CyberSec Research

We present Apertus, a fully open suite of large language models (LLMs) designed to address two systemic shortcomings in today's open model ecosystem: data compliance and multilingual representation. Unlike many prior models that release weights without reproducible data pipelines or regard for content-owner rights, Apertus models are pretrained exclusively on openly available data, retroactively respecting robots.txt exclusions and filtering for non-permissive, toxic, and personally identifiable content. To mitigate risks of memorization, we adopt the Goldfish objective during pretraining, strongly suppressing verbatim recall of data while retaining downstream task performance. The Apertus models also expand multilingual coverage, training on 15T tokens from over 1800 languages, with ~40% of pretraining data allocated to non-English content. Released at 8B and 70B scales, Apertus approaches state-of-the-art results among fully open models on multilingual benchmarks, rivalling or surpassing open-weight counterparts. Beyond model weights, we release all scientific artifacts from our development cycle with a permissive license, including data preparation scripts, checkpoints, evaluation suites, and training code, enabling transparent audit and extension.

We show that language models' activations linearly encode when information was learned during training. Our setup involves creating a model with a known training order by sequentially fine-tuning Llama-3.2-1B on six disjoint but otherwise similar datasets about named entities. We find that the average activations of test samples for the six training datasets encode the training order: when projected into a 2D subspace, these centroids are arranged exactly in the order of training and lie on a straight line. Further, we show that linear probes can accurately (~90%) distinguish "early" vs. "late" entities, generalizing to entities unseen during the probes' own training. The model can also be fine-tuned to explicitly report an unseen entity's training stage (~80% accuracy). Interestingly, this temporal signal does not seem attributable to simple differences in activation magnitudes, losses, or model confidence. Our paper demonstrates that models are capable of differentiating information by its acquisition time, and carries significant implications for how they might manage conflicting data and respond to knowledge modifications.

Generative Engine Marketing (GEM) is an emerging ecosystem for monetizing generative engines, such as LLM-based chatbots, by seamlessly integrating relevant advertisements into their responses. At the core of GEM lies the generation and evaluation of ad-injected responses. However, existing benchmarks are not specifically designed for this purpose, which limits future research. To address this gap, we propose GEM-Bench, the first comprehensive benchmark for ad-injected response generation in GEM. GEM-Bench includes three curated datasets covering both chatbot and search scenarios, a metric ontology that captures multiple dimensions of user satisfaction and engagement, and several baseline solutions implemented within an extensible multi-agent framework. Our preliminary results indicate that, while simple prompt-based methods achieve reasonable engagement such as click-through rate, they often reduce user satisfaction. In contrast, approaches that insert ads based on pre-generated ad-free responses help mitigate this issue but introduce additional overhead. These findings highlight the need for future research on designing more effective and efficient solutions for generating ad-injected responses in GEM.

High temporal resolution is essential for capturing fine-grained details in video understanding. However, current video large language models (VLLMs) and benchmarks mostly rely on low-frame-rate sampling, such as uniform sampling or keyframe selection, discarding dense temporal information. This compromise avoids the high cost of tokenizing every frame, which otherwise leads to redundant computation and linear token growth as video length increases. While this trade-off works for slowly changing content, it fails for tasks like lecture comprehension, where information appears in nearly every frame and requires precise temporal alignment. To address this gap, we introduce Dense Video Understanding (DVU), which enables high-FPS video comprehension by reducing both tokenization time and token overhead. Existing benchmarks are also limited, as their QA pairs focus on coarse content changes. We therefore propose DIVE (Dense Information Video Evaluation), the first benchmark designed for dense temporal reasoning. To make DVU practical, we present Gated Residual Tokenization (GRT), a two-stage framework: (1) Motion-Compensated Inter-Gated Tokenization uses pixel-level motion estimation to skip static regions during tokenization, achieving sub-linear growth in token count and compute. (2) Semantic-Scene Intra-Tokenization Merging fuses tokens across static regions within a scene, further reducing redundancy while preserving dynamic semantics. Experiments on DIVE show that GRT outperforms larger VLLM baselines and scales positively with FPS. These results highlight the importance of dense temporal information and demonstrate that GRT enables efficient, scalable high-FPS video understanding.

We present a large-scale computational analysis of migration-related discourse in UK parliamentary debates spanning over 75 years and compare it with US congressional discourse. Using open-weight LLMs, we annotate each statement with high-level stances toward migrants and track the net tone toward migrants across time and political parties. For the UK, we extend this with a semi-automated framework for extracting fine-grained narrative frames to capture nuances of migration discourse. Our findings show that, while US discourse has grown increasingly polarised, UK parliamentary attitudes remain relatively aligned across parties, with a persistent ideological gap between Labour and the Conservatives, reaching its most negative level in 2025. The analysis of narrative frames in the UK parliamentary statements reveals a shift toward securitised narratives such as border control and illegal immigration, while longer-term integration-oriented frames such as social integration have declined. Moreover, discussions of national law about immigration have been replaced over time by international law and human rights, revealing nuances in discourse trends. Taken together broadly, our findings demonstrate how LLMs can support scalable, fine-grained discourse analysis in political and historical contexts.

Personalized financial advice requires consideration of user goals, constraints, risk tolerance, and jurisdiction. Prior LLM work has focused on support systems for investors and financial planners. Simultaneously, numerous recent studies examine broader personal finance tasks, including budgeting, debt management, retirement, and estate planning, through agentic pipelines that incur high maintenance costs, yielding less than 25% of their expected financial returns. In this study, we introduce a novel and reproducible framework that integrates relevant financial context with behavioral finance studies to construct supervision data for end-to-end advisors. Using this framework, we create a 19k sample reasoning dataset and conduct a comprehensive fine-tuning of the Qwen-3-8B model on the dataset. Through a held-out test split and a blind LLM-jury study, we demonstrate that through careful data curation and behavioral integration, our 8B model achieves performance comparable to significantly larger baselines (14-32B parameters) across factual accuracy, fluency, and personalization metrics while incurring 80% lower costs than the larger counterparts.

Recent advancements in multimodal large language models (MLLMs) have garnered significant attention, offering a promising pathway toward artificial general intelligence (AGI). Among the essential capabilities required for AGI, creativity has emerged as a critical trait for MLLMs, with association serving as its foundation. Association reflects a model' s ability to think creatively, making it vital to evaluate and understand. While several frameworks have been proposed to assess associative ability, they often overlook the inherent ambiguity in association tasks, which arises from the divergent nature of associations and undermines the reliability of evaluations. To address this issue, we decompose ambiguity into two types-internal ambiguity and external ambiguity-and introduce AssoCiAm, a benchmark designed to evaluate associative ability while circumventing the ambiguity through a hybrid computational method. We then conduct extensive experiments on MLLMs, revealing a strong positive correlation between cognition and association. Additionally, we observe that the presence of ambiguity in the evaluation process causes MLLMs' behavior to become more random-like. Finally, we validate the effectiveness of our method in ensuring more accurate and reliable evaluations. See Project Page for the data and codes.

We present CS-FLEURS, a new dataset for developing and evaluating code-switched speech recognition and translation systems beyond high-resourced languages. CS-FLEURS consists of 4 test sets which cover in total 113 unique code-switched language pairs across 52 languages: 1) a 14 X-English language pair set with real voices reading synthetically generated code-switched sentences, 2) a 16 X-English language pair set with generative text-to-speech 3) a 60 {Arabic, Mandarin, Hindi, Spanish}-X language pair set with the generative text-to-speech, and 4) a 45 X-English lower-resourced language pair test set with concatenative text-to-speech. Besides the four test sets, CS-FLEURS also provides a training set with 128 hours of generative text-to-speech data across 16 X-English language pairs. Our hope is that CS-FLEURS helps to broaden the scope of future code-switched speech research. Dataset link: https://huggingface.co/datasets/byan/cs-fleurs.

While virtual reality (VR) excels at simulating physical environments, its effectiveness for training complex interpersonal skills is limited by a lack of psychologically plausible virtual humans. This is a critical gap in high-stakes domains like medical education, where communication is a core competency. This paper introduces a framework that integrates large language models (LLMs) into immersive VR to create medically coherent virtual patients with distinct, consistent personalities, built on a modular architecture that decouples personality from clinical data. We evaluated our system in a mixed-method, within-subjects study with licensed physicians who engaged in simulated consultations. Results demonstrate that the approach is not only feasible but is also perceived by physicians as a highly rewarding and effective training enhancement. Furthermore, our analysis uncovers critical design principles, including a ``realism-verbosity paradox" where less communicative agents can seem more artificial, and the need for challenges to be perceived as authentic to be instructive. This work provides a validated framework and key insights for developing the next generation of socially intelligent VR training environments.

This report introduces Canary-1B-v2, a fast, robust multilingual model for Automatic Speech Recognition (ASR) and Speech-to-Text Translation (AST). Built with a FastConformer encoder and Transformer decoder, it supports 25 languages primarily European. The model was trained on 1.7M hours of total data samples, including Granary and NeMo ASR Set 3.0, with non-speech audio added to reduce hallucinations for ASR and AST. We describe its two-stage pre-training and fine-tuning process with dynamic data balancing, as well as experiments with an nGPT encoder. Results show nGPT scales well with massive data, while FastConformer excels after fine-tuning. For timestamps, Canary-1B-v2 uses the NeMo Forced Aligner (NFA) with an auxiliary CTC model, providing reliable segment-level timestamps for ASR and AST. Evaluations show Canary-1B-v2 outperforms Whisper-large-v3 on English ASR while being 10x faster, and delivers competitive multilingual ASR and AST performance against larger models like Seamless-M4T-v2-large and LLM-based systems. We also release Parakeet-TDT-0.6B-v3, a successor to v2, offering multilingual ASR across the same 25 languages with just 600M parameters.

Chain-of-Thought (CoT) reasoning enhances Large Language Models (LLMs) by prompting intermediate steps, improving accuracy and robustness in arithmetic, logic, and commonsense tasks. However, this benefit comes with high computational costs: longer outputs increase latency, memory usage, and KV-cache demands. These issues are especially critical in software engineering tasks where concise and deterministic outputs are required. To investigate these trade-offs, we conduct an empirical study based on code generation benchmarks. The results reveal that longer CoT does not always help. Excessive reasoning often causes truncation, accuracy drops, and latency up to five times higher, with failed outputs consistently longer than successful ones. These findings challenge the assumption that longer reasoning is inherently better and highlight the need for adaptive CoT control. Motivated by this, we propose SEER (Self-Enhancing Efficient Reasoning), an adaptive framework that compresses CoT while preserving accuracy. SEER combines Best-of-N sampling with task-aware adaptive filtering, dynamically adjusting thresholds based on pre-inference outputs to reduce verbosity and computational overhead. We then evaluate SEER on three software engineering tasks and one math task. On average, SEER shortens CoT by 42.1%, improves accuracy by reducing truncation, and eliminates most infinite loops. These results demonstrate SEER as a practical method to make CoT-enhanced LLMs more efficient and robust, even under resource constraints.

Modern mobile CPU software pose challenges for conventional instruction cache replacement policies due to their complex runtime behavior causing high reuse distance between executions of the same instruction. Mobile code commonly suffers from large amounts of stalls in the CPU frontend and thus starvation of the rest of the CPU resources. Complexity of these applications and their code footprint are projected to grow at a rate faster than available on-chip memory due to power and area constraints, making conventional hardware-centric methods for managing instruction caches to be inadequate. We present a novel software-hardware co-design approach called TRRIP (Temperature-based Re-Reference Interval Prediction) that enables the compiler to analyze, classify, and transform code based on "temperature" (hot/cold), and to provide the hardware with a summary of code temperature information through a well-defined OS interface based on using code page attributes. TRRIP's lightweight hardware extension employs code temperature attributes to optimize the instruction cache replacement policy resulting in the eviction rate reduction of hot code. TRRIP is designed to be practical and adoptable in real mobile systems that have strict feature requirements on both the software and hardware components. TRRIP can reduce the L2 MPKI for instructions by 26.5% resulting in geomean speedup of 3.9%, on top of RRIP cache replacement running mobile code already optimized using PGO.

Sign Language Translation (SLT) bridges the communication gap between deaf people and hearing people, where dialogue provides crucial contextual cues to aid in translation. Building on this foundational concept, this paper proposes Question-based Sign Language Translation (QB-SLT), a novel task that explores the efficient integration of dialogue. Unlike gloss (sign language transcription) annotations, dialogue naturally occurs in communication and is easier to annotate. The key challenge lies in aligning multimodality features while leveraging the context of the question to improve translation. To address this issue, we propose a cross-modality Self-supervised Learning with Sigmoid Self-attention Weighting (SSL-SSAW) fusion method for sign language translation. Specifically, we employ contrastive learning to align multimodality features in QB-SLT, then introduce a Sigmoid Self-attention Weighting (SSAW) module for adaptive feature extraction from question and sign language sequences. Additionally, we leverage available question text through self-supervised learning to enhance representation and translation capabilities. We evaluated our approach on newly constructed CSL-Daily-QA and PHOENIX-2014T-QA datasets, where SSL-SSAW achieved SOTA performance. Notably, easily accessible question assistance can achieve or even surpass the performance of gloss assistance. Furthermore, visualization results demonstrate the effectiveness of incorporating dialogue in improving translation quality.

Generative Large Language Models (LLMs) have demonstrated remarkable performance across a wide range of tasks. Recent research has introduced Multi-Agent Debate (MAD) systems, which leverage multiple LLMs to simulate human debate and thereby improve task performance. However, while some LLMs may possess superior knowledge or reasoning capabilities for specific tasks, they often struggle to clearly communicate this advantage during debates, in part due to a lack of confidence expression. Moreover, inappropriate confidence expression can cause agents in MAD systems to either stubbornly maintain incorrect beliefs or converge prematurely on suboptimal answers, ultimately reducing debate effectiveness and overall system performance. To address these challenges, we propose incorporating confidence expression into MAD systems to allow LLMs to explicitly communicate their confidence levels. To validate this approach, we develop ConfMAD, a MAD framework that integrates confidence expression throughout the debate process. Experimental results demonstrate the effectiveness of our method, and we further analyze how confidence influences debate dynamics, offering insights into the design of confidence-aware MAD systems.

Achieving human-level translations requires leveraging context to ensure coherence and handle complex phenomena like pronoun disambiguation. Sparsity of contextually rich examples in the standard training data has been hypothesized as the reason for the difficulty of context utilization. In this work, we systematically validate this claim in both single- and multilingual settings by constructing training datasets with a controlled proportions of contextually relevant examples. We demonstrate a strong association between training data sparsity and model performance confirming sparsity as a key bottleneck. Importantly, we reveal that improvements in one contextual phenomenon do no generalize to others. While we observe some cross-lingual transfer, it is not significantly higher between languages within the same sub-family. Finally, we propose and empirically evaluate two training strategies designed to leverage the available data. These strategies improve context utilization, resulting in accuracy gains of up to 6 and 8 percentage points on the ctxPro evaluation in single- and multilingual settings respectively.

Machine Translation (MT) has achieved remarkable performance, with growing interest in speech translation and multimodal approaches. However, despite these advancements, MT quality assessment remains largely text centric, typically relying on human experts who read and compare texts. Since many real-world MT applications (e.g Google Translate Voice Mode, iFLYTEK Translator) involve translation being spoken rather printed or read, a more natural way to assess translation quality would be through speech as opposed text-only evaluations. This study compares text-only and audio-based evaluations of 10 MT systems from the WMT General MT Shared Task, using crowd-sourced judgments collected via Amazon Mechanical Turk. We additionally, performed statistical significance testing and self-replication experiments to test reliability and consistency of audio-based approach. Crowd-sourced assessments based on audio yield rankings largely consistent with text only evaluations but, in some cases, identify significant differences between translation systems. We attribute this to speech richer, more natural modality and propose incorporating speech-based assessments into future MT evaluation frameworks.

We present Hala, a family of Arabic-centric instruction and translation models built with our translate-and-tune pipeline. We first compress a strong AR$\leftrightarrow$EN teacher to FP8 (yielding $\sim$2$\times$ higher throughput with no quality loss) and use it to create high-fidelity bilingual supervision. A lightweight language model LFM2-1.2B is then fine-tuned on this data and used to translate high-quality English instruction sets into Arabic, producing a million-scale corpus tailored to instruction following. We train Hala models at 350M, 700M, 1.2B, and 9B parameters, and apply slerp merging to balance Arabic specialization with base-model strengths. On Arabic-centric benchmarks, Hala achieves state-of-the-art results within both the "nano" ($\leq$2B) and "small" (7-9B) categories, outperforming their bases. We release models, data, evaluation, and recipes to accelerate research in Arabic NLP.

Reasoning large language models (LLMs) have demonstrated superior capacities in solving complicated problems by generating long chain-of-thoughts (CoT), but such a lengthy CoT incurs high inference costs. In this study, we introduce ES-CoT, an inference-time method that shortens CoT generation by detecting answer convergence and stopping early with minimal performance loss. At the end of each reasoning step, we prompt the LLM to output its current final answer, denoted as a step answer. We then track the run length of consecutive identical step answers as a measure of answer convergence. Once the run length exhibits a sharp increase and exceeds a minimum threshold, the generation is terminated. We provide both empirical and theoretical support for this heuristic: step answers steadily converge to the final answer, and large run-length jumps reliably mark this convergence. Experiments on five reasoning datasets across three LLMs show that ES-CoT reduces the number of inference tokens by about 41\% on average while maintaining accuracy comparable to standard CoT. Further, ES-CoT integrates seamlessly with self-consistency prompting and remains robust across hyperparameter choices, highlighting it as a practical and effective approach for efficient reasoning.