arXiv subjects

Jordi Ros-Giralt

Publications and source records attributed to Jordi Ros-Giralt.

10 recordsLinked to original sources

arXiv2026

From Chain-of-Thought to Loops: Non-Autoregressive Latent Reasoning via Looped Transformers

Chain-of-thought (CoT) reasoning often improves language-model performance by giving models additional computation before answering. However, explicit CoT expresses this computation as a sequence of autoregressively generated tokens. Latent reasoning replaces these tokens with compact continuous states, but most autoregressive latent-reasoning methods retain a left-to-right dependency among latent vectors. We introduce LLoCoT: a looped latent-reasoning framework that replaces left-to-right latent generation with iterative refinement of a compact latent workspace. A shared transformer is reapplied for a small number of refinement iterations, jointly updating the latent slots based on the prompt and the evolving workspace state. Using the refined state, a probabilistic head predicts a distribution from which latent tokens are sampled in parallel and used to condition an autoregressive decoder for answer generation. Training uses continuous representations derived from explicit CoT together with a final-answer prediction loss and likelihood-based supervision of the latent states. Across HumanEval and MBPP, LLoCoT achieves the highest mean among the evaluated methods, performing on par in accuracy with Reasoning SFT, our explicit-CoT baseline, while outperforming the base model, answer-only SFT and NF-CoT. Relative to Reasoning SFT, LLoCoT reduces time to the first answer token by approximately $36\times$ and reasoning-phase latency by approximately $42\times$, while increasing end-to-end throughput by $9.2\%$. This design replaces serial thought generation with parallel latent-slot refinement while retaining probabilistic latent modeling and autoregressive answer decoding.

cs.AI↗

arXiv2026

Memory-Efficient Looped Transformer: Decoupling Compute from Memory in Looped Language Models

Recurrent LLM architectures have emerged as a promising approach for improving reasoning, as they enable multi-step computation in the embedding space without generating intermediate tokens. Models such as Ouro perform reasoning by iteratively updating internal representations while retaining a standard Key-Value (KV) cache across iterations, causing memory consumption to grow linearly with reasoning depth. Consequently, increasing the number of reasoning iterations can lead to prohibitive memory usage, limiting the practical scalability of such architectures. In this work, we propose Memory-Efficient Looped Transformer (MELT), a novel architecture that decouples reasoning depth from memory consumption. Instead of using a standard KV cache per layer and loop, MELT maintains a single KV cache per layer that is shared across reasoning loops. This cache is updated over time via a learnable gating mechanism. To enable stable and efficient training under this architecture, we propose to train MELT using chunk-wise training in a two phase procedure: interpolated transition, followed by attention-aligned distillation, both from the LoopLM starting model to MELT. Empirically, we show that MELT models fine-tuned from pretrained Ouro parameters outperform standard LLMs of comparable size, while maintaining a memory footprint comparable to those models and dramatically smaller than Ouro's. Overall, MELT achieves constant-memory iterative reasoning without sacrificing LoopLM performance, using only a lightweight post-training procedure.

cs.CL↗

arXiv2026

Optimizing Teacher-Student Partitioning for Scalable Knowledge Distillation on HPC Systems

Knowledge Distillation (KD) enables training smaller student models under the guidance of larger teacher models, and the widely adopted TRL library implements it. Yet, TRL treats both models symmetrically, missing opportunities to exploit their pronounced asymmetry in memory footprint, and communication requirements. This paper presents an HPC-aware methodology for KD that decouples teacher and student partitioning efficiently. Our approach achieves up to 67% higher samples-per-second than TRL by avoiding unnecessary teacher-model data structures and selecting the best split strategy. We combine vertical and horizontal partitioning of models, deriving an analytical expression that identifies the existence of inflection points between splitting regimes. These results showed that exploiting teacher--student asymmetry through topology-aware parallelism notably accelerated GKD training on production HPC clusters at our company

cs.DC↗

arXiv2026

LaneRoPE: Positional Encoding for Collaborative Parallel Reasoning and Generation

Parallel LLM test-time scaling techniques (e.g., best-of-$N$) require drawing $N>1$ sequences conditioned on the same input prompt. These methods boost accuracy while exploiting the computational efficiency of batching $N$ generations. However, each sequence in the batch is traditionally generated independently and hence does not reuse intermediate generations, computations, or observations from other sequences. In this paper, we propose LaneRoPE to enable coordination and collaboration among $N>1$ sequences at generation time. LaneRoPE involves two key ideas: (a) an inter-sequence attention mask to make sampling of sequences dependent on one another; and (b) a RoPE extension that injects positional information that captures relative positions between tokens, both within and outside a particular sequence. We evaluate our approach on mathematical reasoning tasks and find promising results: LaneRoPE enables collaboration among sequences, yielding additional accuracy gains under limited generated sequence length. Importantly, since LaneRoPE enables coordination with minimal changes to the underlying LLM architecture and introduces a negligible overhead at inference time, it is appealing to rapidly incorporate parallel reasoning into existing LLM inference pipelines.

cs.AI↗

arXiv2026

M3Kang: Evaluating Multilingual Multimodal Mathematical Reasoning in Vision-Language Models

Despite state-of-the-art vision-language models (VLMs) have demonstrated strong reasoning capabilities, their performance in multilingual mathematical reasoning remains underexplored, particularly when compared to human performance. To bridge this gap, we introduce M3Kang, the first massively multilingual, multimodal mathematical reasoning dataset for VLMs. It is derived from the Kangaroo Math Competition, the world's largest mathematics contest, which annually engages over six million participants under the age of 18 across more than 90 countries. M3Kang includes 1,747 unique multiple-choice problems organized by grade-level difficulty, with translations into 108 culturally diverse languages, some of them including diagrams essential for solving them. Using this dataset, we conduct extensive benchmarking on both closed- and open-source SOTA models. We observe that, despite recent advances, models still struggle with basic math and diagram-based reasoning, with performance scaling with language presence and model size, but not with grade level. We also find that multilingual techniques can be effectively extended to the multimodal setting, resulting in significant improvements over baseline approaches. Our analysis also incorporates performance data from over 68,000 students, enabling direct comparison with human performance. We are open-sourcing M3Kang, including the English-only subset M2Kang, along with the framework and codebase used to construct the dataset.

cs.CL↗

arXiv2024

Towards Deep Application-Network Integration: Architectures, Progress and Opportunities

With the rise of a new generation of applications (e.g., virtual and augmented reality, artificial intelligence, etc) demanding stringent performance requirements, the need for networking solutions and architectures that can enable a higher Quality of Experience (QoE) is becoming increasingly important. While jointly optimizing application and network may increase the applications' QoE and simultaneously improve the utilization of network resources, such a paradigm has had limited success in real production networks. However, with the combination of revolutionary trends in (1) compute processing demands, (2) networking capabilities, and (3) sustainable business models, it is high time the community explores the full potential of deeper integration between application and network. In this paper, recent trends observed over the past few years are systematically reviewed. These include the paradigm shift in modern communication services towards computing-driven applications, such as on-site AI training, advances in programmable network technologies like Software Defined Networking (SDN), and new business models incentivizing collaboration and cooperation between parties. Following this, successful scenarios that benefit from various forms of deeper network-application integration are reported, highlighting their considerable potential. A unified framework is then introduced, providing an overview of possible architecture paradigms for network-application integration and bringing awareness to existing abstractions, mechanisms, tools, and their potential combinations. The paper concludes with a discussion of several remaining challenges in building practical network-application integrated systems.

cs.NI↗

arXiv2022

A Quantitative Theory of Bottleneck Structures for Data Networks

The conventional view of the congestion control problem in data networks is based on the principle that a flow's performance is uniquely determined by the state of its bottleneck link, regardless of the topological properties of the network. However, recent work has shown that the behavior of congestion-controlled networks is better explained by models that account for the interactions between bottleneck links. These interactions are captured by a latent \textit{bottleneck structure}, a model describing the complex ripple effects that changes in one part of the network exert on the other parts. In this paper, we present a \textit{quantitative} theory of bottleneck structures (QTBS), a mathematical and engineering framework comprising a family of polynomial-time algorithms that can be used to reason about a wide variety of network optimization problems, including routing, capacity planning and flow control. QTBS can contribute to traffic engineering by making clear predictions about the relative performance of alternative flow routes, and by providing numerical recommendations for the optimal rate settings of traffic shapers. A particularly novel result in the domain of capacity planning indicates that previously established rules for the design of folded-Clos networks are suboptimal when flows are congestion controlled. We show that QTBS can be used to derive the optimal rules for this important class of topologies, and empirically demonstrate the correctness and efficacy of these results using the BBR and Cubic congestion-control algorithms.

cs.NI↗

arXiv2018

High Speed Elephant Flow Detection Under Partial Information

In this paper we introduce a new framework to detect elephant flows at very high speed rates and under uncertainty. The framework provides exact mathematical formulas to compute the detection likelihood and introduces a new flow reconstruction lemma under partial information. These theoretical results lead to the design of BubbleCache, a new elephant flow detection algorithm designed to operate near the optimal tradeoff between computational scalability and accuracy by dynamically tracking the traffic's natural cutoff sampling rate. We demonstrate on a real world 100 Gbps network that the BubbleCache algorithm helps reduce the computational cost by a factor of 1000 and the memory requirements by a factor of 100 while detecting the top flows on the network with very high probability.

cs.NI↗

arXiv2017

Algorithms and Data Structures to Accelerate Network Analysis

As the sheer amount of computer generated data continues to grow exponentially, new bottlenecks are unveiled that require rethinking our traditional software and hardware architectures. In this paper we present five algorithms and data structures (long queue emulation, lockless bimodal queues, tail early dropping, LFN tables, and multiresolution priority queues) designed to optimize the process of analyzing network traffic. We integrated these optimizations on R-Scope, a high performance network appliance that runs the Bro network analyzer, and present benchmarks showcasing performance speed ups of 5X at traffic rates of 10 Gbps.

cs.NI↗

arXiv2017

Multiresolution Priority Queues

Priority queues are container data structures essential to many high performance computing (HPC) applications. In this paper, we introduce multiresolution priority queues, a data structure that improves the performance of the standard heap based implementations by trading off a controllable amount of resolution in the space of priorities. The new data structure can reduce the worst case performance of inserting an element from O(log(n)) to O(log(r)), where n is the number of elements in the queue and r is the number of resolution groups in the priority space. The worst case cost of removing the top element is O(1). When the number of elements in the table is high, the amortized cost to insert an element becomes O(1).

cs.DS↗