Stable Preference Optimization for LLMs: A Bilevel Approach Beyond
  Direct Preference Optimization

By: Chengtao Jian, Kai Yang, Ye Ouyang, Xiaozhou Ye

Direct Preference Optimization (DPO) has emerged as a popular and efficient alternative to reward modeling and reinforcement learning for aligning language models with human preferences. Despite its empirical success, the theoretical properties and intrinsic limitations of DPO remain underexplored. In this work, we first present a comprehensive analysis of DPO's dynamics from a probability evolution perspective. Our analysis reveals that DPO ... more
Direct Preference Optimization (DPO) has emerged as a popular and efficient alternative to reward modeling and reinforcement learning for aligning language models with human preferences. Despite its empirical success, the theoretical properties and intrinsic limitations of DPO remain underexplored. In this work, we first present a comprehensive analysis of DPO's dynamics from a probability evolution perspective. Our analysis reveals that DPO is highly sensitive to initialization. It also tends to misallocate probability mass, which can inadvertently shift probability toward irrelevant or undesired responses. This misallocation may unintentionally reinforce model bias, thereby compromising both the stability of model alignment and the consistency with intended preferences. Motivated by these theoretical findings, we propose a theoretically grounded bilevel optimization framework that tightly integrate supervised fine-tuning with an enhanced DPO objective a.k.a. stable preference optimization. Our approach introduces a principled regularization scheme to explicitly encourage absolute probability improvement for preferred outputs, while maintaining stable optimization dynamics. Experiments on challenging reasoning and summarization benchmarks elucidate that our method consistently improves reasoning accuracy and better aligns output distributions with intended preferences, outperforming standard DPO. Stable preference optimization provides new insights into the design of preference-based alignment objectives and opens up new avenues towards more reliable and interpretable language model alignment. less
Dynamic Chunking for End-to-End Hierarchical Sequence Modeling

By: Sukjun Hwang, Brandon Wang, Albert Gu

Despite incredible progress in language models (LMs) in recent years, largely resulting from moving away from specialized models designed for specific tasks to general models based on powerful architectures (e.g. the Transformer) that learn everything from raw data, pre-processing steps such as tokenization remain a barrier to true end-to-end foundation models. We introduce a collection of new techniques that enable a dynamic chunking mechani... more
Despite incredible progress in language models (LMs) in recent years, largely resulting from moving away from specialized models designed for specific tasks to general models based on powerful architectures (e.g. the Transformer) that learn everything from raw data, pre-processing steps such as tokenization remain a barrier to true end-to-end foundation models. We introduce a collection of new techniques that enable a dynamic chunking mechanism which automatically learns content -- and context -- dependent segmentation strategies learned jointly with the rest of the model. Incorporating this into an explicit hierarchical network (H-Net) allows replacing the (implicitly hierarchical) tokenization-LM-detokenization pipeline with a single model learned fully end-to-end. When compute- and data- matched, an H-Net with one stage of hierarchy operating at the byte level outperforms a strong Transformer language model operating over BPE tokens. Iterating the hierarchy to multiple stages further increases its performance by modeling multiple levels of abstraction, demonstrating significantly better scaling with data and matching a token-based Transformer of twice its size. H-Nets pretrained on English show significantly increased character-level robustness, and qualitatively learn meaningful data-dependent chunking strategies without any heuristics or explicit supervision. Finally, the H-Net's improvement over tokenized pipelines is further increased in languages and modalities with weaker tokenization heuristics, such as Chinese and code, or DNA sequences (nearly 4x improvement in data efficiency over baselines), showing the potential of true end-to-end models that learn and scale better from unprocessed data. less
EXPO: Stable Reinforcement Learning with Expressive Policies

By: Perry Dong, Qiyang Li, Dorsa Sadigh, Chelsea Finn

We study the problem of training and fine-tuning expressive policies with online reinforcement learning (RL) given an offline dataset. Training expressive policy classes with online RL present a unique challenge of stable value maximization. Unlike simpler Gaussian policies commonly used in online RL, expressive policies like diffusion and flow-matching policies are parameterized by a long denoising chain, which hinders stable gradient propag... more
We study the problem of training and fine-tuning expressive policies with online reinforcement learning (RL) given an offline dataset. Training expressive policy classes with online RL present a unique challenge of stable value maximization. Unlike simpler Gaussian policies commonly used in online RL, expressive policies like diffusion and flow-matching policies are parameterized by a long denoising chain, which hinders stable gradient propagation from actions to policy parameters when optimizing against some value function. Our key insight is that we can address stable value maximization by avoiding direct optimization over value with the expressive policy and instead construct an on-the-fly RL policy to maximize Q-value. We propose Expressive Policy Optimization (EXPO), a sample-efficient online RL algorithm that utilizes an on-the-fly policy to maximize value with two parameterized policies -- a larger expressive base policy trained with a stable imitation learning objective and a light-weight Gaussian edit policy that edits the actions sampled from the base policy toward a higher value distribution. The on-the-fly policy optimizes the actions from the base policy with the learned edit policy and chooses the value maximizing action from the base and edited actions for both sampling and temporal-difference (TD) backup. Our approach yields up to 2-3x improvement in sample efficiency on average over prior methods both in the setting of fine-tuning a pretrained policy given offline data and in leveraging offline data to train online. less
Skip a Layer or Loop it? Test-Time Depth Adaptation of Pretrained LLMs

By: Ziyue Li, Yang Li, Tianyi Zhou

Can a pretrained neural network adapt its architecture to different inputs without any finetuning? Do we need all layers for simple tasks, and are they adequate for challenging tasks? We found that the layers of a pretrained large language model (LLM) can be manipulated as separate modules to build a better and even shallower model customized for each test sample. In particular, each layer from the pretrained model can be skipped/pruned or re... more
Can a pretrained neural network adapt its architecture to different inputs without any finetuning? Do we need all layers for simple tasks, and are they adequate for challenging tasks? We found that the layers of a pretrained large language model (LLM) can be manipulated as separate modules to build a better and even shallower model customized for each test sample. In particular, each layer from the pretrained model can be skipped/pruned or repeated multiple times as recurrent neural networks (RNN), and stacked with others in arbitrary orders, yielding a chain-of-layers (CoLa) per sample. This compositional space greatly expands the scope of existing works on looped/recurrent pretrained modules, layer pruning, or early-exit networks. We develop a Monte Carlo Tree Search (MCTS) protocol to explore and identify the optimal CoLa for each sample from math and commonsense reasoning benchmarks. Compared to a static model of a fixed depth, CoLa allows shortcut paths (fast thinking), recurrence of the same layer(s) (slow thinking), and combining both, offering more flexible, dynamic architectures for different inputs. We conduct an extensive analysis of the MCTS-optimized CoLa, which leads to two key findings: (1) For >75% of samples with correct predictions by the original LLM, we can find shorter CoLa, suggesting a large space for improving inference efficiency; (2) For >60% of samples with originally incorrect predictions, we can identify CoLa achieving correct predictions, suggesting a large space of performance enhancement. Our results highlight the shortcomings of using a fixed architecture of pre-trained LLMs for inference on different samples and pave the way to unlock the generalization power of test-time depth adaptation. less
Measuring AI Alignment with Human Flourishing

By: Elizabeth Hilliard, Akshaya Jagadeesh, Alex Cook, Steele Billings, Nicholas Skytland, Alicia Llewellyn, Jackson Paull, Nathan Paull, Nolan Kurylo, Keatra Nesbitt, Robert Gruenewald, Anthony Jantzi, Omar Chavez

This paper introduces the Flourishing AI Benchmark (FAI Benchmark), a novel evaluation framework that assesses AI alignment with human flourishing across seven dimensions: Character and Virtue, Close Social Relationships, Happiness and Life Satisfaction, Meaning and Purpose, Mental and Physical Health, Financial and Material Stability, and Faith and Spirituality. Unlike traditional benchmarks that focus on technical capabilities or harm preve... more
This paper introduces the Flourishing AI Benchmark (FAI Benchmark), a novel evaluation framework that assesses AI alignment with human flourishing across seven dimensions: Character and Virtue, Close Social Relationships, Happiness and Life Satisfaction, Meaning and Purpose, Mental and Physical Health, Financial and Material Stability, and Faith and Spirituality. Unlike traditional benchmarks that focus on technical capabilities or harm prevention, the FAI Benchmark measures AI performance on how effectively models contribute to the flourishing of a person across these dimensions. The benchmark evaluates how effectively LLM AI systems align with current research models of holistic human well-being through a comprehensive methodology that incorporates 1,229 objective and subjective questions. Using specialized judge Large Language Models (LLMs) and cross-dimensional evaluation, the FAI Benchmark employs geometric mean scoring to ensure balanced performance across all flourishing dimensions. Initial testing of 28 leading language models reveals that while some models approach holistic alignment (with the highest-scoring models achieving 72/100), none are acceptably aligned across all dimensions, particularly in Faith and Spirituality, Character and Virtue, and Meaning and Purpose. This research establishes a framework for developing AI systems that actively support human flourishing rather than merely avoiding harm, offering significant implications for AI development, ethics, and evaluation. less
AI Should Sense Better, Not Just Scale Bigger: Adaptive Sensing as a
  Paradigm Shift

By: Eunsu Baek, Keondo Park, Jeonggil Ko, Min-hwan Oh, Taesik Gong, Hyung-Sin Kim

Current AI advances largely rely on scaling neural models and expanding training datasets to achieve generalization and robustness. Despite notable successes, this paradigm incurs significant environmental, economic, and ethical costs, limiting sustainability and equitable access. Inspired by biological sensory systems, where adaptation occurs dynamically at the input (e.g., adjusting pupil size, refocusing vision)--we advocate for adaptive s... more
Current AI advances largely rely on scaling neural models and expanding training datasets to achieve generalization and robustness. Despite notable successes, this paradigm incurs significant environmental, economic, and ethical costs, limiting sustainability and equitable access. Inspired by biological sensory systems, where adaptation occurs dynamically at the input (e.g., adjusting pupil size, refocusing vision)--we advocate for adaptive sensing as a necessary and foundational shift. Adaptive sensing proactively modulates sensor parameters (e.g., exposure, sensitivity, multimodal configurations) at the input level, significantly mitigating covariate shifts and improving efficiency. Empirical evidence from recent studies demonstrates that adaptive sensing enables small models (e.g., EfficientNet-B0) to surpass substantially larger models (e.g., OpenCLIP-H) trained with significantly more data and compute. We (i) outline a roadmap for broadly integrating adaptive sensing into real-world applications spanning humanoid, healthcare, autonomous systems, agriculture, and environmental monitoring, (ii) critically assess technical and ethical integration challenges, and (iii) propose targeted research directions, such as standardized benchmarks, real-time adaptive algorithms, multimodal integration, and privacy-preserving methods. Collectively, these efforts aim to transition the AI community toward sustainable, robust, and equitable artificial intelligence systems. less
Meek Models Shall Inherit the Earth

By: Hans Gundlach, Jayson Lynch, Neil Thompson

The past decade has seen incredible scaling of AI systems by a few companies, leading to inequality in AI model performance. This paper argues that, contrary to prevailing intuition, the diminishing returns to compute scaling will lead to a convergence of AI model capabilities. In other words, meek models (those with limited computation budget) shall inherit the earth, approaching the performance level of the best models overall. We develop a... more
The past decade has seen incredible scaling of AI systems by a few companies, leading to inequality in AI model performance. This paper argues that, contrary to prevailing intuition, the diminishing returns to compute scaling will lead to a convergence of AI model capabilities. In other words, meek models (those with limited computation budget) shall inherit the earth, approaching the performance level of the best models overall. We develop a model illustrating that under a fixed-distribution next-token objective, the marginal capability returns to raw compute shrink substantially. Given current scaling practices, we argue that these diminishing returns are strong enough that even companies that can scale their models exponentially faster than other organizations will eventually have little advantage in capabilities. As part of our argument, we give several reasons that proxies like training loss differences capture important capability measures using evidence from benchmark data and theoretical performance models. In addition, we analyze empirical data on the capability difference of AI models over time. Finally, in light of the increasing ability of meek models, we argue that AI strategy and policy require reexamination, and we outline the areas this shift will affect. less
DiffSpectra: Molecular Structure Elucidation from Spectra using
  Diffusion Models

By: Liang Wang, Yu Rong, Tingyang Xu, Zhenyi Zhong, Zhiyuan Liu, Pengju Wang, Deli Zhao, Qiang Liu, Shu Wu, Liang Wang

Molecular structure elucidation from spectra is a foundational problem in chemistry, with profound implications for compound identification, synthesis, and drug development. Traditional methods rely heavily on expert interpretation and lack scalability. Pioneering machine learning methods have introduced retrieval-based strategies, but their reliance on finite libraries limits generalization to novel molecules. Generative models offer a promi... more
Molecular structure elucidation from spectra is a foundational problem in chemistry, with profound implications for compound identification, synthesis, and drug development. Traditional methods rely heavily on expert interpretation and lack scalability. Pioneering machine learning methods have introduced retrieval-based strategies, but their reliance on finite libraries limits generalization to novel molecules. Generative models offer a promising alternative, yet most adopt autoregressive SMILES-based architectures that overlook 3D geometry and struggle to integrate diverse spectral modalities. In this work, we present DiffSpectra, a generative framework that directly infers both 2D and 3D molecular structures from multi-modal spectral data using diffusion models. DiffSpectra formulates structure elucidation as a conditional generation process. Its denoising network is parameterized by Diffusion Molecule Transformer, an SE(3)-equivariant architecture that integrates topological and geometric information. Conditioning is provided by SpecFormer, a transformer-based spectral encoder that captures intra- and inter-spectral dependencies from multi-modal spectra. Extensive experiments demonstrate that DiffSpectra achieves high accuracy in structure elucidation, recovering exact structures with 16.01% top-1 accuracy and 96.86% top-20 accuracy through sampling. The model benefits significantly from 3D geometric modeling, SpecFormer pre-training, and multi-modal conditioning. These results highlight the effectiveness of spectrum-conditioned diffusion modeling in addressing the challenge of molecular structure elucidation. To our knowledge, DiffSpectra is the first framework to unify multi-modal spectral reasoning and joint 2D/3D generative modeling for de novo molecular structure elucidation. less
Squeeze the Soaked Sponge: Efficient Off-policy Reinforcement Finetuning
  for Large Language Model

By: Jing Liang, Hongyao Tang, Yi Ma, Jinyi Liu, Yan Zheng, Shuyue Hu, Lei Bai, Jianye Hao

Reinforcement Learning (RL) has demonstrated its potential to improve the reasoning ability of Large Language Models (LLMs). One major limitation of most existing Reinforcement Finetuning (RFT) methods is that they are on-policy RL in nature, i.e., data generated during the past learning process is not fully utilized. This inevitably comes at a significant cost of compute and time, posing a stringent bottleneck on continuing economic and effi... more
Reinforcement Learning (RL) has demonstrated its potential to improve the reasoning ability of Large Language Models (LLMs). One major limitation of most existing Reinforcement Finetuning (RFT) methods is that they are on-policy RL in nature, i.e., data generated during the past learning process is not fully utilized. This inevitably comes at a significant cost of compute and time, posing a stringent bottleneck on continuing economic and efficient scaling. To this end, we launch the renaissance of off-policy RL and propose Reincarnating Mix-policy Proximal Policy Gradient (ReMix), a general approach to enable on-policy RFT methods like PPO and GRPO to leverage off-policy data. ReMix consists of three major components: (1) Mix-policy proximal policy gradient with an increased Update-To-Data (UTD) ratio for efficient training; (2) KL-Convex policy constraint to balance the trade-off between stability and flexibility; (3) Policy reincarnation to achieve a seamless transition from efficient early-stage learning to steady asymptotic improvement. In our experiments, we train a series of ReMix models upon PPO, GRPO and 1.5B, 7B base models. ReMix shows an average Pass@1 accuracy of 52.10% (for 1.5B model) with 0.079M response rollouts, 350 training steps and achieves 63.27%/64.39% (for 7B model) with 0.007M/0.011M response rollouts, 50/75 training steps, on five math reasoning benchmarks (i.e., AIME'24, AMC'23, Minerva, OlympiadBench, and MATH500). Compared with 15 recent advanced models, ReMix shows SOTA-level performance with an over 30x to 450x reduction in training cost in terms of rollout data volume. In addition, we reveal insightful findings via multifaceted analysis, including the implicit preference for shorter responses due to the Whipping Effect of off-policy discrepancy, the collapse mode of self-reflection behavior under the presence of severe off-policyness, etc. less
Self-Supervised Learning at the Edge: The Cost of Labeling

By: Roberto Pereira, Fernanda Famá, Asal Rangrazi, Marco Miozzo, Charalampos Kalalas, Paolo Dini

Contrastive learning (CL) has recently emerged as an alternative to traditional supervised machine learning solutions by enabling rich representations from unstructured and unlabeled data. However, CL and, more broadly, self-supervised learning (SSL) methods often demand a large amount of data and computational resources, posing challenges for deployment on resource-constrained edge devices. In this work, we explore the feasibility and effici... more
Contrastive learning (CL) has recently emerged as an alternative to traditional supervised machine learning solutions by enabling rich representations from unstructured and unlabeled data. However, CL and, more broadly, self-supervised learning (SSL) methods often demand a large amount of data and computational resources, posing challenges for deployment on resource-constrained edge devices. In this work, we explore the feasibility and efficiency of SSL techniques for edge-based learning, focusing on trade-offs between model performance and energy efficiency. In particular, we analyze how different SSL techniques adapt to limited computational, data, and energy budgets, evaluating their effectiveness in learning robust representations under resource-constrained settings. Moreover, we also consider the energy costs involved in labeling data and assess how semi-supervised learning may assist in reducing the overall energy consumed to train CL models. Through extensive experiments, we demonstrate that tailored SSL strategies can achieve competitive performance while reducing resource consumption by up to 4X, underscoring their potential for energy-efficient learning at the edge. less