Reinforcement learning (RL) has emerged as the predominant paradigm for training large language model (LLM) agents to solve complex, multi-step tasks through environmental interaction. A fundamental challenge in such long-horizon scenarios is credit assignment, as delayed rewards provide inadequate signals for evaluating individual action contributions. Existing methods typically neglect trajectory transition dynamics, which leads to coarse-grained or biased credit assignment. To address these limitations, we introduce SHADOW, a novel framework that systematically incorporates transition dynamics for improved credit assignment. Our framework makes two primary contributions: (i) a dynamics-aware state grouping mechanism that mitigates misleading action comparisons between dynamically inconsistent states, and (ii) a local dynamic advantage estimator that leverages Generalized Advantage Estimation (GAE) to precisely quantify individual action contributions through a fine-grained analysis of transition patterns. Comprehensive experiments conducted with the Qwen2.5-1.5/7B-Instruct agent model demonstrate that our method achieves success rate improvements of 9.4%/7.6% on the ALFworld benchmark and a performance gain of over 5% on WebShop.

In this work, we argue that large language models (LLMs), though trained to predict only the next token, exhibit emergent planning behaviors: $\textbf{their hidden representations encode future outputs beyond the next token}$. Through simple probing, we demonstrate that LLM prompt representations encode global attributes of their entire responses, including $\textit{structure attributes}$ (e.g., response length, reasoning steps), $\textit{content attributes}$ (e.g., character choices in storywriting, multiple-choice answers at the end of response), and $\textit{behavior attributes}$ (e.g., answer confidence, factual consistency). In addition to identifying response planning, we explore how it scales with model size across tasks and how it evolves during generation. The findings that LLMs plan ahead for the future in their hidden representations suggest potential applications for improving transparency and generation control.

One way to mitigate risks in vision-language models (VLMs) is to censor dangerous samples from their training data. However, data moderation can be easily bypassed when harmful images are split into small, benign-looking patches, scattered across many training samples. VLMs may then learn to piece these fragments together and generate harmful responses at inference, either from full images or text references. For instance, if trained on image patches from a bloody scene paired with the descriptions "safe," VLMs may later describe, the full image or a text reference to the scene, as "safe." We define the core ability of VLMs enabling this attack as $\textit{visual stitching}$—the ability to integrate visual information spread across multiple training samples that share the same textual descriptions. In our work, we first demonstrate visual stitching abilities in common open-source VLMs on three datasets where each image is labeled with a unique synthetic ID. We split each $(\texttt{image}, \texttt{ID})$ pair into $\{(\texttt{patch}, \texttt{ID})\}$ pairs at different granularities for finetuning, and we find that models can verbalize the correct IDs from full images or text reference. Building on this, we simulate the adversarial data poisoning scenario mentioned above by using patches from dangerous images and replacing IDs with text descriptions like "safe" or "unsafe", demonstrating how harmful content can evade moderation in patches and later be reconstructed through visual stitching, posing serious VLM safety risks.

Recent advancements in offline reinforcement learning (RL) have underscored the capabilities of Return-Conditioned Supervised Learning (RCSL), a paradigm that learns the action distribution based on target returns for each state in a supervised manner. However, prevailing RCSL methods largely focus on deterministic trajectory modeling, disregarding stochastic state transitions and the diversity of future trajectory distributions. A fundamental challenge arises from the inconsistency between the sampled returns within individual trajectories and the expected returns across multiple trajectories. Fortunately, value-based methods offer a solution by leveraging a value function to approximate the expected returns, thereby addressing the inconsistency effectively. Building upon these insights, we propose a novel approach, termed the Critic-Guided Decision Transformer (CGDT), which combines the predictability of long-term returns from value-based methods with the trajectory modeling capability of the Decision Transformer. By incorporating a learned value function, known as the critic, CGDT ensures a direct alignment between the specified target returns and the expected returns of actions. This integration bridges the gap between the deterministic nature of RCSL and the probabilistic characteristics of value-based methods. Empirical evaluations on stochastic environments and D4RL benchmark datasets demonstrate the superiority of CGDT over traditional RCSL methods. These results highlight the potential of CGDT to advance the state of the art in offline RL and extend the applicability of RCSL to a wide range of RL tasks.

Elucidating the reasoning process with structured explanations from question to answer is crucial, as it significantly enhances the interpretability, traceability, and trustworthiness of question-answering (QA) systems.However, structured explanations demand models to perform intricately structured reasoning, which poses great challenges.Most existing methods focus on single-step reasoning through supervised learning, ignoring logical dependencies between steps.Moreover, existing reinforcement learning (RL) based methods overlook the structured relationships, underutilizing the potential of RL in structured reasoning.In this paper, we propose SEER, a novel method that maximizes a structure-based return to facilitate structured reasoning and explanation.Our proposed structure-based return precisely describes the hierarchical and branching structure inherent in structured reasoning, effectively capturing the intricate relationships between different reasoning steps.In addition, we introduce a fine-grained reward function to meticulously delineate diverse reasoning steps.Extensive experiments show that SEER significantly outperforms state-of-theart methods, achieving an absolute improvement of 6.9% over RL-based methods on En-tailmentBank, a 4.4% average improvement on STREET benchmark, and exhibiting outstanding efficiency and cross-dataset generalization performance.Our code is available at https://github.com/Chen-GX/SEER.

Zhanhui Zhou, Jie Liu, Zhichen Dong, Jiaheng Liu, Chao Yang, Wanli Ouyang, Yu Qiao. Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2024.

Significant progress in video question answering (VideoQA) have been made thanks to thriving large image-language pretraining frameworks. Although image-language models can efficiently represent both video and language branches, they typically employ goal-free vision perception and do not interact vision with language well during the answer generation, thus omitting crucial visual cues. In this paper, we are inspired by the human recognition and learning pattern and propose VideoDistill, a framework with language-aware (i.e., goal-driven) behavior in both vision perception and answer generation. VideoDistill generates answers only from question-related visual embeddings and follows a thinking-observing-answering approach that closely resembles human behavior, distinguishing it from previous research. Specifically, we develop a language-aware gating mechanism to replace the standard cross-attention, avoiding language's direct fusion into visual representations. We incorporate this mechanism into two key components of the entire framework. The first component is a differentiable sparse sampling module, which selects frames containing the necessary dynamics and semantics relevant to the questions. The second component is a vision refinement module that merges existing spatial-temporal attention layers to ensure extracting multi-grained visual semantics associated with the questions. We conduct evaluations on various challenging video question-answering benchmarks, and VideoDistill achieves state-of-the-art performance in both general and long-form VideoQA datasets. In Addition, we verify that VideoDistill can effectively alleviate the utilization of language shortcut solutions in the EgoTaskQA dataset.

Existing methods to fine-tune LLMs, like Adapter, Prefix-tuning, and LoRA, which introduce extra modules or ad-ditional input sequences to inject new skills or knowledge, may compromise the innate abilities of LLMs. In this paper, we propose LLaMA-Excitor, a lightweight method that stimulates the LLMs' potential to better follow instructions by gradually paying more attention to worthwhile information. Specifically, LLaMA-Excitor does not directly change the intermediate hidden state during the self-attention calculation. We designed the Excitor block as a bypass module that reconstructs Keys and changes the importance of Values in self-attention using learnable prompts. LLaMA-Excitor ensures a self-adaptive allocation of additional attention to input instructions, thus effectively preserving LLMs' pre-trained knowledge when fine-tuning LLMs on low-quality instruction-following datasets. Furthermore, we unify the modeling of multi-modal and language-only tuning, extending LLaMA-Excitor to a powerful visual instruction follower without the need for complex multi-modal alignment. Our approach is evaluated in language-only and multi-modal scenarios. Compared with the original LLaMA-7B, LLaMA-Excitor is the only PEFT method that maintains basic capabilities and achieves +3.12% relative improvement on the MMLU benchmark. In the visual instruction tuning, we achieve a new state-of-the-art image captioning performance on MSCOCO (157.5 CIDEr), and a comparable performance on ScienceQA (88.39%) to cutting-edge models with more parameters and extensive vision-language pertaining. The code will be available at https://zoubo9034.github.io/Excitor/.

Language-conditioned robot behavior plays a vital role in executing complex tasks by associating human commands or instructions with perception and actions. The ability to compose long-horizon tasks based on unconstrained language instructions necessitates the acquisition of a diverse set of general-purpose skills.However, acquiring inherent primitive skills in a coupled and long-horizon environment without external rewards or human supervision presents significant challenges. In this paper, we evaluate the relationship between skills and language instructions from a mathematical perspective, employing two forms of mutual information within the framework of language-conditioned policy learning.To maximize the mutual information between language and skills in an unsupervised manner, we propose an end-to-end imitation learning approach known as Language Conditioned Skill Discovery (LCSD). Specifically, we utilize vector quantization to learn discrete latent skills and leverage skill sequences of trajectories to reconstruct high-level semantic instructions.Through extensive experiments on language-conditioned robotic navigation and manipulation tasks, encompassing BabyAI, LORel, and Calvin, we demonstrate the superiority of our method over prior works. Our approach exhibits enhanced generalization capabilities towards unseen tasks, improved skill interpretability, and notably higher rates of task completion success.

Robotic behavior synthesis, the problem of understanding multimodal inputs and generating precise physical control for robots, is an important part of Embodied AI. Despite successes in applying multimodal large language models for high-level understanding, it remains challenging to translate these conceptual understandings into detailed robotic actions while achieving generalization across various scenarios. In this paper, we propose a tree-structured multimodal code generation framework for generalized robotic behavior synthesis, termed RoboCodeX. RoboCodeX decomposes high-level human instructions into multiple object-centric manipulation units consisting of physical preferences such as affordance and safety constraints, and applies code generation to introduce generalization ability across various robotics platforms. To further enhance the capability to map conceptual and perceptual understanding into control commands, a specialized multimodal reasoning dataset is collected for pre-training and an iterative self-updating methodology is introduced for supervised fine-tuning. Extensive experiments demonstrate that RoboCodeX achieves state-of-the-art performance in both simulators and real robots on four different kinds of manipulation tasks and one embodied navigation task.

Zhichen Dong, Zhanhui Zhou, Chao Yang, Jing Shao, Yu Qiao. Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers). 2024.

Large language models are usually fine-tuned to align with human preferences. However, fine-tuning a large language model can be challenging. In this work, we introduce $\textit{weak-to-strong search}$, framing the alignment of a large language model as a test-time greedy search to maximize the log-probability difference between small tuned and untuned models while sampling from the frozen large model. This method serves both as (1) a compute-efficient model up-scaling strategy that avoids directly tuning the large model and as (2) an instance of weak-to-strong generalization that enhances a strong model with weak test-time guidance. Empirically, we demonstrate the flexibility of weak-to-strong search across different tasks. In controlled-sentiment generation and summarization, we use tuned and untuned $\texttt{gpt2}$s to improve the alignment of large models without additional training. Crucially, in a more difficult instruction-following benchmark, AlpacaEval 2.0, we show that reusing off-the-shelf small models (e.g., $\texttt{zephyr-7b-beta}$ and its untuned version) can improve the length-controlled win rates of both white-box and black-box large models against $\texttt{gpt-4-turbo}$ (e.g., $34.4\% \rightarrow 37.9\%$ for $\texttt{Llama-3-70B-Instruct}$ and $16.0\% \rightarrow 20.1\%$ for $\texttt{gpt-3.5-turbo-instruct}$), despite the small models' low win rates $\approx 10.0\%$.

Keypoint-based representation has proven advantageous in various visual and robotic tasks. However, the existing 2D and 3D methods for detecting keypoints mainly rely on geometric consistency to achieve spatial alignment, neglecting temporal consistency. To address this issue, the Transporter method was introduced for 2D data, which reconstructs the target frame from the source frame to incorporate both spatial and temporal information. However, the direct application of the Transporter to 3D point clouds is infeasible due to their structural differences from 2D images. Thus, we propose the first 3D version of the Transporter, which leverages hybrid 3D representation, cross attention, and implicit reconstruction. We apply this new learning system on 3D articulated objects and non-rigid animals (humans and rodents) and show that learned keypoints are spatio-temporally consistent. Additionally, we propose a closed-loop control strategy that utilizes the learned keypoints for 3D object manipulation and demonstrate its superior performance. Codes are available at https://github.com/zhongcl-thu/3D-Implicit-Transporter.

The tremendous progress of vision-to-language retrieval over these years is fueled by contrastive vision-language pretraining (VLP), such as CLIP. Although, contrastive methods do not exhibit the same level of performance on other downstream tasks (e.g., video question answering and natural language grounding). One possible reason is they ignore the misalignment between vision and language, especially the absence of spatial information in language. To mitigate this issue, We start from a new perspective and propose a contrastive VLP framework with spatial reconstruction on text (SpaceCLIP). Specifically, we introduce a unique reconstruction method to assign text representations into the same spatial structure with images or videos and a pretraining objective, SpatialNCE, to reduce the computational overhead and ensure performance on downstream tasks. Empirically, we show SpaceCLIP outperforms other methods with performance gains ranging from 2.1% up to 9.0% on MSRVTT and EgoCLIP multiple-choice questions answering, 2.5% up to 11.0% on EPIC-KITCHENS-100 and MSRVTT multi-instance retrieval, and 0.31% up to 7.2% on Ego4D natural language query benchmark.

Visual navigation involves a movable robotic agent striving to reach a point goal (target location) using vision sensory input. While navigation with ideal visibility has seen plenty of success, it becomes challenging in suboptimal visual conditions like poor illumination, where traditional approaches suffer from severe performance degradation. We propose E3VN (echo-enhanced embodied visual navigation) to effectively perceive the surroundings even under poor visibility to mitigate this problem. This is made possible by adopting an echoer that actively perceives the environment via auditory signals. E3VN models the robot agent as playing a cooperative Markov game with that echoer. The action policies of robot and echoer are jointly optimized to maximize the reward in a two-stream actor-critic architecture. During optimization, the reward is also adaptively decomposed into the robot and echoer parts. Our experiments and ablation studies show that E3VN is consistently effective and robust in point goal navigation tasks, especially under nonideal visibility.

Keypoint detection and description play a central role in computer vision. Most existing methods are in the form of scene-level prediction, without returning the object classes of different keypoints. In this paper, we propose the object-centric formulation, which, beyond the conventional setting, requires further identifying which object each interest point belongs to. With such fine-grained information, our framework enables more downstream potentials, such as object-level matching and pose estimation in a clustered environment. To get around the difficulty of label collection in the real world, we develop a sim2real contrastive learning mechanism that can generalize the model trained in simulation to real-world applications. The novelties of our training method are three-fold: (i) we integrate the uncertainty into the learning framework to improve feature description of hard cases, e.g., less-textured or symmetric patches; (ii) we decouple the object descriptor into two independent branches, intra-object salience and inter-object distinctness, resulting in a better pixel-wise description; (iii) we enforce cross-view semantic consistency for enhanced robustness in representation learning. Comprehensive experiments on image matching and 6D pose estimation verify the encouraging generalization ability of our method. Particularly for 6D pose estimation, our method significantly outperforms typical unsupervised/sim2real methods, achieving a closer gap with the fully supervised counterpart.

Deep reinforcement learning has made significant progress in robotic manipulation tasks and it works well in the ideal disturbance-free environment. However, in a real-world environment, both internal and external disturbances are inevitable, thus the performance of the trained policy will dramatically drop. To improve the robustness of the policy, we introduce the adversarial training mechanism to the robotic manipulation tasks in this paper, and an adversarial skill learning algorithm based on soft actor-critic (SAC) is proposed for robust manipulation. Extensive experiments are conducted to demonstrate that the learned policy is robust to internal and external disturbances. Additionally, the proposed algorithm is evaluated in both the simulation environment and on the real robotic platform.

In this paper, we study Reinforcement Learning from Demonstrations (RLfD) that improves the exploration efficiency of Reinforcement Learning (RL) by providing expert demonstrations. Most of existing RLfD methods require demonstrations to be perfect and sufficient, which yet is unrealistic to meet in practice. To work on imperfect demonstrations, we first define an imperfect expert setting for RLfD in a formal way, and then point out that previous methods suffer from two issues in terms of optimality and convergence, respectively. Upon the theoretical findings we have derived, we tackle these two issues by regarding the expert guidance as a soft constraint on regulating the policy exploration of the agent, which eventually leads to a constrained optimization problem. We further demonstrate that such problem is able to be addressed efficiently by performing a local linear search on its dual form. Considerable empirical evaluations on a comprehensive collection of benchmarks indicate our method attains consistent improvement over other RLfD counterparts.

Satellite image dehazing aims at precisely retrieving the real situations of the obscured parts from the hazy remote sensing (RS) images, which is a challenging task since the hazy regions contain both ground features and haze components. Many approaches of removing haze focus on processing multi-spectral or RGB images, whereas few of them utilize multi-sensor data. The multi-sensor data fusion is significant to provide auxiliary information since RGB images are sensitive to atmospheric conditions. In this paper, a dataset called SateHaze1k is established and composed of 1200 pairs clear Synthetic Aperture Radar (SAR), hazy RGB, and corresponding ground truth images, which are divided into three degrees of the haze, i.e. thin, moderate, and thick fog. Moreover, we propose a novel fusion dehazing method to directly restore the haze-free RS images by using an end-to-end conditional generative adversarial network(cGAN). The proposed network combines the information of both RGB and SAR images to eliminate the image blurring. Besides, the dilated residual blocks of the generator can also sufficiently improve the dehazing effects. Our experiments demonstrate that the proposed method, which fuses the information of different sensors applied to the cloudy conditions, can achieve more precise results than other baseline models.

This paper studies Learning from Observations (LfO) for imitation learning with access to state-only demonstrations. In contrast to Learning from Demonstration (LfD) that involves both action and state supervisions, LfO is more practical in leveraging previously inapplicable resources (e.g., videos), yet more challenging due to the incomplete expert guidance. In this paper, we investigate LfO and its difference with LfD in both theoretical and practical perspectives. We first prove that the gap between LfD and LfO actually lies in the disagreement of inverse dynamics models between the imitator and expert, if following the modeling approach of GAIL. More importantly, the upper bound of this gap is revealed by a negative causal entropy which can be minimized in a model-free way. We term our method as Inverse-Dynamics-Disagreement-Minimization (IDDM) which enhances the conventional LfO method through further bridging the gap to LfD. Considerable empirical results on challenging benchmarks indicate that our method attains consistent improvements over other LfO counterparts.

Humans' fingertips can perceive not only the magnitude and the direction of force but also the texture of object. When we grasp an object, the surface texture sensing of the fingertip helps us recognize the object and the force feeling that is parallel to the skin helps us grasp stably. Focusing on these points, we have developed a dual-modal vision-based tactile sensor that can measure the texture of object and a distribution of force vectors. The tactile sensor consists of a transparent elastomer, a camera, a piece of transparent acrylic board, LEDs and supporting structures. A reflective membrane and markers array are on the surface of the elastomer. An applied force on the elastic body results in movements of the markers, which are acquired by the CCD camera. In addition, the shape and texture of the object's contact surface can be reflected by the membrane deformations. The distribution of force vectors is determined by the BP neural network. The local binary pattern algorithm using captured images calculates the texture information. This paper reports experimental evaluation results concerning accuracy of determination of magnitude, direction of force, and texture recognition rate.

Insufficient training data is a serious problem in all domains related to bioinformatics. Large-scale annotated electroencephalography (EEG) datasets are almost impossible to acquire because biological data acquisition is challenging and quality annotation is costly. Transfer learning relaxes the hypothesis that the training data must be independent and identically distributed (i.i. d.) with the test data, which motivates us to use transfer learning to solve the problem of insufficient training data in bioinformatics. We propose a new approach to transfer knowledge via a deep transfer learning framework, which includes an adaptive sample selection algorithm and a joint adversarial training algorithm. The adaptive sample selection algorithm dynamically adjusts the sample weights during the training process according to the distance between the source domain and the target domain. The joint adversarial training algorithm forces the network to learn a feature extractor suitable for the target domain based on a dataset from the source domain by using an adversarial network and a specific loss function. The experiments demonstrate that our approach has many advantages, such as robustness and accuracy, when applied to EEG classification tasks.