ICLR 2025 Papers

Large Language Models fine-tuned with Reinforcement Learning from Human Feedback (RLHF-LLMs) can over-rely on aligned preferences without truly gaining self-knowledge, leading to hallucination and biases. If an LLM can better access its knowledge and know what it knows, it can avoid making false or unsupported claims. Therefore, it is crucial to evaluate whether LLMs have the ability to know what they know, as it can help to ensure accuracy and faithfulness in real-world applications. Inspired by research in Educational Psychology, surface learners who don’t really know are easily affected by teacher and peer guidance, we treat LLM as a student, incorporate role guidance in prompts to explore whether LLMs really know. Specifically, we propose a novel strategy called Role-Guided and Self-Reflection (RoSe) to fully assess whether LLM “knows it knows”. We introduce multiple combinations of different roles and strong reminder in prompts combined with self-reflection to explore what local information in prompt LLMs rely on and whether LLMs remain unaffected by external guidance with varying roles. Our findings reveal that LLMs are very sensitive to the strong reminder information. Role guidance can help LLMs reduce their reliance on strong reminder. Meanwhile, LLMs tend to trust the role of authority more when guided by different roles. Following these findings, we propose a double-calibrated strategy with verbalized confidence to extract well-calibrated data from closed-source LLM and fine-tune open-source LLMs. Extensive experiments conducted on fine-tuning open-source LLMs demonstrate the effectiveness of double-calibrated strategy in mitigating the reliance of LLMs on local information. For a thorough comparison, we not only employ public JEC-QA and openBookQA datasets, but also construct EG-QA which contains English Grammar multiple-choice question-answering and 14 key knowledge points for assessing self-knowledge and logical reasoning.

In the study of LLMs, sycophancy represents a prevalent hallucination that poses significant challenges to these models. Specifically, LLMs often fail to adhere to original correct responses, instead blindly agreeing with users' opinions, even when those opinions are incorrect or malicious. However, research on sycophancy in visual language models (VLMs) has been scarce. In this work, we extend the exploration of sycophancy from LLMs to VLMs, introducing the MM-SY benchmark to evaluate this phenomenon. We present evaluation results from multiple representative models, addressing the gap in sycophancy research for VLMs. To mitigate sycophancy, we propose a synthetic dataset for training and employ methods based on prompts, supervised fine-tuning, and DPO. Our experiments demonstrate that these methods effectively alleviate sycophancy in VLMs. Additionally, we probe VLMs to assess the semantic impact of sycophancy and analyze the attention distribution of visual tokens. Our findings indicate that the ability to prevent sycophancy is predominantly observed in higher layers of the model. The lack of attention to image knowledge in these higher layers may contribute to sycophancy, and enhancing image attention at high layers proves beneficial in mitigating this issue.

RNA is a dynamic biomolecule crucial for cellular regulation, with its function largely determined by its folding into complex structures, while misfolding can lead to multifaceted biological sequelae. During the folding process, RNA traverses through a series of intermediate structural states, with each transition occurring at variable rates that collectively influence the time required to reach the functional form. Understanding these folding kinetics is vital for predicting RNA behavior and optimizing applications in synthetic biology and drug discovery. While in silico kinetic RNA folding simulators are often computationally intensive and time-consuming, accurate approximations of the folding times can already be very informative to assess the efficiency of the folding process. In this work, we present KinPFN, a novel approach that leverages prior-data fitted networks to directly model the posterior predictive distribution of RNA folding times. By training on synthetic data representing arbitrary prior folding times, KinPFN efficiently approximates the cumulative distribution function of RNA folding times in a single forward pass, given only a few initial folding time examples. Our method offers a modular extension to existing RNA kinetics algorithms, promising significant computational speed-ups orders of magnitude faster, while achieving comparable results. We showcase the effectiveness of KinPFN through extensive evaluations and real-world case studies, demonstrating its potential for RNA folding kinetics analysis, its practical relevance, and generalization to other biological data.

Diffusion transformers have demonstrated remarkable performance in visual generation tasks, such as generating realistic images or videos based on textual instructions. However, larger model sizes and multi-frame processing for video generation lead to increased computational and memory costs, posing challenges for practical deployment on edge devices. Post-Training Quantization (PTQ) is an effective method for reducing memory costs and computational complexity. When quantizing diffusion transformers, we find that existing quantization methods face challenges when applied to text-to-image and video tasks. To address these challenges, we begin by systematically analyzing the source of quantization error and conclude with the unique challenges posed by DiT quantization. Accordingly, we design an improved quantization scheme: ViDiT-Q (**V**ideo \& **I**mage **Di**ffusion **T**ransformer **Q**uantization), tailored specifically for DiT models. We validate the effectiveness of ViDiT-Q across a variety of text-to-image and video models, achieving W8A8 and W4A8 with negligible degradation in visual quality and metrics. Additionally, we implement efficient GPU kernels to achieve practical 2-2.5x memory optimization and a 1.4-1.7x end-to-end latency speedup.

We train, for the first time, large language models using FP8 precision on datasets up to 2 trillion tokens --- a 20-fold increase over previous limits. Through these extended training runs, we uncover critical instabilities in FP8 training that were not observable in earlier works with shorter durations. We trace these instabilities to outlier amplification by the SwiGLU activation function. Interestingly, we show, both analytically and empirically, that this amplification happens only over prolonged training periods, and link it to a SwiGLU weight alignment process. To address this newly identified issue, we introduce Smooth-SwiGLU, a novel modification that ensures stable FP8 training without altering function behavior. We also demonstrate, for the first time, FP8 quantization of both Adam optimizer moments. Combining these innovations, we successfully train a 7B parameter model using FP8 precision on 256 Intel Gaudi2 accelerators, achieving on-par results with the BF16 baseline while delivering up to a $\sim$ 34 % throughput improvement. A reference implementation is supplied in https://github.com/Anonymous1252022/Megatron-DeepSpeed

Imitation learning with a privileged teacher has proven effective for learning complex control behaviors from high-dimensional inputs, such as images. In this framework, a teacher is trained with privileged task information, while a student tries to predict the actions of the teacher with more limited observations, e.g., in a robot navigation task, the teacher might have access to distances to nearby obstacles, while the student only receives visual observations of the scene. However, privileged imitation learning faces a key challenge: the student might be unable to imitate the teacher's behavior due to partial observability. This problem arises because the teacher is trained without considering if the student is capable of imitating the learned behavior. To address this teacher-student asymmetry, we propose a framework for joint training of the teacher and student policies, encouraging the teacher to learn behaviors that can be imitated by the student despite the latters' limited access to information and its partial observability. Based on the performance bound in imitation learning, we add (i) the approximated action difference between teacher and student as a penalty term to the reward function of the teacher, and (ii) a supervised teacher-student alignment step. We motivate our method with a maze navigation task and demonstrate its effectiveness on complex vision-based quadrotor flight and manipulation tasks.

Boosting has garnered significant interest across both machine learning and statistical communities. Traditional boosting algorithms, designed for fully observed random samples, often struggle with real-world problems, particularly with interval-censored data. This type of data is common in survival analysis and time-to-event studies where exact event times are unobserved but fall within known intervals. Effective handling of such data is crucial in fields like medical research, reliability engineering, and social sciences. In this work, we introduce novel nonparametric boosting methods for regression and classification tasks with interval-censored data. Our approaches leverage censoring unbiased transformations to adjust loss functions and impute transformed responses while maintaining model accuracy. Implemented via functional gradient descent, these methods ensure scalability and adaptability. We rigorously establish their theoretical properties, including optimality and mean squared error trade-offs. Our proposed methods not only offer a robust framework for enhancing predictive accuracy in domains where interval-censored data are common but also complement existing work, expanding the applicability of existing boosting techniques. Empirical studies demonstrate robust performance across various finite-sample scenarios, highlighting the practical utility of our approaches.

We present the first mechanistic evidence that model-free reinforcement learning agents can learn to plan. This is achieved by applying a methodology based on concept-based interpretability to a model-free agent in Sokoban -- a commonly used benchmark for studying planning. Specifically, we demonstrate that DRC, a generic model-free agent introduced by [Guez et al. (2019)](https://arxiv.org/abs/1901.03559), uses learned concept representations to internally formulate plans that both predict the long-term effects of actions on the environment and influence action selection. Our methodology involves: (1) probing for planning-relevant concepts, (2) investigating plan formation within the agent's representations, and (3) verifying that discovered plans (in the agent's representations) have a causal effect on the agent's behavior through interventions. We also show that the emergence of these plans coincides with the emergence of a planning-like property: the ability to benefit from additional test-time compute. Finally, we perform a qualitative analysis of the planning algorithm learned by the agent and discover a strong resemblance to parallelized bidirectional search. Our findings advance understanding of the internal mechanisms underlying planning behavior in agents, which is important given the recent trend of emergent planning and reasoning capabilities in LLMs through RL.

Understanding and modeling pedestrian movements in the real world is crucial for applications like motion forecasting and scene simulation. Many factors influence pedestrian movements, such as scene context, individual characteristics, and goals, which are often ignored by the existing human generation methods. Web videos contain natural pedestrian behavior and rich motion context, but annotating them with pre-trained predictors leads to noisy labels. In this work, we propose learning diverse pedestrian movements from web videos. We first curate a large-scale dataset called CityWalkers that captures diverse real-world pedestrian movements in urban scenes. Then, based on CityWalkers, we propose a generative model called PedGen for diverse pedestrian movement generation. PedGen introduces automatic label filtering to remove the low-quality labels and a mask embedding to train with partial labels. It also contains a novel context encoder that lifts the 2D scene context to 3D and can incorporate various context factors in generating realistic pedestrian movements in urban scenes. Experiments show that PedGen outperforms existing baseline methods for pedestrian movement generation by learning from noisy labels and incorporating the context factors. In addition, PedGen achieves zero-shot generalization in both real-world and simulated environments. The code, model, and data are available at https://genforce.github.io/PedGen/.

We propose a simple heuristic privacy analysis of noisy clipped stochastic gradient descent (DP-SGD) in the setting where only the last iterate is released and the intermediate iterates remain hidden. Namely, our heuristic assumes a linear structure for the model. We show experimentally that our heuristic is predictive of the outcome of privacy auditing applied to various training procedures. Thus it can be used prior to training as a rough estimate of the final privacy leakage. We also probe the limitations of our heuristic by providing some artificial counterexamples where it underestimates the privacy leakage. The standard composition-based privacy analysis of DP-SGD effectively assumes that the adversary has access to all intermediate iterates, which is often unrealistic. However, this analysis remains the state of the art in practice. While our heuristic does not replace a rigorous privacy analysis, it illustrates the large gap between the best theoretical upper bounds and the privacy auditing lower bounds and sets a target for further work to improve the theoretical privacy analyses.

Neural networks are often challenging to work with due to their large size and complexity. To address this, various methods aim to reduce model size by sparsifying or decomposing weight matrices, such as magnitude pruning and low-rank or block-diagonal factorization. In this work, we present Double Sparse Factorization (DSF), where we factorize each weight matrix into two sparse matrices. Although solving this problem exactly is computationally infeasible, we propose an efficient heuristic based on alternating minimization via ADMM that achieves state-of-the-art results, enabling unprecedented sparsification of neural networks. For instance, in a one-shot pruning setting, our method can reduce the size of the LLaMA2-13B model by 50% while maintaining better performance than the dense LLaMA2-7B model. We also compare favorably with Optimal Brain Compression, the state-of-the-art layer-wise pruning approach for convolutional neural networks. Furthermore, accuracy improvements of our method persist even after further model fine-tuning. Code available at: https://github.com/usamec/double_sparse

Controlled text generation allows for enforcing user-defined constraints on large language model outputs, an increasingly important field as LLMs become more prevalent in everyday life. One common approach uses energy-based decoding, which defines a target distribution through an energy function that combines multiple constraints into a weighted average. However, these methods often struggle to balance fluency with constraint satisfaction, even with extensive tuning of the energy function's coefficients. In this paper, we identify that this suboptimal balance arises from sampling in continuous space rather than the natural discrete space of text tokens. To address this, we propose \emph{Discrete Auto-regressive Biasing}, a controlled decoding algorithm that leverages gradients while operating entirely in the discrete text domain. Specifically, we introduce a new formulation for controlled text generation by defining a joint distribution over the generated sequence and an auxiliary bias sequence. To efficiently sample from this joint distribution, we propose a Langevin-within-Gibbs sampling algorithm using gradient-based discrete MCMC. Our method significantly improves constraint satisfaction while maintaining comparable or better fluency, all with lower computational costs. We demonstrate the advantages of our controlled decoding method on sentiment control, language detoxification, and keyword-guided generation.

Room layout estimation from multiple-perspective images is poorly investigated due to the complexities that emerge from multi-view geometry, which requires muti-step solutions such as camera intrinsic and extrinsic estimation, image matching, and triangulation. However, in 3D reconstruction, the advancement of recent 3D foundation models such as DUSt3R has shifted the paradigm from the traditional multi-step structure-from-motion process to an end-to-end single-step approach. To this end, we introduce Plane-DUSt3R, a novel method for multi-view room layout estimation leveraging the 3D foundation model DUSt3R. Plane-DUSt3R incorporates the DUSt3R framework and fine-tunes on a room layout dataset (Structure3D) with a modified objective to estimate structural planes. By generating uniform and parsimonious results, Plane-DUSt3R enables room layout estimation with only a single post-processing step and 2D detection results. Unlike previous methods that rely on single-perspective or panorama image, Plane-DUSt3R extends the setting to handle multiple-perspective images. Moreover, it offers a streamlined, end-to-end solution that simplifies the process and reduces error accumulation. Experimental results demonstrate that Plane-DUSt3R not only outperforms state-of-the-art methods on the synthetic dataset but also proves robust and effective on in the wild data with different image styles such as cartoon. Our code is available at: https://github.com/justacar/Plane-DUSt3R

Skins wrapping around our bodies, leathers covering over the sofa, sheet metal coating the car – it suggests that objects are enclosed by a series of continuous surfaces, which provides us with informative geometry prior for objectness deduction. In this paper, we propose Gaussian-Det which leverages Gaussian Splatting as surface representation for multi-view based 3D object detection. Unlike existing monocular or NeRF-based methods which depict the objects via discrete positional data, Gaussian-Det models the objects in a continuous manner by formulating the input Gaussians as feature descriptors on a mass of partial surfaces. Furthermore, to address the numerous outliers inherently introduced by Gaussian splatting, we accordingly devise a Closure Inferring Module (CIM) for the comprehensive surface-based objectness deduction. CIM firstly estimates the probabilistic feature residuals for partial surfaces given the underdetermined nature of Gaussian Splatting, which are then coalesced into a holistic representation on the overall surface closure of the object proposal. In this way, the surface information Gaussian-Det exploits serves as the prior on the quality and reliability of objectness and the information basis of proposal refinement. Experiments on both synthetic and real-world datasets demonstrate that Gaussian-Det outperforms various existing approaches, in terms of both average precision and recall.

While internet-scale data often come in pairs (e.g., audio+image, image+text), we often want to perform inferences over modalities unseen together in the training data (e.g., audio+text). Prior work has addressed this issue by learning multiple contrastive embedding spaces between existing modality pairs, implicitly hoping that unseen modality pairs will end up being aligned. This theoretical paper proves that this hope is well founded, under certain assumptions. Starting with the proper Bayesian approach of integrating out intermediate modalities, we show that directly comparing the representations of data from unpaired modalities can recover the same likelihood ratio. Our analysis builds on prior work on the geometry and probabilistic interpretation of contrastive representations, showing how these representations can answer many of the same inferences as probabilistic graphical models. Our analysis suggests two new ways of using contrastive representations: in settings with pre-trained contrastive models, and for handling language ambiguity in reinforcement learning. Our numerical experiments study the importance of our assumptions and demonstrate these new applications.

RNA secondary structure prediction is critical for understanding RNA function but remains challenging due to complex structural elements like pseudoknots and limited training data. We introduce DEPfold, a novel deep learning approach that re-frames RNA secondary structure prediction as a dependency parsing problem. DEPfold presents three key innovations: (1) a biologically motivated transformation of RNA structures into labeled dependency trees, (2) a biaffine attention mechanism for joint prediction of base pairings and their types, and (3) an optimal tree decoding algorithm that enforces valid RNA structural constraints. Unlike traditional energy-based methods, DEPfold learns directly from annotated data and leverages pretrained language models to predict RNA structure. We evaluate DEPfold on both within-family and cross-family RNA datasets, demonstrating significant performance improvements over existing methods. DEPfold shows strong performance in cross-family generalization when trained on data augmented by traditional energy-based models, outperforming existing methods on the bpRNAnew dataset. This demonstrates DEPfold’s ability to effectively learn structural information beyond what traditional methods capture. Our approach bridges natural language processing (NLP) with RNA biology, providing a computationally efficient and adaptable tool for advancing RNA structure prediction and analysis

Do diverse perspectives help students learn better? Multi-teacher knowledge distillation, which is a more effective technique than traditional single-teacher methods, supervises the student from different perspectives (i.e., teacher). While effective, multi-teacher, teacher ensemble, or teaching assistant-based approaches are computationally expensive and resource-intensive, as they require training multiple teacher networks. These concerns raise a question: can we supervise the student with diverse perspectives using only a single teacher? We, as the pioneer, demonstrate TeKAP, a novel teacher knowledge augmentation technique that generates multiple synthetic teacher knowledge by perturbing the knowledge of a single pretrained teacher i.e., Teacher Knowledge Augmentation via Perturbation, at both the feature and logit levels. These multiple augmented teachers simulate an ensemble of models together. The student model is trained on both the actual and augmented teacher knowledge, benefiting from the diversity of an ensemble without the need to train multiple teachers. TeKAP significantly reduces training time and computational resources, making it feasible for large-scale applications and easily manageable. Experimental results demonstrate that our proposed method helps existing state-of-the-art knowledge distillation techniques achieve better performance, highlighting its potential as a cost-effective alternative. The source code can be found in the supplementary.

Ensembling strategies for Large Language Models (LLMs) have demonstrated significant potential in improving performance across various tasks by combining the strengths of individual models. However, identifying the most effective ensembling method remains an open challenge, as neither maximizing output consistency through self-consistency decoding nor enhancing model diversity via frameworks like "Mixture of Agents" has proven universally optimal. Motivated by this, we propose a unified framework to examine the trade-offs between task performance, model diversity, and output consistency in ensembles. More specifically, we introduce a consistency score that defines a gating mechanism for mixtures of agents and an algorithm for mixture refinement to investigate these trade-offs at the semantic and model levels, respectively. We incorporate our insights into a novel inference-time LLM ensembling strategy called the Dynamic Mixture of Agents (DMoA) and demonstrate that it achieves a new state-of-the-art result in the challenging Big Bench Hard mixed evaluations benchmark. Our analysis reveals that cross-validation bias can enhance performance, contingent on the expertise of the constituent models. We further demonstrate that distinct reasoning tasks—such as arithmetic reasoning, commonsense reasoning, and instruction following—require different model capabilities, leading to inherent task-dependent trade-offs that DMoA balances effectively.

An increasing interest in learning to forecast for time-series of high-dimensional observations is the ability to adapt to systems with diverse underlying dynamics. Access to observations that define a stationary distribution of these systems is often unattainable, as the underlying dynamics may change over time. Naively training or retraining models at each shift may lead to catastrophic forgetting about previously-seen systems. We present a new continual meta-learning (CML) framework to realize continual slow-and fast adaptation of latent dynamics (CoSFan). We leverage a feed-forward meta-model to infer *what* the current system is and *how* to adapt a latent dynamics function to it, enabling *fast adaptation* to specific dynamics. We then develop novel strategies to automatically detect *when* a shift of data distribution occurs, with which to identify its underlying dynamics and its relation with previously-seen dynamics. In combination with fixed-memory experience replay mechanisms, this enables continual *slow update* of the *what-how* meta-model. Empirical studies demonstrated that both the meta- and continual-learning component was critical for learning to forecast across non-stationary distributions of diverse dynamics systems, and the feed-forward meta-model combined with task-aware/-relational continual learning strategies significantly outperformed existing CML alternatives.