Large Language Models (LLMs) as stochastic systems may generate numbers that deviate from available data, a failure known as \emph{numeric hallucination}. Existing safeguards -- retrieval-augmented generation, citations, and uncertainty estimation -- improve transparency but cannot guarantee fidelity: fabricated or misquoted values may still be displayed as if correct. We propose \textbf{Proof-Carrying Numbers (PCN)}, a presentation-layer protocol that enforces numeric fidelity through mechanical verification. Under PCN, numeric spans are emitted as \emph{claim-bound tokens} tied to structured claims, and a verifier checks each token under a declared policy (e.g., exact equality, rounding, aliases, or tolerance with qualifiers). Crucially, PCN places verification in the \emph{renderer}, not the model: only claim-checked numbers are marked as verified, and all others default to unverified. This separation prevents spoofing and guarantees fail-closed behavior. We formalize PCN and prove soundness, completeness under honest tokens, fail-closed behavior, and monotonicity under policy refinement. PCN is lightweight and model-agnostic, integrates seamlessly into existing applications, and can be extended with cryptographic commitments. By enforcing verification as a mandatory step before display, PCN establishes a simple contract for numerically sensitive settings: \emph{trust is earned only by proof}, while the absence of a mark communicates uncertainty.

Encoder-only languages models are frequently used for a variety of standard machine learning tasks, including classification and retrieval. However, there has been a lack of recent research for encoder models, especially with respect to multilingual models. We introduce mmBERT, an encoder-only language model pretrained on 3T tokens of multilingual text in over 1800 languages. To build mmBERT we introduce several novel elements, including an inverse mask ratio schedule and an inverse temperature sampling ratio. We add over 1700 low-resource languages to the data mix only during the decay phase, showing that it boosts performance dramatically and maximizes the gains from the relatively small amount of training data. Despite only including these low-resource languages in the short decay phase we achieve similar classification performance to models like OpenAI's o3 and Google's Gemini 2.5 Pro. Overall, we show that mmBERT significantly outperforms the previous generation of models on classification and retrieval tasks -- on both high and low-resource languages.

Scaling up test-time compute, by generating multiple independent solutions and selecting or aggregating among them, has become a central paradigm for improving large language models (LLMs) on challenging reasoning tasks. While most prior work relies on simple majority voting or reward model ranking to aggregate solutions, these approaches may only yield limited benefits. In this work, we propose to learn aggregation as an explicit reasoning skill: given a set of candidate solutions, we train an aggregator model to review, reconcile, and synthesize a final, correct answer using reinforcement learning from verifiable rewards. A key ingredient is careful balancing of easy and hard training examples, allowing the model to learn both to recover minority-but-correct answers as well as easy majority-correct answers. Empirically, we find our method, AggLM, outperforms both strong rule-based and reward-model baselines, across multiple benchmarks. Furthermore, it generalizes effectively to solutions from differing models, including stronger ones than contained in the training data, all while requiring substantially fewer tokens than majority voting with larger numbers of solutions.

Test-time scaling increases inference-time computation by allowing models to generate long reasoning chains, and has shown strong performance across many domains. However, in this work, we show that this approach is not yet effective for knowledge-intensive tasks, where high factual accuracy and low hallucination rates are essential. We conduct a comprehensive evaluation of test-time scaling using 12 reasoning models on two knowledge-intensive benchmarks. Our results reveal that increasing test-time computation does not consistently improve accuracy and, in many cases, it even leads to more hallucinations. We then analyze how extended reasoning affects hallucination behavior. We find that reduced hallucinations often result from the model choosing to abstain after thinking more, rather than from improved factual recall. Conversely, for some models, longer reasoning encourages attempts on previously unanswered questions, many of which result in hallucinations. Case studies show that extended reasoning can induce confirmation bias, leading to overconfident hallucinations. Despite these limitations, we observe that compared to non-thinking, enabling thinking remains beneficial. Code and data are available at https://github.com/XuZhao0/tts-knowledge

Large Language Models are increasingly used to simulate human opinion dynamics, yet the effect of genuine interaction is often obscured by systematic biases. We present a Bayesian framework to disentangle and quantify three such biases: (i) a topic bias toward prior opinions in the training data; (ii) an agreement bias favoring agreement irrespective of the question; and (iii) an anchoring bias toward the initiating agent's stance. Applying this framework to multi-step dialogues reveals that opinion trajectories tend to quickly converge to a shared attractor, with the influence of the interaction fading over time, and the impact of biases differing between LLMs. In addition, we fine-tune an LLM on different sets of strongly opinionated statements (incl. misinformation) and demonstrate that the opinion attractor shifts correspondingly. Exposing stark differences between LLMs and providing quantitative tools to compare them to human subjects in the future, our approach highlights both chances and pitfalls in using LLMs as proxies for human behavior.

Recent advancements in image customization exhibit a wide range of application prospects due to stronger customization capabilities. However, since we humans are more sensitive to faces, a significant challenge remains in preserving consistent identity while avoiding identity confusion with multi-reference images, limiting the identity scalability of customization models. To address this, we present UMO, a Unified Multi-identity Optimization framework, designed to maintain high-fidelity identity preservation and alleviate identity confusion with scalability. With "multi-to-multi matching" paradigm, UMO reformulates multi-identity generation as a global assignment optimization problem and unleashes multi-identity consistency for existing image customization methods generally through reinforcement learning on diffusion models. To facilitate the training of UMO, we develop a scalable customization dataset with multi-reference images, consisting of both synthesised and real parts. Additionally, we propose a new metric to measure identity confusion. Extensive experiments demonstrate that UMO not only improves identity consistency significantly, but also reduces identity confusion on several image customization methods, setting a new state-of-the-art among open-source methods along the dimension of identity preserving. Code and model: https://github.com/bytedance/UMO

The scarcity of high-quality, logically sound data is a critical bottleneck for advancing the mathematical reasoning of Large Language Models (LLMs). Our work confronts this challenge by turning decades of automated theorem proving research into a scalable data engine. Rather than relying on error-prone LLMs or complex proof-assistant syntax like Lean and Isabelle, our framework leverages E-prover's saturation capabilities on the vast TPTP axiom library to derive a massive, guaranteed-valid corpus of theorems. Our pipeline is principled and simple: saturate axioms, filter for "interesting" theorems, and generate tasks. With no LLMs in the loop, we eliminate factual errors by construction. This purely symbolic data is then transformed into three difficulty-controlled challenges: entailment verification, premise selection, and proof reconstruction. Our zero-shot experiments on frontier models reveal a clear weakness: performance collapses on tasks requiring deep, structural reasoning. Our framework provides both the diagnostic tool to measure this gap and a scalable source of symbolic training data to address it. We make the code and data publicly available. https://github.com/sileod/reasoning_core https://hf.co/datasets/reasoning-core/rc1

Instruction Fine-Tuning (IFT) has been widely adopted as an effective post-training strategy to enhance various abilities of Large Language Models (LLMs). However, prior studies have shown that IFT can significantly compromise LLMs' safety, particularly their ability to refuse malicious instructions, raising significant concerns. Recent research into the internal mechanisms of LLMs has identified the refusal direction (r-direction) in the hidden states, which plays a pivotal role in governing refusal behavior. Building on this insight, our study reveals that the r-direction tends to drift during training, which we identify as one of the causes of the associated safety risks. To mitigate such drift, our proposed ProCon method introduces a projection-constrained loss term that regularizes the projection magnitude of each training sample's hidden state onto the r-direction. Our initial analysis shows that applying an appropriate constraint can effectively mitigate the refusal direction drift and associated safety risks, but remains limited by overall performance barriers. To overcome this barrier, informed by our observation of early-stage sharp drift and a data-driven perspective, we introduce a warm-up strategy that emphasizes early-stage strong constraints and broaden the data distribution to strengthen constraint signals, leading to an enhanced ProCon method. Experimental results under various datasets, scenarios, and LLMs demonstrate that our method can significantly mitigate safety risks posed by IFT while preserving task performance gains. Even compared with strong baselines, our method consistently delivers superior overall performance. Crucially, our analysis indicates that ProCon can contribute to stabilizing the r-direction during training, while such an interpretability-driven exploration of LLMs' internal mechanisms lays a solid foundation for future safety research.

Large Vision-Language Models (LVLMs) or multimodal large language models represent a significant advancement in artificial intelligence, enabling systems to understand and generate content across both visual and textual modalities. While large-scale pretraining has driven substantial progress, fine-tuning these models for aligning with human values or engaging in specific tasks or behaviors remains a critical challenge. Deep Reinforcement Learning (DRL) and Direct Preference Optimization (DPO) offer promising frameworks for this aligning process. While DRL enables models to optimize actions using reward signals instead of relying solely on supervised preference data, DPO directly aligns the policy with preferences, eliminating the need for an explicit reward model. This overview explores paradigms for fine-tuning LVLMs, highlighting how DRL and DPO techniques can be used to align models with human preferences and values, improve task performance, and enable adaptive multimodal interaction. We categorize key approaches, examine sources of preference data, reward signals, and discuss open challenges such as scalability, sample efficiency, continual learning, generalization, and safety. The goal is to provide a clear understanding of how DRL and DPO contribute to the evolution of robust and human-aligned LVLMs.

Intelligent vehicle cockpits present unique challenges for API Agents, requiring coordination across tightly-coupled subsystems that exceed typical task environments' complexity. Traditional Function Calling (FC) approaches operate statelessly, requiring multiple exploratory calls to build environmental awareness before execution, leading to inefficiency and limited error recovery. We introduce VehicleWorld, the first comprehensive environment for the automotive domain, featuring 30 modules, 250 APIs, and 680 properties with fully executable implementations that provide real-time state information during agent execution. This environment enables precise evaluation of vehicle agent behaviors across diverse, challenging scenarios. Through systematic analysis, we discovered that direct state prediction outperforms function calling for environmental control. Building on this insight, we propose State-based Function Call (SFC), a novel approach that maintains explicit system state awareness and implements direct state transitions to achieve target conditions. Experimental results demonstrate that SFC significantly outperforms traditional FC approaches, achieving superior execution accuracy and reduced latency. We have made all implementation code publicly available on Github https://github.com/OpenMOSS/VehicleWorld.