Company
Mistral AI
AI lab building frontier and open-weight language models for enterprise and developer use.
Recent coverage
Related events
Related papers
arXiv cs.AI
MathAgent: Adversarial Evolution of Constraint Graphs for Mathematical Reasoning Data Synthesis
Synthesizing high-quality mathematical reasoning data without human priors remains a significant challenge. Current approaches typically rely on seed data mutation or simple prompt engineering, often suffering from mode collapse and limited logical complexity. This paper proposes a hierarchical synthesis framework that formulates data synthesis as an unsupervised optimization problem over a constraint graph followed by semantic instantiation, rather than treating it as a direct text generation task. We introduce a Legislator-Executor paradigm: The Legislator adversarially evolves structured generation blueprints encoding the constraints of the problem, while the Executor instantiates these specifications into diverse natural language scenarios. This decoupling of skeleton design from linguistic realization enables a prioritized focus on constructing complex and diverse logical structures, thereby guiding high-quality data synthesis. Experiments conducted on a total of 10 models across the Qwen, Llama, Mistral, and Gemma series demonstrate that our method achieves notable results: models fine-tuned on 1K synthesized samples outperform widely-used datasets of comparable scale (LIMO, s1K) across eight mathematical benchmarks, exhibiting superior out-of-distribution generalization.
arXiv cs.AI
Shared Emotion Geometry Across Small Language Models: A Cross-Architecture Study of Representation, Behavior, and Methodological Confounds
We extract 21-emotion vector sets from twelve small language models (six architectures x base/instruct, 1B-8B parameters) under a unified comprehension-mode pipeline at fp16 precision, and compare the resulting geometries via representational similarity analysis on raw cosine RDMs. The five mature architectures (Qwen 2.5 1.5B, SmolLM2 1.7B, Llama 3.2 3B, Mistral 7B v0.3, Llama 3.1 8B) share nearly identical 21-emotion geometry, with pairwise RDM Spearman correlations of 0.74-0.92. This universality persists across diametrically opposed behavioral profiles: Qwen 2.5 and Llama 3.2 occupy opposite poles of MTI Compliance facets yet produce nearly identical emotion RDMs (rho = 0.81), so behavioral facet differences arise above the shared emotion representation. Gemma-3 1B base, the one immature case in our dataset, exhibits extreme residual-stream anisotropy (0.997) and is restructured by RLHF across all geometric descriptors, whereas the five already-mature families show within-family base x instruct RDM correlations of rho >= 0.92 (Mistral 7B v0.3 at rho = 0.985), suggesting RLHF restructures only representations that are not yet organized. Methodologically, we show that what prior work has read as a single comprehension-vs-generation method effect in fact decomposes into four distinct layers -- a coarse method-dependent dissociation, robust sub-parameter sensitivity within generation, a true precision (fp16 vs INT8) effect, and a conflated cross-experiment bias that distorts in opposite directions for different models -- so that a single rho between two prior emotion-vector studies is not a safe basis for interpretation without the layered decomposition.
arXiv cs.AI
Quantization Dominates Rank Reduction for KV-Cache Compression
We compare two strategies for compressing the KV cache in transformer inference: rank reduction (discard dimensions) and quantization (keep all dimensions, reduce precision). At matched storage budgets across five models (124M-14B, MHA and GQA), we find that quantization consistently outperforms rank reduction by 4-364 PPL depending on model and compression level. The gap persists even when rank reduction is combined with quantization in hybrid baselines, and it grows with GQA aggressiveness. On LAMBADA, INT4 matches FP16 accuracy (+0.23 PPL on Mistral 7B, +0.58 on GPT-2) while rank-32 at identical storage collapses to 0.4%. We trace this gap to a structural asymmetry: under softmax attention routing, removing a dimension can flip which token is attended (a discrete failure), while quantization noise is bounded and typically preserves score ordering. We formalize this via a perturbation result showing projection damage exceeds quantization damage by 3 x 2^(2b) per direction under the softmax Fisher metric. A basis ablation confirms the finding is basis-independent (spread <0.4 PPL), establishing that the advantage comes from preserving dimensions, not from a better coordinate system. Joint K+V INT4 quantization achieves 75% total KV reduction at only +0.18 PPL on Mistral 7B.
arXiv cs.AI
Advancing Polish Language Modeling through Tokenizer Optimization in the Bielik v3 7B and 11B Series
The development of the Bielik v3 PL series, encompassing both the 7B and 11B parameter variants, represents a significant milestone in the field of language-specific large language model (LLM) optimization. While general-purpose models often demonstrate impressive multilingual capabilities, they frequently suffer from a fundamental architectural inefficiency: the use of universal tokenizers. These tokenizers, typically designed to cover a broad spectrum of languages, often fail to capture the morphological nuances of specific languages like Polish, leading to higher fertility ratios, increased inference costs, and restricted effective context windows. This report details the transition from the universal Mistral-based tokenization to a dedicated Polish-optimized vocabulary for the Bielik v3 models, exploring the FOCUS-based embedding initialization, the multi-stage pretraining curriculum, and the subsequent post-training alignment involving Supervised Fine-Tuning, Direct Preference Optimization, and Reinforcement Learning through Group Relative Policy Optimization with verifiable rewards.
arXiv cs.AI
Dead Weights, Live Signals: Feedforward Graphs of Frozen Language Models
We present a feedforward graph architecture in which heterogeneous frozen large language models serve as computational nodes, communicating through a shared continuous latent space via learned linear projections. Building on recent work demonstrating geometric compatibility between independently trained LLM latent spaces~\cite{armstrong2026thinking}, we extend this finding from static two-model steering to end-to-end trainable multi-node graphs, where projection matrices are optimized jointly via backpropagation through residual stream injection hooks. Three small frozen models (Llama-3.2-1B, Qwen2.5-1.5B, Gemma-2-2B) encode the input into a shared latent space whose aggregate signal is injected into two larger frozen models (Phi-3-mini, Mistral-7B), whose representations feed a lightweight cross-attention output node. With only 17.6M trainable parameters against approximately 12B frozen, the architecture achieves 87.3\% on ARC-Challenge, 82.8\% on OpenBookQA, and 67.2\% on MMLU, outperforming the best single constituent model by 11.4, 6.2, and 1.2 percentage points respectively, and outperforming parameter-matched learned classifiers on frozen single models by 9.1, 5.2, and 6.7 points. Gradient flow through multiple frozen model boundaries is empirically verified to be tractable, and the output node develops selective routing behavior across layer-2 nodes without explicit supervision.
arXiv cs.AI
The AI Skills Shift: Mapping Skill Obsolescence, Emergence, and Transition Pathways in the LLM Era
As Large Language Models reshape the global labor market, policymakers and workers need empirical data on which occupational skills may be most susceptible to automation. We present the Skill Automation Feasibility Index (SAFI), benchmarking four frontier LLMs -- LLaMA 3.3 70B, Mistral Large, Qwen 2.5 72B, and Gemini 2.5 Flash -- across 263 text-based tasks spanning all 35 skills in the U.S. Department of Labor's O*NET taxonomy (1,052 total model calls, 0% failure rate). Cross-referencing with real-world AI adoption data from the Anthropic Economic Index (756 occupations, 17,998 tasks), we propose an AI Impact Matrix -- an interpretive framework that positions skills along four quadrants: High Displacement Risk, Upskilling Required, AI-Augmented, and Lower Displacement Risk. Key findings: (1) Mathematics (SAFI: 73.2) and Programming (71.8) receive the highest automation feasibility scores; Active Listening (42.2) and Reading Comprehension (45.5) receive the lowest; (2) a "capability-demand inversion" where skills most demanded in AI-exposed jobs are those LLMs perform least well at in our benchmark; (3) 78.7% of observed AI interactions are augmentation, not automation; (4) all four models converge to similar skill profiles (3.6-point spread), suggesting that text-based automation feasibility may be more skill-dependent than model-dependent. SAFI measures LLM performance on text-based representations of skills, not full occupational execution. All data, code, and model responses are open-sourced.
arXiv cs.AI
Digital Skin, Digital Bias: Uncovering Tone-Based Biases in LLMs and Emoji Embeddings
Skin-toned emojis are crucial for fostering personal identity and social inclusion in online communication. As AI models, particularly Large Language Models (LLMs), increasingly mediate interactions on web platforms, the risk that these systems perpetuate societal biases through their representation of such symbols is a significant concern. This paper presents the first large-scale comparative study of bias in skin-toned emoji representations across two distinct model classes. We systematically evaluate dedicated emoji embedding models (emoji2vec, emoji-sw2v) against four modern LLMs (Llama, Gemma, Qwen, and Mistral). Our analysis first reveals a critical performance gap: while LLMs demonstrate robust support for skin tone modifiers, widely-used specialized emoji models exhibit severe deficiencies. More importantly, a multi-faceted investigation into semantic consistency, representational similarity, sentiment polarity, and core biases uncovers systemic disparities. We find evidence of skewed sentiment and inconsistent meanings associated with emojis across different skin tones, highlighting latent biases within these foundational models. Our findings underscore the urgent need for developers and platforms to audit and mitigate these representational harms, ensuring that AI's role on the web promotes genuine equity rather than reinforcing societal biases.
arXiv cs.AI
LatentAudit: Real-Time White-Box Faithfulness Monitoring for Retrieval-Augmented Generation with Verifiable Deployment
Retrieval-augmented generation (RAG) mitigates hallucination but does not eliminate it: a deployed system must still decide, at inference time, whether its answer is actually supported by the retrieved evidence. We introduce LatentAudit, a white-box auditor that pools mid-to-late residual-stream activations from an open-weight generator and measures their Mahalanobis distance to the evidence representation. The resulting quadratic rule requires no auxiliary judge model, runs at generation time, and is simple enough to calibrate on a small held-out set. We show that residual-stream geometry carries a usable faithfulness signal, that this signal survives architecture changes and realistic retrieval failures, and that the same rule remains amenable to public verification. On PubMedQA with Llama-3-8B, LatentAudit reaches 0.942 AUROC with 0.77,ms overhead. Across three QA benchmarks and five model families (Llama-2/3, Qwen-2.5/3, Mistral), the monitor remains stable; under a four-way stress test with contradictions, retrieval misses, and partial-support noise, it reaches 0.9566--0.9815 AUROC on PubMedQA and 0.9142--0.9315 on HotpotQA. At 16-bit fixed-point precision, the audit rule preserves 99.8% of the FP16 AUROC, enabling Groth16-based public verification without revealing model weights or activations. Together, these results position residual-stream geometry as a practical basis for real-time RAG faithfulness monitoring and optional verifiable deployment.
arXiv cs.AI
MARL-GPT: Foundation Model for Multi-Agent Reinforcement Learning
Recent advances in multi-agent reinforcement learning (MARL) have demonstrated success in numerous challenging domains and environments, but typically require specialized models for each task. In this work, we propose a coherent methodology that makes it possible for a single GPT-based model to learn and perform well across diverse MARL environments and tasks, including StarCraft Multi-Agent Challenge, Google Research Football and POGEMA. Our method, MARL-GPT, applies offline reinforcement learning to train at scale on the expert trajectories (400M for SMACv2, 100M for GRF, and 1B for POGEMA) combined with a single transformer-based observation encoder that requires no task-specific tuning. Experiments show that MARL-GPT achieves competitive performance compared to specialized baselines in all tested environments. Thus, our findings suggest that it is, indeed, possible to build a multi-task transformer-based model for a wide variety of (significantly different) multi-agent problems paving the way to the fundamental MARL model (akin to ChatGPT, Llama, Mistral etc. in natural language modeling).