Think Smarter not Harder: Adaptive Reasoning with Inference Aware Optimization

Solving mathematics problems has been an intriguing capability of large language models, and many efforts have been made to improve reasoning by extending reasoning length, such as through self-correction and extensive long chain-of-thoughts. While promising in problem-solving, advanced long reasoning chain models exhibit an undesired single-modal behavior, where trivial questions require unnecessarily tedious long chains of thought. In this work, we propose a way to allow models to be aware of inference budgets by formulating it as utility maximization with respect to an inference budget constraint, hence naming our algorithm Inference Budget-Constrained Policy Optimization (IBPO). In a nutshell, models fine-tuned through IBPO learn to ``understand'' the difficulty of queries and allocate inference budgets to harder ones. With different inference budgets, our best models are able to have a $4.14$\% and $5.74$\% absolute improvement ($8.08$\% and $11.2$\% relative improvement) on MATH500 using $2.16$x and $4.32$x inference budgets respectively, relative to LLaMA3.1 8B Instruct. These improvements are approximately $2$x those of self-consistency under the same budgets.

Step-KTO: Optimizing Mathematical Reasoning through Stepwise Binary Feedback

Large language models (LLMs) have recently demonstrated remarkable success in mathematical reasoning. Despite progress in methods like chain-of-thought prompting and self-consistency sampling, these advances often focus on final correctness without ensuring that the underlying reasoning process is coherent and reliable. This paper introduces Step-KTO, a training framework that combines process-level and outcome-level binary feedback to guide LLMs toward more trustworthy reasoning trajectories. By providing binary evaluations for both the intermediate reasoning steps and the final answer, Step-KTO encourages the model to adhere to logical progressions rather than relying on superficial shortcuts. Our experiments on challenging mathematical benchmarks show that Step-KTO significantly improves both final answer accuracy and the quality of intermediate reasoning steps. For example, on the MATH-500 dataset, Step-KTO achieves a notable improvement in Pass@1 accuracy over strong baselines. These results highlight the promise of integrating stepwise process feedback into LLM training, paving the way toward more interpretable and dependable reasoning capabilities.

Unifying Generative and Dense Retrieval for Sequential Recommendation

Sequential dense retrieval models utilize advanced sequence learning techniques to compute item and user representations, which are then used to rank relevant items for a user through inner product computation between the user and all item representations. However, this approach requires storing a unique representation for each item, resulting in significant memory requirements as the number of items grow. In contrast, the recently proposed generative retrieval paradigm offers a promising alternative by directly predicting item indices using a generative model trained on semantic IDs that encapsulate items' semantic information. Despite its potential for large-scale applications, a comprehensive comparison between generative retrieval and sequential dense retrieval under fair conditions is still lacking, leaving open questions regarding performance, and computation trade-offs. To address this, we compare these two approaches under controlled conditions on academic benchmarks and propose LIGER (LeveragIng dense retrieval for GEnerative Retrieval), a hybrid model that combines the strengths of these two widely used methods. LIGER integrates sequential dense retrieval into generative retrieval, mitigating performance differences and enhancing cold-start item recommendation in the datasets evaluated. This hybrid approach provides insights into the trade-offs between these approaches and demonstrates improvements in efficiency and effectiveness for recommendation systems in small-scale benchmarks.

Multi-IF: Benchmarking LLMs on Multi-Turn and Multilingual Instructions Following

Large Language Models (LLMs) have demonstrated impressive capabilities in various tasks, including instruction following, which is crucial for aligning model outputs with user expectations. However, evaluating LLMs' ability to follow instructions remains challenging due to the complexity and subjectivity of human language. Current benchmarks primarily focus on single-turn, monolingual instructions, which do not adequately reflect the complexities of real-world applications that require handling multi-turn and multilingual interactions. To address this gap, we introduce Multi-IF, a new benchmark designed to assess LLMs' proficiency in following multi-turn and multilingual instructions. Multi-IF, which utilizes a hybrid framework combining LLM and human annotators, expands upon the IFEval by incorporating multi-turn sequences and translating the English prompts into another 7 languages, resulting in a dataset of 4,501 multilingual conversations, where each has three turns. Our evaluation of 14 state-of-the-art LLMs on Multi-IF reveals that it presents a significantly more challenging task than existing benchmarks. All the models tested showed a higher rate of failure in executing instructions correctly with each additional turn. For example, o1-preview drops from 0.877 at the first turn to 0.707 at the third turn in terms of average accuracy over all languages. Moreover, languages with non-Latin scripts (Hindi, Russian, and Chinese) generally exhibit higher error rates, suggesting potential limitations in the models' multilingual capabilities. We release Multi-IF prompts and the evaluation code base to encourage further research in this critical area.

The Perfect Blend: Redefining RLHF with Mixture of Judges

Reinforcement learning from human feedback (RLHF) has become the leading approach for fine-tuning large language models (LLM). However, RLHF has limitations in multi-task learning (MTL) due to challenges of reward hacking and extreme multi-objective optimization (i.e., trade-off of multiple and/or sometimes conflicting objectives). Applying RLHF for MTL currently requires careful tuning of the weights for reward model and data combinations. This is often done via human intuition and does not generalize. In this work, we introduce a novel post-training paradigm which we called Constrained Generative Policy Optimization (CGPO). The core of CGPO is Mixture of Judges (MoJ) with cost-efficient constrained policy optimization with stratification, which can identify the perfect blend in RLHF in a principled manner. It shows strong empirical results with theoretical guarantees, does not require extensive hyper-parameter tuning, and is plug-and-play in common post-training pipelines. Together, this can detect and mitigate reward hacking behaviors while reaching a pareto-optimal point across an extremely large number of objectives. Our empirical evaluations demonstrate that CGPO significantly outperforms standard RLHF algorithms like PPO and DPO across various tasks including general chat, STEM questions, instruction following, and coding. Specifically, CGPO shows improvements of 7.4% in AlpacaEval-2 (general chat), 12.5% in Arena-Hard (STEM & reasoning), and consistent gains in other domains like math and coding. Notably, PPO, while commonly used, is prone to severe reward hacking in popular coding benchmarks, which CGPO successfully addresses. This breakthrough in RLHF not only tackles reward hacking and extreme multi-objective optimization challenges but also advances the state-of-the-art in aligning general-purpose LLMs for diverse applications.

PromptAttack: Probing Dialogue State Trackers with Adversarial Prompts

A key component of modern conversational systems is the Dialogue State Tracker (or DST), which models a user's goals and needs. Toward building more robust and reliable DSTs, we introduce a prompt-based learning approach to automatically generate effective adversarial examples to probe DST models. Two key characteristics of this approach are: (i) it only needs the output of the DST with no need for model parameters, and (ii) it can learn to generate natural language utterances that can target any DST. Through experiments over state-of-the-art DSTs, the proposed framework leads to the greatest reduction in accuracy and the best attack success rate while maintaining good fluency and a low perturbation ratio. We also show how much the generated adversarial examples can bolster a DST through adversarial training. These results indicate the strength of prompt-based attacks on DSTs and leave open avenues for continued refinement.

Grad-PU: Arbitrary-Scale Point Cloud Upsampling via Gradient Descent with Learned Distance Functions

Most existing point cloud upsampling methods have roughly three steps: feature extraction, feature expansion and 3D coordinate prediction. However,they usually suffer from two critical issues: (1)fixed upsampling rate after one-time training, since the feature expansion unit is customized for each upsampling rate; (2)outliers or shrinkage artifact caused by the difficulty of precisely predicting 3D coordinates or residuals of upsampled points. To adress them, we propose a new framework for accurate point cloud upsampling that supports arbitrary upsampling rates. Our method first interpolates the low-res point cloud according to a given upsampling rate. And then refine the positions of the interpolated points with an iterative optimization process, guided by a trained model estimating the difference between the current point cloud and the high-res target. Extensive quantitative and qualitative results on benchmarks and downstream tasks demonstrate that our method achieves the state-of-the-art accuracy and efficiency.

Evolution of Popularity Bias: Empirical Study and Debiasing

Popularity bias is a long-standing challenge in recommender systems. Such a bias exerts detrimental impact on both users and item providers, and many efforts have been dedicated to studying and solving such a bias. However, most existing works situate this problem in a static setting, where the bias is analyzed only for a single round of recommendation with logged data. These works fail to take account of the dynamic nature of real-world recommendation process, leaving several important research questions unanswered: how does the popularity bias evolve in a dynamic scenario? what are the impacts of unique factors in a dynamic recommendation process on the bias? and how to debias in this long-term dynamic process? In this work, we aim to tackle these research gaps. Concretely, we conduct an empirical study by simulation experiments to analyze popularity bias in the dynamic scenario and propose a dynamic debiasing strategy and a novel False Positive Correction method utilizing false positive signals to debias, which show effective performance in extensive experiments.

Density-preserving Deep Point Cloud Compression

Local density of point clouds is crucial for representing local details, but has been overlooked by existing point cloud compression methods. To address this, we propose a novel deep point cloud compression method that preserves local density information. Our method works in an auto-encoder fashion: the encoder downsamples the points and learns point-wise features, while the decoder upsamples the points using these features. Specifically, we propose to encode local geometry and density with three embeddings: density embedding, local position embedding and ancestor embedding. During the decoding, we explicitly predict the upsampling factor for each point, and the directions and scales of the upsampled points. To mitigate the clustered points issue in existing methods, we design a novel sub-point convolution layer, and an upsampling block with adaptive scale. Furthermore, our method can also compress point-wise attributes, such as normal. Extensive qualitative and quantitative results on SemanticKITTI and ShapeNet demonstrate that our method achieves the state-of-the-art rate-distortion trade-off.

MetaBalance: Improving Multi-Task Recommendations via Adapting Gradient Magnitudes of Auxiliary Tasks

In many personalized recommendation scenarios, the generalization ability of a target task can be improved via learning with additional auxiliary tasks alongside this target task on a multi-task network. However, this method often suffers from a serious optimization imbalance problem. On the one hand, one or more auxiliary tasks might have a larger influence than the target task and even dominate the network weights, resulting in worse recommendation accuracy for the target task. On the other hand, the influence of one or more auxiliary tasks might be too weak to assist the target task. More challenging is that this imbalance dynamically changes throughout the training process and varies across the parts of the same network. We propose a new method: MetaBalance to balance auxiliary losses via directly manipulating their gradients w.r.t the shared parameters in the multi-task network. Specifically, in each training iteration and adaptively for each part of the network, the gradient of an auxiliary loss is carefully reduced or enlarged to have a closer magnitude to the gradient of the target loss, preventing auxiliary tasks from being so strong that dominate the target task or too weak to help the target task. Moreover, the proximity between the gradient magnitudes can be flexibly adjusted to adapt MetaBalance to different scenarios. The experiments show that our proposed method achieves a significant improvement of 8.34% in terms of NDCG@10 upon the strongest baseline on two real-world datasets. The code of our approach can be found at here: https://github.com/facebookresearch/MetaBalance