Large language models (LLMs) are increasingly used across research and industry applications, yet their inference efficiency remains a significant challenge. As the computational power of modern GPU architectures continuously improves, their memory bandwidth and capacity have not scaled proportionally, creating a critical bottleneck during inference. To address this, we investigate ternary language models (TriLMs) that employ quantization-aware training to significantly reduce memory requirements. We first analyze the scalability of TriLMs by conducting a scaling law analysis, revealing that TriLMs benefit more from increasing training data than from scaling model parameters. Based on this observation, we introduce Spectra-1.1, an open suite of TriLMs trained on up to 1.2 trillion tokens, demonstrating sustained performance gains at scale. Furthermore, to improve inference efficiency, we propose novel 2-bit and 1.6-bit packing schemes for ternary weights, which demonstrate accelerated inference across various CPU architectures. Also, building on the 2-bit packing, we develop a GPU kernel called TriRun that accelerates end-to-end model inference by up to 5 times compared to floating-point baselines. To encourage further exploration and development of TriLMs, we will release the Spectra-1.1 suite and TriRun inference kernels. Overall, our work lays the foundation for building and deploying efficient LLMs, providing a valuable resource for the research community.

The ability to process information from multiple modalities and to reason through it step-by-step remains a critical challenge in advancing artificial intelligence. However, existing reasoning benchmarks focus on text-only reasoning, or employ multimodal questions that can be answered by directly retrieving information from a non-text modality. Thus, complex reasoning remains poorly understood in multimodal domains. Here, we present MARBLE, a challenging multimodal reasoning benchmark that is designed to scrutinize multimodal language models (MLLMs) in their ability to carefully reason step-by-step through complex multimodal problems and environments. MARBLE is composed of two highly challenging tasks, M-Portal and M-Cube, that require the crafting and understanding of multistep plans under spatial, visual, and physical constraints. We find that current MLLMs perform poorly on MARBLE -- all the 12 advanced models obtain near-random performance on M-Portal and 0% accuracy on M-Cube. Only in simplified subtasks some models outperform the random baseline, indicating that complex reasoning is still a challenge for existing MLLMs. Moreover, we show that perception remains a bottleneck, where MLLMs occasionally fail to extract information from the visual inputs. By shedding a light on the limitations of MLLMs, we hope that MARBLE will spur the development of the next generation of models with the ability to reason and plan across many, multimodal reasoning steps.

Existing video segmenter and grounder approaches, exemplified by Sa2VA, directly fuse features within segmentation models. This often results in an undesirable entanglement of dynamic visual information and static semantics, thereby degrading segmentation accuracy. To systematically mitigate this issue, we propose DeSa2VA, a decoupling-enhanced prompting scheme integrating text pre-training and a linear decoupling module to address the information processing limitations inherent in SAM-2. Specifically, first, we devise a pre-training paradigm that converts textual ground-truth labels into point-level prompts while generating corresponding text masks. These masks are refined through a hybrid loss function to strengthen the model's semantic grounding capabilities. Next, we employ linear projection to disentangle hidden states that generated by a large language model into distinct textual and visual feature subspaces. Finally, a dynamic mask fusion strategy synergistically combines these decoupled features through triple supervision from predicted text/visual masks and ground-truth annotations. Extensive experiments demonstrate state-of-the-art performance across diverse tasks, including image segmentation, image question answering, video segmentation, and video question answering. Our codes are available at https://github.com/longmalongma/DeSa2VA.

Large language models (LLMs) have recently been applied to dialog systems. Despite making progress, LLMs are prone to errors in knowledge-intensive scenarios. Recently, approaches based on retrieval augmented generation (RAG) and agent have emerged to improve the factual accuracy by enhancing the LLMs with knowledge retrieved from external knowledge bases (KBs). This is mostly implemented by prompting the LLMs with instructions, examples and the retrieved knowledge. However, LLMs may have difficulty using the retrieved knowledge effectively for response generation, because they are not well trained to do such generation for specific domains. To mitigate this problem, we propose to finetune the LLMs in the RAG-based and agent-based systems with domain-specific data, together with domain-specific external knowledge, which is called knowledge augmented finetuning (KAFT). We base our study on the MobileCS2 dataset, a real-life customer service dialog dataset that features intensive knowledge interactions, to systematically compare the prompting and KAFT techniques in the RAG-based and agent-based systems. Experiment results show that KAFT substantially surpasses prompting in both RAG and agent systems, particularly in terms of factual accuracy. To the best of our knowledge, this paper represents the first solid empirical work to investigate the KAFT idea.

--Large Language Models (LLMs) like GPT -3.5-T urbo are increasingly used to assist software development, yet they often produce incomplete code or incorrect imports, especially when lacking access to external or project-specific documentation. We introduce RAILS (Retrieval-Augmented Intelligence for Learning Software Development), a framework that augments LLM prompts with semantically retrieved context from curated Java resources using F AISS and OpenAI embeddings. RAILS incorporates an iterative validation loop guided by compiler feedback to refine suggestions. We evaluated RAILS on 78 real-world Java import error cases spanning standard libraries, GUI APIs, external tools, and custom utilities. Despite using the same LLM, RAILS outperforms baseline prompting by preserving intent, avoiding hallucinations, and surfacing correct imports even when libraries are unavailable locally. Future work will integrate symbolic filtering via PostgreSQL and extend support to other languages and IDEs.

We present our submission to the LiveRAG Challenge 2025, which evaluates retrieval-augmented generation (RAG) systems on dynamic test sets using the FineWeb-10BT corpus. Our final hybrid approach combines sparse (BM25) and dense (E5) retrieval methods and then aims to generate relevant and faithful answers with Falcon3-10B-Instruct. Through systematic evaluation on 200 synthetic questions generated with DataMorgana across 64 unique question-user combinations, we demonstrate that neural re-ranking with RankLLaMA improves MAP from 0.523 to 0.797 (52% relative improvement) but introduces prohibitive computational costs (84s vs 1.74s per question). While DSPy-optimized prompting strategies achieved higher semantic similarity (0.771 vs 0.668), their 0% refusal rates raised concerns about over-confidence and generalizability. Our submitted hybrid system without re-ranking achieved 4th place in faithfulness and 11th place in correctness among 25 teams. Analysis across question categories reveals that vocabulary alignment between questions and documents was the strongest predictor of performance on our development set, with document-similar phrasing improving cosine similarity from 0.562 to 0.762.

--This paper introduces a new framework, Drive-BLIP2, built upon the BLIP2-OPT architecture, to generate accurate and contextually relevant explanations for emerging driving scenarios. While existing vision-language models perform well in general tasks, they encounter difficulties in understanding complex, multi-object environments, particularly in real-time applications such as autonomous driving, where the rapid identification of key objects is crucial. T o address this limitation, an Attention Map Generator is proposed to highlight significant objects relevant to driving decisions within critical video frames. By directing the model's focus to these key regions, the generated attention map helps produce clear and relevant explanations, enabling drivers to better understand the vehicle's decision-making process in critical situations. Evaluations on the DRAMA dataset reveal significant improvements in explanation quality, as indicated by higher BLEU, ROUGE, CIDEr, and SPICE scores compared to baseline models. These findings underscore the potential of targeted attention mechanisms in vision-language models for enhancing explainability in real-time autonomous driving. In recent years, large language models (LLMs) have made significant progress in natural language processing, powering a wide range of applications from machine translation to text generation. These models, such as GPT [1], [21] and BERT [2], have demonstrated remarkable abilities to understand and generate human-like text for a wide range of tasks.

-- The android robot Andrea was set up at a public museum in Germany for six consecutive days to have conversations with visitors, fully autonomously. No specific context was given, so visitors could state their opinions regarding possible use-cases in structured interviews, without any bias. Additionally the 44 interviewees were asked for their general opinions of the robot, their reasons (not) to interact with it and necessary improvements for future use. The android's voice and wig were changed between different days of operation to give varying cues regarding its gender . This did not have a significant impact on the positive overall perception of the robot. Most visitors want the robot to provide information about exhibits in the future, while opinions on other roles, like a receptionist, were both wanted and explicitly not wanted by different visitors. Speaking more languages (than only English) and faster response times were the improvements most desired. These findings from the interviews are in line with an analysis of the system logs, which revealed, that after chitchat and personal questions, most of the 4436 collected requests asked for information related to the museum and to converse in a different language. The valuable insights gained from these real-world interactions are now used to improve the system to become a useful real-world application. An android robot's outer appearance is explicitly designed to resemble a human as closely as possible.

Large Language Models (LLMs) have shown impressive performance on a range of educational tasks, but are still understudied for their potential to solve mathematical problems. In this study, we compare three prominent LLMs, including GPT-4o, DeepSeek-V3, and Gemini-2.0, on three mathematics datasets of varying complexities (GSM8K, MATH500, and MIT Open Courseware datasets). We take a five-dimensional approach based on the Structured Chain-of-Thought (SCoT) framework to assess final answer correctness, step completeness, step validity, intermediate calculation accuracy, and problem comprehension. The results show that GPT-4o is the most stable and consistent in performance across all the datasets, but particularly it performs outstandingly in high-level questions of the MIT Open Courseware dataset. DeepSeek-V3 is competitively strong in well-structured domains such as optimisation, but suffers from fluctuations in accuracy in statistical inference tasks. Gemini-2.0 shows strong linguistic understanding and clarity in well-structured problems but performs poorly in multi-step reasoning and symbolic logic. Our error analysis reveals particular deficits in each model: GPT-4o is at times lacking in sufficient explanation or precision; DeepSeek-V3 leaves out intermediate steps; and Gemini-2.0 is less flexible in mathematical reasoning in higher dimensions.

Apple is considering using artificial intelligence technology from Anthropic or OpenAI to power a new version of Siri, sidelining its own in-house models in a potentially blockbuster move aimed at turning around its flailing AI effort. The iPhone maker has talked with both companies about using their large language models for Siri, according to people familiar with the discussions. It has asked them to train versions of their models that could run on Apple's cloud infrastructure for testing, said the people, who asked not to be identified discussing private deliberations. If Apple ultimately moves forward, it would represent a monumental reversal. The company currently powers most of its AI features with homegrown technology that it calls Apple Foundation Models and had been planning a new version of its voice assistant that runs on that technology for 2026.