Speculative decoding is a powerful technique that accelerates Large Language Model (LLM) inference by leveraging a lightweight speculative draft model. However, existing designs suffers in performance due to misalignment between training and inference. Recent methods have tried to solve this issue by adopting a multi-step training strategy, but the complex inputs of different training steps make it harder for the draft model to converge. To address this, we propose CORAL, a novel framework that improves both accuracy and efficiency in speculative drafting. CORAL introduces Cross-Step Representation Alignment, a method that enhances consistency across multiple training steps, significantly improving speculative drafting performance. Additionally, we identify the LM head as a major bottleneck in the inference speed of the draft model. We introduce a weight-grouping mechanism that selectively activates a subset of LM head parameters during inference, substantially reducing the latency of the draft model. We evaluate CORAL on three LLM families and three benchmark datasets, achieving speedup ratios of 2.50x-4.07x, outperforming state-of-the-art methods such as EAGLE-2 and HASS. Our results demonstrate that CORAL effectively mitigates training-inference misalignment and delivers significant speedup for modern LLMs with large vocabularies.

Vision-Language Models (VLMs) show promise for autonomous driving, yet their struggle with hallucinations, inefficient reasoning, and limited real-world validation hinders accurate perception and robust step-by-step reasoning. To overcome this, we introduce AgentThink, a pioneering unified framework that, for the first time, integrates Chain-of-Thought (CoT) reasoning with dynamic, agent-style tool invocation for autonomous driving tasks. AgentThink’s core innovations include: (i) Structured Data Generation, by establishing an autonomous driving tool library to automatically construct structured, self-verified reasoning data explicitly incorporating tool usage for diverse driving scenarios; (ii) A Two-stage Training Pipeline, employing Supervised Fine-Tuning (SFT) with Group Relative Policy Optimization (GRPO) to equip VLMs with the capability for autonomous tool invocation; and (iii) Agent-style Tool-Usage Evaluation, introducing a novel multi-tool assessment protocol to rigorously evaluate the model’s tool invocation and utilization. Experiments on the DriveLMM-o1 benchmark demonstrate AgentThink significantly boosts overall reasoning scores by 53.91% and enhances answer accuracy by 33.54%, while markedly improving reasoning quality and consistency. Furthermore, ablation studies and robust zero-shot/few-shot generalization experiments across various benchmarks underscore its powerful capabilities. These findings highlight a promising trajectory for developing trustworthy and tool-aware autonomous driving models.

To enable finer-grained linguistic analysis, we propose a method for the separation of lexical and grammatical information within contextualized word embeddings. Using CamemBERT embeddings for French, we apply our method to 14,472 inflected word forms extracted from the Lexical Network of French ( LN-fr ), covering 1,468 nouns, 202 adjectives and 299 verbs inflected via 14 distinct grammatical feature values. Our iterative distillation alternates two steps until convergence: (i) estimating lexical or grammatical vectors by averaging the embeddings of words that share the same lexeme or grammatical feature value, and (ii) isolating the complementary component of each word embedding by subtracting the estimated vector. To assess the quality of the decomposition, we measure whether the resulting lexical and grammatical vectors form more compact clusters within their respective groups and whether their sum better reconstructs the original word embeddings. All evaluations rely on L2 distance. The observed improvements in both clustering and reconstruction accuracy demonstrate the effectiveness of our approach.

BERT-like language models have been demonstrated to capture the idiomatic meaning of multiword expressions. Linguists have also shown that idioms have varying degrees of idiomaticity. In this paper, we assess CamemBERT’s sensitivity to the degree of idiomaticity within idioms, as well as the dependency of this sensitivity on part of speech and idiom length. We used a demasking task on tokens from 3127 idioms and 22551 tokens corresponding to simple lexemes taken from the French Lexical Network (LN-fr), and observed that CamemBERT performs distinctly on tokens embedded within idioms compared to simple ones. When demasking tokens within idioms, the model is not proficient in discerning their level of idiomaticity. Moreover, regardless of idiomaticity, CamemBERT excels at handling function words. The length of idioms also impacts CamemBERT’s performance to a certain extent. The last two observations partly explain the difference between the model’s performance on idioms versus simple lexemes. We conclude that the model treats idioms differently from simple lexemes, but that it does not capture the difference in compositionality between subclasses of idioms.

Dans cet article nous proposons une nouvelle solution pour extraire le contour interne des lèvres d’un locuteur sans utiliser d’artifices. La méthode s’appuie sur un algorithme récent d’extraction du contour de visage développé en vision par ordinateur, CLNF pour Constrained Local Neural Field. Cet algorithme fournit en particulier 8 points caractéristiques délimitant le contour interne des lèvres. Appliqué directement à nos données audio-visuelles du locuteur, le CLNF donne de très bons résultats dans environ 70% des cas. Des erreurs subsistent cependant pour le reste des cas. Nous proposons des solutions pour estimer un contour raisonnable des lèvres à partir des points fournis par CLNF utilisant l’interpolation par spline permettant de corriger ses erreurs et d’extraire correctement les paramètres labiaux classiques. Les évaluations sur une base de données de 179 images confirment les performances de notre algorithme.