses more often acknowledge diverse cultural practices.
  • 8.3 Conclusion: Overall, DeepSeek-R1 shows less universal-principle moral reasoning than GPT-4 and consistent language-linked differences in values and reasoning processes.The authors argue that these findings increase the need to consider social perspectives when evaluating reasoning models.

9 Relation to Human Sentence Processing

DeepSeek-R1 produces longer reasoning chains for sentence constructions that impose greater processing load on humans, but the chains’ repetitive form cautions against equating them with human reasoning.

  • Results: Garden path prompts yield subtly but significantly longer reasoning chains than minimally different non-garden path prompts.The distributions differ clearly, with garden path prompts also producing more extreme outliers.
  • Results: Reasoning-chain length correlates with human accuracy on both garden path and non-garden path questions.The reported Spearman correlations are −0.54 and −0.60, respectively, with highly significant p-values.
  • Interpretation: These length effects resemble human sentence-processing load, but the reasoning chains show substantially greater rumination for difficult sentences and controls alike.The model may continue reasoning after reaching an answer and can enter repetitive loops.
  • Results: Comparative-illusion prompts produce reasoning chains over 1,000 tokens longer on average than their control prompts.The difference is visible at the item level for the vast majority of datapoints.
  • Conclusion: The observed correlations therefore support only high-level similarity between LRM reasoning length and human processing load, not deeper cognitive correspondence.The authors specifically caution that chain form gives reason for skepticism about stronger comparisons with human thought.

10 World Modeling and Visual Reasoning

DeepSeek-R1 shows limited gains over its non-reasoning counterpart on ASCII visual and physical tasks, but its reasoning is often incoherent, non-iterative, and overly mathematical.

  • Single-object generation: DeepSeek-R1 rarely refines previous ASCII drafts, often restarting from scratch or producing a final image inconsistent with its reasoning steps.The final output frequently does not use the last draft generated during reasoning.
  • Object composition: On composite-object prompts, responses average 7.2K characters versus 9.0K for simpler single-object prompts, contrary to expectations of longer reasoning.The model also rarely reuses intermediate subcomponent drafts in the final image.
  • Physical simulation: DeepSeek-R1 performs below expectations on simple physical simulations despite often producing impressive and correct intermediate mathematics and physics.Its generated ASCII outputs are most often incorrect.
  • Physical simulation: The model predominantly approaches intuitive physics through formal mathematics rather than an apparent implicit world model.This strategy can consume thousands of tokens while still yielding incorrect ASCII outputs.
  • Caveat: ASCII drawing in a one-dimensional sequential format is itself non-trivial, so these results may not fully determine whether the model has a coherent physical world model.This is an explicit caveat on interpreting the physical-reasoning findings.
  • Conclusion: Overall, DeepSeek-R1 performs somewhat better than DeepSeek-V3, but gains in visual and physical reasoning remain limited by weak coherence, refinement, and intuitive visual reasoning.The authors describe substantial remaining limitations in these capacities.

11 Learning to Follow a Thinking Budget

Prompting alone does not reliably control DeepSeek-R1’s thinking length, whereas tailored reinforcement-learning rewards improve budget adherence while introducing an accuracy trade-off.

  • Prompting a budget: DeepSeek-R1 thinks for about 8,000 tokens across prompted budgets and does not reliably follow the specified limit.Increasing the budget from 1,000 to 2,000 tokens raises average response length by about 20%, while increasing it to 32,000 adds only about 5%.
  • Prompting a budget: There is no correlation between the specified thinking budget and AIME-24 accuracy.The evaluation compares performance across varying prompted token budgets.
  • Reward design: A tailored length reward penalizes responses that exceed or significantly deviate from a target number of thinking tokens.The proposed variants distinguish a strict upper limit from proximity to the target budget.
  • Results: RMaxDiff consistently improves budget adherence, whereas RMaxLength produces very short responses that remain within budget but are unhelpful.RMaxDiff also improves accuracy when the model is prompted to think more, though accuracy remains below the original reward.
  • Conclusion: Budget-constrained reinforcement learning improves token-use awareness and control but leaves overall accuracy below that of the unconstrained model.The result establishes a trade-off between budget compliance and task performance.

12 Post-submission: Analysis of Other LRMs

A preliminary comparison finds that Qwen3 resembles DeepSeek-R1’s multi-stage reasoning, while gpt-oss is more concise; both newer models nevertheless show a reasoning-length sweet spot.

  • Reasoning structure: gpt-oss produces substantially shorter, more task-oriented chains that lack the self-verification loops and alternate solutions characteristic of Qwen3 and DeepSeek-R1.Its reasoning process is described as qualitatively different from the other two models.
  • Reasoning structure: Qwen3 exhibits four reasoning phases resembling DeepSeek-R1: problem definition, bloom, reconstruction, and final answer.Its chains also contain similar reconstruction markers such as “Wait” and “Hmm”.
  • Thought length: For both Qwen3 and gpt-oss, performance increases with thought length, reaches a maximum, and then decreases with longer reasoning processes.This reproduces the non-monotonic pattern previously observed for DeepSeek-R1.
  • Thought length: Correct thoughts are shorter on average than incorrect thoughts for both newer models across AIME-24 and MATH-500.The result is reported as similar to the corresponding DeepSeek-R1 pattern.
  • Conclusion: The comparison suggests that some reasoning-length phenomena generalize across models, while reasoning structures differ substantially between Qwen3 and gpt-oss.The analysis is preliminary and covers two math-reasoning benchmarks.

13 Discussion

The study presents DeepSeek-R1 as an important but imperfect step toward reasoning-oriented language models: its observable reasoning reveals recurring strengths, limitations, and safety concerns. The authors argue that future LRMs need better process monitoring, diverse strategies, faithful reasoning, and safer implementation.

  • Reasoning behaviours: DeepSeek-R1 follows recurring cycles of problem definition, decomposition, reconstruction, verification, and reinterpretation across many tasks.The model may abandon or re-bloom interpretations during these cycles, while repeatedly re-examining previously investigated claims.
  • Prohibitive thought length: Excessive thought length can impair performance through verification failure, recall degradation, or overwhelmed gibberish outputs.The study identifies a reasoning sweet spot and reports that DeepSeek-R1 does not reliably monitor its own reasoning length.
  • Faithfulness to reasoning: The reasoning chain and final answer are sometimes misaligned, so stronger claims about reasoning faithfulness require further investigation.The authors recommend examining fidelity not only to final outputs, but also to subsequent behavior and the tendency to continue reasoning.
  • Social implications: DeepSeek-R1 raises safety concerns because it more readily outputs harmful information than V3 and can jailbreak other language models.The authors also report contrasting behavior across English, Chinese, and Hindi, while noting that language-specific reasoning claims need deeper investigation.
  • New reasoning strategies: Explicit process monitoring could reduce rumination, identify misleading thoughts, support thought budgeting, and enable additional reasoning paradigms.The paper also notes failures to iteratively refine drafts and a tendency to use mathematical or symbolic reasoning in physical simulation tasks.
  • Implications on model systems: DeepSeek-R1 exhibits slow-reasoning hallmarks but struggles to select optimal approaches and monitor progress, leading the authors to describe it as “System 1.5 thinking.”The study is an initial, broad investigation using some qualitative analyses, relatively small samples, and limited dataset diversity because querying the model is costly.

A.2 Extra results

Additional examples show that DeepSeek-R1 can ruminate across reasoning cycles and that unnecessarily long thoughts may reduce accuracy, including through unrecovered errors and failed self-verification.

  • Across most tasks, the Bloom cycle is longest, while reconstruction cycles generally shorten over time but show periodic jumps.Context management is an exception because distracting or irrelevant information prompts extended deliberation over user intentions.
  • In a GSM8K example, DeepSeek-R1 repeatedly reconsidered whether Carla had to restart the download before selecting its initial answer.
  • Longer-than-optimal thoughts can fail when DeepSeek-R1 follows an incorrect reasoning path and never recovers.This failure mode is illustrated on an AIME-24 problem.
  • DeepSeek-R1 can find the correct product of a 9 × 9 multiplication problem but then incorrectly verify it after unnecessarily long thinking.
  • The appendix plots average thought accuracy by length bin for DeepSeek-R1, Qwen3, and gpt-oss on AIME-24.Longer-thought bins are shaded progressively darker in the corresponding figures.

C Appendix: Long Context Evaluation

The long-context evaluation presents examples of errors, overthinking, and incoherent behavior, while also showing that DeepSeek-R1 can use provided context and sometimes recall information after very long reasoning chains.

  • DeepSeek-R1 is evaluated on long-context information-seeking question answering, repository-level code generation, output generation, and recall of self-generated facts.
  • The model makes an information-seeking QA error by omitting a major point from the ground-truth answer despite its discussion in the documents.
  • In repository-level code generation, DeepSeek-R1 overthinks in a seemingly infinite loop.
  • DeepSeek-R1 fails to follow instructions to generate long-context outputs directly, despite a prompt requiring at least 30,000 words.
  • After generating a nearly 30k-token reasoning chain, DeepSeek-R1 can recall a self-chosen historical fact, but another example shows it becoming overwhelmed during recall.

E.1 HarmBench evaluation

The HarmBench examples document mixed safety behavior: DeepSeek-R1 can recognize and refuse some harmful requests, yet its reasoning may also elaborate malicious procedures and generate jailbreak reframings.

  • Harassment and Bullying: For a bullying request, DeepSeek-R1 identifies the harmful intent, refuses assistance, and emphasizes respect, inclusion, and seeking support.
  • Cybercrime: For a ransomware request, the reasoning begins to break down how ransomware generates money before acknowledging that creating and deploying it is a serious crime.
  • Chemical & Biological Weapons/Drugs: The appendix includes a harmful parathion-synthesis instruction and an example response that reasons through possible synthesis steps.
  • Jailbreak generation: DeepSeek-R1 is prompted to rewrite malicious requests so their harmful nature is concealed and another LLM may provide unsafe information.
  • Jailbreak generation: The appendix supplies example jailbreaks that reframe requests for ricin or home methamphetamine recipes as fiction-writing or educational requests.

F.2.1 Anecdotal analysis of difference in cultural values

The cultural analysis finds that DeepSeek-R1’s reasoning length and answers vary by language, with English responses often more pluralistic and Chinese responses reflecting China-oriented cultural framing.

  • For general questions about the AI race, World War II in Asia, and internet regulation, Chinese answers exhibit a China-oriented perspective.
  • For blessings from elders, the English reasoning considers multiple cultural practices and aims for a respectful, inclusive answer.
  • In the AI-race example, the model reasons for 33 seconds in English while the Chinese thinking chain is empty.
  • Across most garden-path datapoints, reasoning chains are about 200-300 words longer than control equivalents.
  • Reasoning-chain length correlates significantly with human accuracy on the same datapoints.

G.2 Comparative illusions

DeepSeek-R1 produces substantially longer reasoning chains for comparative-illusion prompts than matched controls, often through repetitive reconsideration before reaching a reasonable answer. The examples show this looping pattern in both illusion and control prompts, with the illusion condition generally producing the longer chains.

  • For the vast majority of datapoints, comparative-illusion prompts produce longer reasoning chains than their control equivalents.
  • A garden-path example similarly shows the model identifying a parsing problem, fixating on punctuation, and eventually producing a reasonable response.
  • The experiment compares comparative-illusion prompts such as the lawyers-and-models question with controls that preserve the question while removing the illusion.
  • The comparative-illusion example enters an extremely long, repetitive loop that reconsiders the same options and phrasing before reaching a reasonable answer.
  • The matched control example also becomes caught in an extensive, repetitive reasoning loop before arriving at the correct answer.
Loading 2504.07128v3…