The paper presents two approaches submitted to the WMT 2025 Automated Translation Quality Evaluation Systems Task 3 - Quality Estimation (QE)-informed Segment-level Error Correction. While jointly training QE systems with Automatic Post-Editing (APE) has shown improved performance for both tasks, APE systems are still known to overcorrect the output of Machine Translation (MT), leading to a degradation in performance. We investigate a simple training-free approach - QE-informed Retranslation, and compare it with another within the same training-free paradigm. Our winning approach selects the highest-quality translation from multiple candidates generated by different LLMs. The second approach, more akin to APE, instructs an LLM to replace error substrings as specified in the provided QE explanation(s). A conditional heuristic was employed to minimise the number of edits, with the aim of maximising the Gain-to-Edit ratio. The two proposed approaches achieved a Delta COMET score of 0.0201 and -0.0108, respectively, leading the first approach to achieve the winning position on the subtask leaderboard.

Reinforcement learning(RL) post-training has become essential for aligning large language models (LLMs), yet its efficiency is increasingly constrained by the rollout phase, where long trajectories are generated token by token. We identify a major bottleneck:the long-tail distribution of rollout lengths, where a small fraction of long generations dominates wall clock time and a complementary opportunity; the availability of historical rollouts that reveal stable prompt level patterns across training epochs. Motivated by these observations, we propose DAS, a Distribution Aware Speculative decoding framework that accelerates RL rollouts without altering model outputs. DAS integrates two key ideas: an adaptive, nonparametric drafter built from recent rollouts using an incrementally maintained suffix tree, and a length aware speculation policy that allocates more aggressive draft budgets to long trajectories that dominate makespan. This design exploits rollout history to sustain acceptance while balancing base and token level costs during decoding. Experiments on math and code reasoning tasks show that DAS reduces rollout time up to 50% while preserving identical training curves, demonstrating that distribution-aware speculative decoding can significantly accelerate RL post training without compromising learning quality.

Application of phylogenetic methods to textual traditions has traditionally treated all changes as equivalent even though it is widely recognized that certain types of variants were more likely to be introduced than others. While it is possible to give weights to certain changes using a maximum parsimony evaluation criterion, it is difficult to state a priori what these weights should be. Probabilistic methods, such as Bayesian phylogenetics, allow users to create categories of changes, and the transition rates for each category can be estimated as part of the analysis. This classification of types of changes in readings also allows for inspecting the probability of these categories across each branch in the resulting trees. However, classification of readings is time-consuming, as it requires categorizing each reading against every other reading at each variation unit, presenting a significant barrier to entry for this kind of analysis. This paper presents Rdgai, a software package that automates this classification task using multi-lingual large language models (LLMs). The tool allows users to easily manually classify changes in readings and then it uses these annotations in the prompt for an LLM to automatically classify the remaining reading transitions. These classifications are stored in TEI XML and ready for downstream phylogenetic analysis. This paper demonstrates the application with data an Arabic translation of the Gospels.

Backdoor attacks pose a serious threat to the security of large language models (LLMs), causing them to exhibit anomalous behavior under specific trigger conditions. The design of backdoor triggers has evolved from fixed triggers to dynamic or implicit triggers. This increased flexibility in trigger design makes it challenging for defenders to identify their specific forms accurately. Most existing backdoor defense methods are limited to specific types of triggers or rely on an additional clean model for support. To address this issue, we propose a backdoor detection method based on attention similarity, enabling backdoor detection without prior knowledge of the trigger. Our study reveals that models subjected to backdoor attacks exhibit unusually high similarity among attention heads when exposed to triggers. Based on this observation, we propose an attention safety alignment approach combined with head-wise fine-tuning to rectify potentially contaminated attention heads, thereby effectively mitigating the impact of backdoor attacks. Extensive experimental results demonstrate that our method significantly reduces the success rate of backdoor attacks while preserving the model's performance on downstream tasks.

Large language models (LLMs) and vision language models (VLMs), such as DeepSeek R1,OpenAI o3, and Gemini 2.5 Pro, have demonstrated remarkable reasoning capabilities across logical inference, problem solving, and decision making. However, spatial reasoning:a fundamental component of human cognition that includes mental rotation, navigation, and spatial relationship comprehension remains a significant challenge for current advanced VLMs. We hypothesize that imagination, the internal simulation of spatial states, is the dominant reasoning mechanism within a spatial world model. To test this hypothesis and systematically probe current VLM spatial reasoning mechanisms, we introduce SpatiaLite, a fully synthetic benchmark that jointly measures spatial reasoning accuracy and reasoning efficiency. Comprehensive experiments reveal three key findings. First, advanced VLMs predominantly rely on linguistic representations for reasoning and imagination, resulting in significant deficiencies on visual centric tasks that demand perceptual spatial relations and 3D geometry transformations such as mental rotation or projection prediction. Second, advanced VLMs exhibit severe inefficiency in their current spatial reasoning mechanisms, with token usage growing rapidly as transformation complexity increases. Third, we propose an Imagery Driven Framework (IDF) for data synthesis and training, which can implicitly construct an internal world model that is critical for spatial reasoning in VLMs. Building on SpatiaLite, this work delineates the spatial reasoning limits and patterns of advanced VLMs, identifies key shortcomings, and informs future advances

Understanding the legally relevant factual basis of an event and conveying it through text is a key skill of legal professionals. This skill is important for preparing forms (e.g., insurance claims) or other legal documents (e.g., court claims), but often presents a challenge for laypeople. Current AI approaches aim to bridge this gap, but mostly rely on the user to articulate what has happened in text, which may be challenging for many. Here, we investigate the capability of large language models (LLMs) to understand and summarize events occurring in videos. We ask an LLM to summarize and draft legal letters, based on 120 YouTube videos showing legal issues in various domains. Overall, 71.7\% of the summaries were rated as of high or medium quality, which is a promising result, opening the door to a number of applications in e.g. access to justice.

Large Language Models (LLMs) are increasingly vulnerable to adversarial attacks that can subtly manipulate their outputs. While various defense mechanisms have been proposed, many operate as black boxes, lacking transparency in their decision-making. This paper introduces ExplainableGuard, an interpretable adversarial defense framework leveraging the chain-of-thought (CoT) reasoning capabilities of DeepSeek-Reasoner. Our approach not only detects and neutralizes adversarial perturbations in text but also provides step-by-step explanations for each defense action. We demonstrate how tailored CoT prompts guide the LLM to perform a multi-faceted analysis (character, word, structural, and semantic) and generate a purified output along with a human-readable justification. Preliminary results on the GLUE Benchmark and IMDB Movie Reviews dataset show promising defense efficacy. Additionally, a human evaluation study reveals that ExplainableGuard's explanations outperform ablated variants in clarity, specificity, and actionability, with a 72.5% deployability-trust rating, underscoring its potential for more trustworthy LLM deployments.

Offline imitation learning (offline IL) enables training effective policies without requiring explicit reward annotations. Recent approaches attempt to estimate rewards for unlabeled datasets using a small set of expert demonstrations. However, these methods often assume that the similarity between a trajectory and an expert demonstration is positively correlated with the reward, which oversimplifies the underlying reward structure. We propose PROF, a novel framework that leverages large language models (LLMs) to generate and improve executable reward function codes from natural language descriptions and a single expert trajectory. We propose Reward Preference Ranking (RPR), a novel reward function quality assessment and ranking strategy without requiring environment interactions or RL training. RPR calculates the dominance scores of the reward functions, where higher scores indicate better alignment with expert preferences. By alternating between RPR and text-based gradient optimization, PROF fully automates the selection and refinement of optimal reward functions for downstream policy learning. Empirical results on D4RL demonstrate that PROF surpasses or matches recent strong baselines across numerous datasets and domains, highlighting the effectiveness of our approach.

Although LLM training is typically centralized with high-bandwidth interconnects and large compute budgets, emerging methods target communication-constrained training in distributed environments. The model trade-offs introduced by this shift remain underexplored, and our goal is to study them. We use the open-source nanochat project, a compact 8K-line full-stack ChatGPT-like implementation containing tokenization, pretraining, fine-tuning, and serving, as a controlled baseline. We implement the DiLoCo algorithm as a lightweight wrapper over nanochat's training loop, performing multiple local steps per worker before synchronization with an outer optimizer, effectively reducing communication by orders of magnitude. This inner-outer training is compared against a standard data-parallel (DDP) setup. Because nanochat is small and inspectable, it enables controlled pipeline adaptations and allows direct comparison with the conventional centralized baseline. DiLoCo achieves stable convergence and competitive loss in pretraining but yields worse MMLU, GSM8K, and HumanEval scores after mid-training and SFT. We discover that using DiLoCo-pretrained weights and running mid- and post-training with DDP fails to recover performance, revealing irreversible representation drift from asynchronous updates that impairs downstream alignment. We provide this implementation as an official fork of nanochat on GitHub.

We propose RT (Refine Thought), a method that can enhance the semantic rea-soning ability of text embedding models. The method obtains the final semanticrepresentation by running multiple forward passes of the text embedding model.Experiments show that RT achieves significant improvements on semantic reason-ing tasks in BRIGHT and the person job matching benchmark PJBenchmark1, while maintaining consistent performance on general-purpose semantic under-standing tasks such as C-MTEB. Our results indicate that RT is effective becauseit further activates the semantic reasoning ability learned during pretraining bydecoder-only text embedding models(e.g., Qwen3-Embedding-8B). RT canbe seen as a test-time inference method.