Example table: reasoning in LLMs | Collection partagée

Créer un compte gratuit pour copier ces articles Se connecter

Articles

Chain of Thought Prompting Elicits Reasoning in Large Language Models
Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Ed Chi, Quoc Le, Denny Zhou · arXiv: 2201.11903v1 · 2022 · Advances in Neural Information Processing Systems 35

Publication retrouvée
Self-Consistency Improves Chain of Thought Reasoning in Language Models
Xuezhi Wang, Jason Wei, Dale Schuurmans, Quoc Le, Ed Chi, Denny Zhou · arXiv: 2203.11171v1 · 2022

Prépublication retrouvée ; aucune publication en revue confirmée
STaR: Bootstrapping Reasoning With Reasoning
Eric Zelikman, Yuhuai Wu, Noah D. Goodman · arXiv: 2203.14465v1 · 2022
Let's Verify Step by Step
Hunter Lightman, Vineet Kosaraju, Yura Burda, Harri Edwards, Bowen Baker, Teddy Lee, Jan Leike, John Schulman · arXiv: 2305.20050v1 · 2023

Prépublication retrouvée ; aucune publication en revue confirmée
DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning
DeepSeek-AI, Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Ruoyu Zhang, Runxin Xu, Qihao Zhu · arXiv: 2501.12948v1 · 2025 · Nature

Publication retrouvée

Tableau comparatif

23 cases remplies sur 23 s’appuient sur un passage de l’article.

Article	Question	Méthode	Données	Résultats principaux	Limites
1 Chain of Thought Prompting Elicits Reasoning in Large Language Models	Étudier si l’incitation à produire un raisonnement intermédiaire améliore les performances des grands modèles de langage sur les tâches de raisonnement. Source · Abstract In this paper, we explore chain of thought prompting as a method for improving the ability of language models to perform reasoning tasks.	Utiliser des exemples de raisonnements intermédiaires dans les invites, sans ajuster les paramètres des modèles. Source · Abstract First, we provide the exact input prompts for all tasks in Table 19–Table 26 in Appendix C (and emphasize that we do not perform any ﬁnetuning and only apply prompting to off-the-shelf language models).	Les expériences portent sur des tâches de raisonnement arithmétique, symbolique et de sens commun. La taille des jeux de données n’est pas précisée dans l’extrait. Source · Abstract Through experiments on arith- metic, symbolic, and commonsense reasoning, we ﬁnd that chain of thought processing is an emergent property of model scale that allows sufﬁciently large language models to perform reasoning tasks that otherwise have ﬂat scaling curves.	Les raisonnements intermédiaires améliorent les performances sur diverses tâches où les exemples classiques ne suffisent pas. Les bénéfices apparaissent à une échelle d’environ 100 milliards de paramètres. Source · Abstract Chain of thought prompting is simple and improves performance across various reasoning tasks where standard few-shot prompting is insufﬁcient. Crucially, our experi- ments suggest that successful chain of thought prompting is an emergent property of model scale—that is, the beneﬁts of chain of thought prompting only materialize at sufﬁcient model scale (around 100B parameters).	Les modèles principaux utilisés ne sont pas accessibles publiquement, ce qui limite la reproductibilité directe. Source · Abstract As our results make use of a set of large language models that is not publicly available, we take the following actions to facilitate reproducibility.
2 Self-Consistency Improves Chain of Thought Reasoning in Language Models	Évaluer si l’agrégation de plusieurs réponses générées améliore la précision du raisonnement des grands modèles de langage. Source · Abstract We explore a simple ensemble strategy, self-consistency, that signiﬁcantly improves the reasoning accuracy of large language models.	Échantillonner plusieurs raisonnements à partir d’une invite contenant des exemples, puis retenir la réponse finale la plus fréquente. Source · Abstract The method we propose ﬁrst prompts the language model with example chains of thought, then generates a diverse set of reasoning paths by sampling from the model’s decoder. Each reasoning path might lead to a different ﬁnal answer, so we determine the optimal answer by taking a plurality or majority vote—i.e., the most commonly occurring answer (corresponding to a majority vote in the special case of only two alternatives).	L’évaluation utilise les jeux de raisonnement arithmétique et de sens commun GSM8K, SVAMP, MultiArith, CommonsenseQA et ARC, dans ses versions facile et difficile. Source · Abstract For arithmetic and commonsense reasoning benchmarks we ﬁnd that self-consistency yields signiﬁcant accuracy improvements in a variety of datasets, such as GSM8K (+10%), SVAMP (+14%), MultiArith (+24%), CommonsenseQA (+5%) and ARC (easy +4%, challenge +5%).	Les gains de précision annoncés sont de +10 % sur GSM8K, +14 % sur SVAMP, +24 % sur MultiArith et +5 % sur CommonsenseQA. Sur ARC, ils atteignent +4 % pour la version facile et +5 % pour la version difficile. Source · Abstract For arithmetic and commonsense reasoning benchmarks we ﬁnd that self-consistency yields signiﬁcant accuracy improvements in a variety of datasets, such as GSM8K (+10%), SVAMP (+14%), MultiArith (+24%), CommonsenseQA (+5%) and ARC (easy +4%, challenge +5%).	Non précisé
3 STaR: Bootstrapping Reasoning With Reasoning Résumé d’auteur et notre synthèse	Permettre à un modèle d’améliorer ses capacités de raisonnement en apprenant de ses propres justifications, à partir de peu d’exemples annotés. Source · Abstract We propose a technique to iteratively leverage a small number of rationale examples and a large dataset without rationales, to bootstrap the ability to perform successively more complex reasoning.	STaR génère des justifications et, en cas d’erreur, recommence en fournissant la bonne réponse. Le modèle est ensuite ajusté sur les justifications aboutissant à une réponse correcte, puis le cycle est répété. Source · Abstract This technique, the "Self-Taught Reasoner" (STaR), relies on a simple loop: generate rationales to answer many questions, prompted with a few rationale examples; if the generated answers are wrong, try again to generate a rationale given the correct answer; fine-tune on all the rationales that ultimately yielded correct answers; repeat.	La méthode exploite quelques exemples avec justifications et un grand jeu de données sans justifications. L’évaluation couvre plusieurs jeux de données, dont CommonsenseQA ; leurs tailles ne sont pas précisées. Source · Abstract We propose a technique to iteratively leverage a small number of rationale examples and a large dataset without rationales, to bootstrap the ability to perform successively more complex reasoning. This technique, the "Self-Taught Reasoner" (STaR), relies on a simple loop: generate rationales to answer many questions, prompted with a few rationale examples; if the generated answers are wrong, try again to generate a rationale given the correct answer; fine-tune on all the rationales that ultimately yielded correct answers; repeat.	STaR améliore les performances par rapport à un ajustement visant directement les réponses finales. Sur CommonsenseQA, ses performances sont comparables à celles d’un modèle de pointe 30 fois plus grand ajusté sur cette tâche. Source · Abstract We show that STaR significantly improves performance on multiple datasets compared to a model fine-tuned to directly predict final answers, and performs comparably to fine-tuning a 30$\times$ larger state-of-the-art language model on CommensenseQA.	Non précisé
4 Let's Verify Step by Step	Comparer la supervision des étapes intermédiaires à celle du seul résultat final pour améliorer la fiabilité des solutions mathématiques. Source · Abstract, ¶1 We conduct our own investigation, finding that process supervision significantly outperforms outcome supervision for training models to solve problems from the challenging MATH dataset. Ouvrir dans l’article	Entraîner des modèles de récompense évaluant soit le résultat final, soit chaque étape, puis les utiliser pour sélectionner des solutions. Le générateur reste fixe et n’est pas amélioré par apprentissage par renforcement. Source · 2.1 Scope, ¶1 At each model scale, we use a single fixed model to generate all solutions. We call this model the generator. Ouvrir dans l’article	L’étude utilise le jeu MATH et publie PRM800K, composé de 800 000 annotations humaines au niveau des étapes pour entraîner le meilleur modèle de récompense. Source · Abstract, ¶1 We conduct our own investigation, finding that process supervision significantly outperforms outcome supervision for training models to solve problems from the challenging MATH dataset. Our process-supervised model solves 78% of problems from a representative subset of the MATH test set. Absent de ce passage : 800000 Ouvrir dans l’article	Le modèle supervisé au niveau des étapes résout 78 % des problèmes d’un sous-ensemble représentatif du test MATH. L’apprentissage actif améliore également l’efficacité de cette supervision. Source · Abstract, ¶1 Our process-supervised model solves 78% of problems from a representative subset of the MATH test set. Additionally, we show that active learning significantly improves the efficacy of process supervision. Ouvrir dans l’article	À grande échelle, les jeux d’entraînement des deux types de modèles de récompense ne sont pas directement comparables. Cette limite empêche une comparaison strictement contrôlée dans ce régime. Source · 2 Methods, ¶2 Unfortunately the training sets for these reward models are not directly comparable, for reasons we will discuss in Section 3. These models are therefore not ideal for making an apples-to-apples comparison of outcome and process supervision. Ouvrir dans l’article
5 DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning Résumé d’auteur et notre synthèse	Développer des modèles de raisonnement et explorer l’émergence de ces capacités par apprentissage par renforcement à grande échelle, sans ajustement supervisé préalable. Source · Abstract We introduce our first-generation reasoning models, DeepSeek-R1-Zero and DeepSeek-R1. DeepSeek-R1-Zero, a model trained via large-scale reinforcement learning (RL) without supervised fine-tuning (SFT) as a preliminary step, demonstrates remarkable reasoning capabilities.	R1-Zero est entraîné par apprentissage par renforcement sans ajustement supervisé préalable. R1 ajoute des données d’amorçage et un entraînement en plusieurs étapes avant le renforcement. Source · Abstract To address these issues and further enhance reasoning performance, we introduce DeepSeek-R1, which incorporates multi-stage training and cold-start data before RL.	R1 utilise des données d’amorçage avant l’apprentissage par renforcement. Leur contenu et leur taille ne sont pas précisés dans l’extrait. Source · Abstract To address these issues and further enhance reasoning performance, we introduce DeepSeek-R1, which incorporates multi-stage training and cold-start data before RL.	R1 atteint des performances comparables à celles d’OpenAI-o1-1217 sur les tâches de raisonnement. Aucun score chiffré n’est fourni dans l’extrait. Source · Abstract DeepSeek-R1 achieves performance comparable to OpenAI-o1-1217 on reasoning tasks.	R1-Zero présente des problèmes de lisibilité et de mélange des langues. Source · Abstract However, it encounters challenges such as poor readability, and language mixing.

Où ils s'accordent

Les cinq études rapportent des résultats favorables à des approches qui exploitent le raisonnement intermédiaire, mais leurs protocoles différents ne permettent pas de comparer directement l’ampleur des bénéfices. [1, 2, 3, 4, 5]
10 passages derrière ce constat

1 · Méthode (Abstract): First, we provide the exact input prompts for all tasks in Table 19–Table 26 in Appendix C (and emphasize that we do not perform any ﬁnetuning and only apply prompting to off-the-shelf language models).
1 · Résultats principaux (Abstract): Chain of thought prompting is simple and improves performance across various reasoning tasks where standard few-shot prompting is insufﬁcient. Crucially, our experi- ments suggest that successful chain of thought prompting is an emergent property of model scale—that is, the beneﬁts of chain of thought prompting only materialize at sufﬁcient model scale (around 100B parameters).
2 · Méthode (Abstract): The method we propose ﬁrst prompts the language model with example chains of thought, then generates a diverse set of reasoning paths by sampling from the model’s decoder. Each reasoning path might lead to a different ﬁnal answer, so we determine the optimal answer by taking a plurality or majority vote—i.e., the most commonly occurring answer (corresponding to a majority vote in the special case of only two alternatives).
2 · Résultats principaux (Abstract): For arithmetic and commonsense reasoning benchmarks we ﬁnd that self-consistency yields signiﬁcant accuracy improvements in a variety of datasets, such as GSM8K (+10%), SVAMP (+14%), MultiArith (+24%), CommonsenseQA (+5%) and ARC (easy +4%, challenge +5%).
3 · Méthode (Abstract): This technique, the "Self-Taught Reasoner" (STaR), relies on a simple loop: generate rationales to answer many questions, prompted with a few rationale examples; if the generated answers are wrong, try again to generate a rationale given the correct answer; fine-tune on all the rationales that ultimately yielded correct answers; repeat.
3 · Résultats principaux (Abstract): We show that STaR significantly improves performance on multiple datasets compared to a model fine-tuned to directly predict final answers, and performs comparably to fine-tuning a 30$\times$ larger state-of-the-art language model on CommensenseQA.
4 · Méthode (2.1 Scope, ¶1): At each model scale, we use a single fixed model to generate all solutions. We call this model the generator.
4 · Résultats principaux (Abstract, ¶1): Our process-supervised model solves 78% of problems from a representative subset of the MATH test set. Additionally, we show that active learning significantly improves the efficacy of process supervision.
5 · Méthode (Abstract): To address these issues and further enhance reasoning performance, we introduce DeepSeek-R1, which incorporates multi-stage training and cold-start data before RL.
5 · Résultats principaux (Abstract): DeepSeek-R1 achieves performance comparable to OpenAI-o1-1217 on reasoning tasks.
Les deux études sur l’incitation montrent qu’il est possible d’améliorer les performances au moment de la génération : l’une fournit des exemples de raisonnements sans modifier les paramètres, l’autre agrège plusieurs raisonnements pour retenir la réponse finale la plus fréquente. [1, 2]
4 passages derrière ce constat

1 · Méthode (Abstract): First, we provide the exact input prompts for all tasks in Table 19–Table 26 in Appendix C (and emphasize that we do not perform any ﬁnetuning and only apply prompting to off-the-shelf language models).
1 · Résultats principaux (Abstract): Chain of thought prompting is simple and improves performance across various reasoning tasks where standard few-shot prompting is insufﬁcient. Crucially, our experi- ments suggest that successful chain of thought prompting is an emergent property of model scale—that is, the beneﬁts of chain of thought prompting only materialize at sufﬁcient model scale (around 100B parameters).
2 · Méthode (Abstract): The method we propose ﬁrst prompts the language model with example chains of thought, then generates a diverse set of reasoning paths by sampling from the model’s decoder. Each reasoning path might lead to a different ﬁnal answer, so we determine the optimal answer by taking a plurality or majority vote—i.e., the most commonly occurring answer (corresponding to a majority vote in the special case of only two alternatives).
2 · Résultats principaux (Abstract): For arithmetic and commonsense reasoning benchmarks we ﬁnd that self-consistency yields signiﬁcant accuracy improvements in a variety of datasets, such as GSM8K (+10%), SVAMP (+14%), MultiArith (+24%), CommonsenseQA (+5%) and ARC (easy +4%, challenge +5%).
Les raisonnements intermédiaires constituent aussi une ressource d’apprentissage : les justifications générées servent à ajuster un modèle, tandis que les annotations humaines des étapes servent à entraîner un modèle de récompense. Les deux études rapportent des bénéfices de ces usages. [3, 4]
6 passages derrière ce constat

3 · Méthode (Abstract): This technique, the "Self-Taught Reasoner" (STaR), relies on a simple loop: generate rationales to answer many questions, prompted with a few rationale examples; if the generated answers are wrong, try again to generate a rationale given the correct answer; fine-tune on all the rationales that ultimately yielded correct answers; repeat.
3 · Données (Abstract): We propose a technique to iteratively leverage a small number of rationale examples and a large dataset without rationales, to bootstrap the ability to perform successively more complex reasoning. This technique, the "Self-Taught Reasoner" (STaR), relies on a simple loop: generate rationales to answer many questions, prompted with a few rationale examples; if the generated answers are wrong, try again to generate a rationale given the correct answer; fine-tune on all the rationales that ultimately yielded correct answers; repeat.
3 · Résultats principaux (Abstract): We show that STaR significantly improves performance on multiple datasets compared to a model fine-tuned to directly predict final answers, and performs comparably to fine-tuning a 30$\times$ larger state-of-the-art language model on CommensenseQA.
4 · Méthode (2.1 Scope, ¶1): At each model scale, we use a single fixed model to generate all solutions. We call this model the generator.
4 · Données (Abstract, ¶1): We conduct our own investigation, finding that process supervision significantly outperforms outcome supervision for training models to solve problems from the challenging MATH dataset. Our process-supervised model solves 78% of problems from a representative subset of the MATH test set.
4 · Résultats principaux (Abstract, ¶1): Our process-supervised model solves 78% of problems from a representative subset of the MATH test set. Additionally, we show that active learning significantly improves the efficacy of process supervision.

Où ils divergent

Le tableau ne permet pas d’établir de contradiction directe entre les résultats : les études évaluent des interventions, des modèles et des tâches différents, sans comparaison commune aux cinq approches. [1, 2, 3, 4, 5]
15 passages derrière ce constat

1 · Méthode (Abstract): First, we provide the exact input prompts for all tasks in Table 19–Table 26 in Appendix C (and emphasize that we do not perform any ﬁnetuning and only apply prompting to off-the-shelf language models).
1 · Données (Abstract): Through experiments on arith- metic, symbolic, and commonsense reasoning, we ﬁnd that chain of thought processing is an emergent property of model scale that allows sufﬁciently large language models to perform reasoning tasks that otherwise have ﬂat scaling curves.
1 · Résultats principaux (Abstract): Chain of thought prompting is simple and improves performance across various reasoning tasks where standard few-shot prompting is insufﬁcient. Crucially, our experi- ments suggest that successful chain of thought prompting is an emergent property of model scale—that is, the beneﬁts of chain of thought prompting only materialize at sufﬁcient model scale (around 100B parameters).
2 · Méthode (Abstract): The method we propose ﬁrst prompts the language model with example chains of thought, then generates a diverse set of reasoning paths by sampling from the model’s decoder. Each reasoning path might lead to a different ﬁnal answer, so we determine the optimal answer by taking a plurality or majority vote—i.e., the most commonly occurring answer (corresponding to a majority vote in the special case of only two alternatives).
2 · Données (Abstract): For arithmetic and commonsense reasoning benchmarks we ﬁnd that self-consistency yields signiﬁcant accuracy improvements in a variety of datasets, such as GSM8K (+10%), SVAMP (+14%), MultiArith (+24%), CommonsenseQA (+5%) and ARC (easy +4%, challenge +5%).
2 · Résultats principaux (Abstract): For arithmetic and commonsense reasoning benchmarks we ﬁnd that self-consistency yields signiﬁcant accuracy improvements in a variety of datasets, such as GSM8K (+10%), SVAMP (+14%), MultiArith (+24%), CommonsenseQA (+5%) and ARC (easy +4%, challenge +5%).
3 · Méthode (Abstract): This technique, the "Self-Taught Reasoner" (STaR), relies on a simple loop: generate rationales to answer many questions, prompted with a few rationale examples; if the generated answers are wrong, try again to generate a rationale given the correct answer; fine-tune on all the rationales that ultimately yielded correct answers; repeat.
3 · Données (Abstract): We propose a technique to iteratively leverage a small number of rationale examples and a large dataset without rationales, to bootstrap the ability to perform successively more complex reasoning. This technique, the "Self-Taught Reasoner" (STaR), relies on a simple loop: generate rationales to answer many questions, prompted with a few rationale examples; if the generated answers are wrong, try again to generate a rationale given the correct answer; fine-tune on all the rationales that ultimately yielded correct answers; repeat.
3 · Résultats principaux (Abstract): We show that STaR significantly improves performance on multiple datasets compared to a model fine-tuned to directly predict final answers, and performs comparably to fine-tuning a 30$\times$ larger state-of-the-art language model on CommensenseQA.
4 · Méthode (2.1 Scope, ¶1): At each model scale, we use a single fixed model to generate all solutions. We call this model the generator.
4 · Données (Abstract, ¶1): We conduct our own investigation, finding that process supervision significantly outperforms outcome supervision for training models to solve problems from the challenging MATH dataset. Our process-supervised model solves 78% of problems from a representative subset of the MATH test set.
4 · Résultats principaux (Abstract, ¶1): Our process-supervised model solves 78% of problems from a representative subset of the MATH test set. Additionally, we show that active learning significantly improves the efficacy of process supervision.
5 · Méthode (Abstract): To address these issues and further enhance reasoning performance, we introduce DeepSeek-R1, which incorporates multi-stage training and cold-start data before RL.
5 · Données (Abstract): To address these issues and further enhance reasoning performance, we introduce DeepSeek-R1, which incorporates multi-stage training and cold-start data before RL.
5 · Résultats principaux (Abstract): DeepSeek-R1 achieves performance comparable to OpenAI-o1-1217 on reasoning tasks.
Les bénéfices de l’agrégation varient selon les jeux de données : les gains de précision annoncés sont de +10 % sur GSM8K, +14 % sur SVAMP, +24 % sur MultiArith, +5 % sur CommonsenseQA, +4 % sur ARC facile et +5 % sur ARC difficile. [2]
2 passages derrière ce constat

2 · Données (Abstract): For arithmetic and commonsense reasoning benchmarks we ﬁnd that self-consistency yields signiﬁcant accuracy improvements in a variety of datasets, such as GSM8K (+10%), SVAMP (+14%), MultiArith (+24%), CommonsenseQA (+5%) and ARC (easy +4%, challenge +5%).
2 · Résultats principaux (Abstract): For arithmetic and commonsense reasoning benchmarks we ﬁnd that self-consistency yields signiﬁcant accuracy improvements in a variety of datasets, such as GSM8K (+10%), SVAMP (+14%), MultiArith (+24%), CommonsenseQA (+5%) and ARC (easy +4%, challenge +5%).
Les constats sur l’échelle portent sur des comparaisons différentes : les bénéfices de l’incitation apparaissent vers 100 milliards de paramètres, tandis que l’apprentissage itératif des justifications atteint sur CommonsenseQA des performances comparables à celles d’un modèle ajusté 30 fois plus grand. Ces résultats ne démontrent donc pas un seuil de taille commun. [1, 3]
4 passages derrière ce constat

1 · Méthode (Abstract): First, we provide the exact input prompts for all tasks in Table 19–Table 26 in Appendix C (and emphasize that we do not perform any ﬁnetuning and only apply prompting to off-the-shelf language models).
1 · Résultats principaux (Abstract): Chain of thought prompting is simple and improves performance across various reasoning tasks where standard few-shot prompting is insufﬁcient. Crucially, our experi- ments suggest that successful chain of thought prompting is an emergent property of model scale—that is, the beneﬁts of chain of thought prompting only materialize at sufﬁcient model scale (around 100B parameters).
3 · Méthode (Abstract): This technique, the "Self-Taught Reasoner" (STaR), relies on a simple loop: generate rationales to answer many questions, prompted with a few rationale examples; if the generated answers are wrong, try again to generate a rationale given the correct answer; fine-tune on all the rationales that ultimately yielded correct answers; repeat.
3 · Résultats principaux (Abstract): We show that STaR significantly improves performance on multiple datasets compared to a model fine-tuned to directly predict final answers, and performs comparably to fine-tuning a 30$\times$ larger state-of-the-art language model on CommensenseQA.
Les résultats ne sont pas exprimés sur une base commune : la supervision des étapes atteint 78 % de problèmes résolus sur un sous-ensemble représentatif du test MATH, tandis que le modèle entraîné par renforcement est annoncé comparable à OpenAI-o1-1217 sans score chiffré dans le tableau. Aucun classement quantitatif entre ces approches n’est possible ici. [4, 5]
4 passages derrière ce constat

4 · Données (Abstract, ¶1): We conduct our own investigation, finding that process supervision significantly outperforms outcome supervision for training models to solve problems from the challenging MATH dataset. Our process-supervised model solves 78% of problems from a representative subset of the MATH test set.
4 · Résultats principaux (Abstract, ¶1): Our process-supervised model solves 78% of problems from a representative subset of the MATH test set. Additionally, we show that active learning significantly improves the efficacy of process supervision.
5 · Données (Abstract): To address these issues and further enhance reasoning performance, we introduce DeepSeek-R1, which incorporates multi-stage training and cold-start data before RL.
5 · Résultats principaux (Abstract): DeepSeek-R1 achieves performance comparable to OpenAI-o1-1217 on reasoning tasks.
Les approches diffèrent par ce qu’elles modifient : l’incitation laisse les paramètres inchangés, l’agrégation sélectionne une réponse parmi plusieurs générations, l’apprentissage des justifications ajuste le modèle, la supervision des étapes entraîne un évaluateur en gardant le générateur fixe, et la dernière étude entraîne les modèles par renforcement. [1, 2, 3, 4, 5]
5 passages derrière ce constat

1 · Méthode (Abstract): First, we provide the exact input prompts for all tasks in Table 19–Table 26 in Appendix C (and emphasize that we do not perform any ﬁnetuning and only apply prompting to off-the-shelf language models).
2 · Méthode (Abstract): The method we propose ﬁrst prompts the language model with example chains of thought, then generates a diverse set of reasoning paths by sampling from the model’s decoder. Each reasoning path might lead to a different ﬁnal answer, so we determine the optimal answer by taking a plurality or majority vote—i.e., the most commonly occurring answer (corresponding to a majority vote in the special case of only two alternatives).
3 · Méthode (Abstract): This technique, the "Self-Taught Reasoner" (STaR), relies on a simple loop: generate rationales to answer many questions, prompted with a few rationale examples; if the generated answers are wrong, try again to generate a rationale given the correct answer; fine-tune on all the rationales that ultimately yielded correct answers; repeat.
4 · Méthode (2.1 Scope, ¶1): At each model scale, we use a single fixed model to generate all solutions. We call this model the generator.
5 · Méthode (Abstract): To address these issues and further enhance reasoning performance, we introduce DeepSeek-R1, which incorporates multi-stage training and cold-start data before RL.

Ce qu'aucun n'aborde

À budget de calcul, modèles de départ et jeux de données identiques, quelle approche offre le meilleur compromis entre précision, coût d’entraînement et coût de génération ? Le tableau ne présente aucune comparaison contrôlée de ce type.
Les raisonnements produits reflètent-ils fidèlement les mécanismes qui déterminent les réponses, ou peuvent-ils fournir une justification plausible sans être causalement déterminants ? Aucune évaluation de cette fidélité n’est décrite.
Dans quelle mesure les gains résistent-ils à des changements de distribution, à des entrées adversariales ou à des tâches éloignées des évaluations retenues ? Aucun protocole explicite de ce type n’est présenté.
Quels gains supplémentaires, ou quelles interférences, résulteraient d’une combinaison de l’agrégation des réponses, de l’apprentissage itératif des justifications, de la supervision des étapes et du renforcement ? Le tableau ne décrit pas d’évaluation conjointe de ces composantes.
Les améliorations se maintiennent-elles de façon comparable selon les langues ? Le mélange des langues est signalé comme une limite, mais aucune évaluation comparative multilingue n’est rapportée.