Weaknesses: W1 The submission claims that existing approaches only capture spatial appearance (line 42), but the one that is compared with [2] is actually based on RNNs, that have the potential to capture motion information across a sequence of frames. W2 While the work acknowledges the challenges of of motion blurs and fine-grained gesture details (line 40), it does not address them in the proposed approach. W3 The quantitative gains in terms of BLEU (9.58 to 13.41) and ROUGE (31.80 to 34.96) scores are not outstanding. W4 The results of [2] by exploiting the glosses available in the dataset are better than the ones in this submission. Given that the contributions of the work address the visual representation, it is not argues why the proposed techniques are also assess with the Sign-to-Gloss-to-Text set up considered in [2].

The reviewers were positive about the ideas in the paper and mostly debated the merits of the evaluation. For one they were not fully convinced about the arguments in the rebuttal about the differences between the sharpness of boundaries for action localization and sign language translation. For camera ready I would suggest better addressing this point, as well as comparing or justifying differences to "Sign Language Transformers: Joint End-to-end Sign Language Recognition and Translation", Camgoz et al, CVPR 2020. One final suggestion is to add results with one more video encoder in addition to I3D.

Weaknesses: One of the strengths of generative algorithms for molecules is to capture a hard-to-describe statistical distribution of plausible molecules that can be made, paid for, stored in a vial, etc. There are many graphs that are formally valid according to valence rules (and their rdkit implementation) that could not exist as molecules because they are not stable. The premise of using generative models for molecular design is sampling natural-looking molecules. Just like generative models for faces, one just needs to look at this to judge whether the model has learned a richer chemistry than the hard-coded rules of RDKit. Very few molecules are shown from what the model produces.

The paper concerns a new online learning problem subject to the constraint of individual fairness. It provides a framework that reduces online classification in the considered model to standard online classification, obtaining an algorithm with sublinear regret both in terms of accuracy and fairness, as well as strong generalization bounds in the i.i.d. All the reviewers liked the paper and the proposed metric-free approach. The appreciated an interesting problem formulation and a clean reduction technique to a known online learning problem. The paper received very high uniform scores of 8 from each reviewer. The reviewers found some issues with the presentation, and I hope the authors will address them in the final version of the manuscript.

The submission studies the adversarial online learning in episodic loop-free Markov decision processes. The importance of this work is that it is the first to provide the understanding to an adversarial online learning problem where the transition function is unknown, the loss functions are changing, and each feedback is bandit. The related work clearly describe the line of this research field from fixing an unknown transition and an unknown loss function to the setting studied in this submission. Although the MDPs considered in the submission is L-layered and loop-free, the results and the analysis pave the way for general MDPs. The main idea is the design of the confidence sets to include the optimal occupancy measure which induces the optimal policy.

This work considers the relationship between convex surrogate loss and learning problem such as classification and ranking. The authors prove that this approach is equivalent, in a strong sense, to working with polyhedral (piecewise linear convex) losses, and give a construction of a link function through which L is a consistent surrogate for the loss it embeds. Some examples are presented to verify the theoretical analysis. This is an interesting direction in learning theory, while I have some concerns as follows: 1) What's the motivation of polyhedral losses? The authors should present some real applications and shows its importance, especially for some new learning problems and settings.

We propose a new finetuning method to provide pre-trained large language models (LMs) the ability to scale test-time compute through the diffusion framework. By increasing the number of diffusion steps, we show our finetuned models achieve monotonically increasing accuracy, directly translating to improved performance across downstream tasks. Furthermore, our finetuned models can expertly answer questions on specific topics by integrating powerful guidance techniques, and autonomously determine the compute required for a given problem by leveraging adaptive ODE solvers. Our method is universally applicable to any foundation model pre-trained with a cross-entropy loss and does not modify any of its original weights, fully preserving its strong single-step generation capabilities. We show our method is more effective and fully compatible with traditional finetuning approaches, introducing an orthogonal new direction to unify the strengths of the autoregressive and diffusion frameworks.

Active Learning (AL) is a sequential learning approach aiming at selecting the most informative data for model training. In many systems, safety constraints appear during data evaluation, requiring the development of safe AL methods. Key challenges of AL are the repeated model training and acquisition optimization required for data selection, which become particularly restrictive under safety constraints. This repeated effort often creates a bottleneck, especially in physical systems requiring real-time decision-making. In this paper, we propose a novel amortized safe AL framework. By leveraging a pretrained neural network policy, our method eliminates the need for repeated model training and acquisition optimization, achieving substantial speed improvements while maintaining competitive learning outcomes and safety awareness. The policy is trained entirely on synthetic data utilizing a novel safe AL objective. The resulting policy is highly versatile and adapts to a wide range of systems, as we demonstrate in our experiments. Furthermore, our framework is modular and we empirically show that we also achieve superior performance for unconstrained time-sensitive AL tasks if we omit the safety requirement.

This study investigates the potential of hybrid metaheuristic algorithms to enhance the training of Probabilistic Neural Networks (PNNs) by leveraging the complementary strengths of multiple optimisation strategies. Traditional learning methods, such as gradient-based approaches, often struggle to optimise high-dimensional and uncertain environments, while single-method metaheuristics may fail to exploit the solution space fully. To address these challenges, we propose the constrained Hybrid Metaheuristic (cHM) algorithm, a novel approach that combines multiple population-based optimisation techniques into a unified framework. The proposed procedure operates in two phases: an initial probing phase evaluates multiple metaheuristics to identify the best-performing one based on the error rate, followed by a fitting phase where the selected metaheuristic refines the PNN to achieve optimal smoothing parameters. This iterative process ensures efficient exploration and convergence, enhancing the network's generalisation and classification accuracy. cHM integrates several popular metaheuristics, such as BAT, Simulated Annealing, Flower Pollination Algorithm, Bacterial Foraging Optimization, and Particle Swarm Optimisation as internal optimisers. To evaluate cHM performance, experiments were conducted on 16 datasets with varying characteristics, including binary and multiclass classification tasks, balanced and imbalanced class distributions, and diverse feature dimensions. The results demonstrate that cHM effectively combines the strengths of individual metaheuristics, leading to faster convergence and more robust learning. By optimising the smoothing parameters of PNNs, the proposed method enhances classification performance across diverse datasets, proving its application flexibility and efficiency.

Language models are prone to occasionally undesirable generations, such as harmful or toxic content, despite their impressive capability to produce texts that appear accurate and coherent. This paper presents a new two-stage approach to detect and mitigate undesirable content generations by rectifying activations. First, we train an ensemble of layerwise classifiers to detect undesirable content using activations by minimizing a smooth surrogate of the risk-aware score. Then, for contents that are detected as undesirable, we propose layerwise distributional intervention policies that perturb the attention heads minimally while guaranteeing probabilistically the effectiveness of the intervention. Benchmarks on several language models and datasets show that our method outperforms baselines in reducing the generation of undesirable output.