EPO: Explicit Policy Optimization for Strategic Reasoning in LLMs via Reinforcement Learning

Large Language Models (LLMs) have shown impressive reasoning capabilities in well-defined problems with clear solutions, such as mathematics and coding. However, they still struggle with complex real-world scenarios like business negotiations, which require strategic reasoning-an ability to navigate dynamic environments and align long-term goals amidst uncertainty. Existing methods for strategic reasoning face challenges in adaptability, scalability, and transferring strategies to new contexts. To address these issues, we propose explicit policy optimization (EPO) for strategic reasoning, featuring an LLM that provides strategies in open-ended action space and can be plugged into arbitrary LLM agents to motivate goal-directed behavior. To improve adaptability and policy transferability, we train the strategic reasoning model via multi-turn reinforcement learning (RL) using process rewards and iterative self-play, without supervised fine-tuning (SFT) as a preliminary step. Experiments across social and physical domains demonstrate EPO's ability of long-term goal alignment through enhanced strategic reasoning, achieving state-of-the-art performance on social dialogue and web navigation tasks. Our findings reveal various collaborative reasoning mechanisms emergent in EPO and its effectiveness in generating novel strategies, underscoring its potential for strategic reasoning in real-world applications.

EquiBench: Benchmarking Code Reasoning Capabilities of Large Language Models via Equivalence Checking

Equivalence checking, i.e., determining whether two programs produce identical outputs for all possible inputs, underpins a broad range of applications, including software refactoring, testing, and optimization. We present the task of equivalence checking as a new way to evaluate the code reasoning abilities of large language models (LLMs). We introduce EquiBench, a dataset of 2400 program pairs spanning four programming languages and six equivalence categories. These pairs are systematically generated through program analysis, compiler scheduling, and superoptimization, covering nontrivial structural transformations that demand deep semantic reasoning beyond simple syntactic variations. Our evaluation of 17 state-of-the-art LLMs shows that OpenAI o3-mini achieves the highest overall accuracy of 78.0%. In the most challenging categories, the best accuracies are 62.3% and 68.8%, only modestly above the 50% random baseline for binary classification, indicating significant room for improvement in current models' code reasoning capabilities.

SLVR: Securely Leveraging Client Validation for Robust Federated Learning

Federated Learning (FL) enables collaborative model training while keeping client data private. However, exposing individual client updates makes FL vulnerable to reconstruction attacks. Secure aggregation mitigates such privacy risks but prevents the server from verifying the validity of each client update, creating a privacy-robustness tradeoff. Recent efforts attempt to address this tradeoff by enforcing checks on client updates using zero-knowledge proofs, but they support limited predicates and often depend on public validation data. We propose SLVR, a general framework that securely leverages clients' private data through secure multi-party computation. By utilizing clients' data, SLVR not only eliminates the need for public validation data, but also enables a wider range of checks for robustness, including cross-client accuracy validation. It also adapts naturally to distribution shifts in client data as it can securely refresh its validation data up-to-date. Our empirical evaluations show that SLVR improves robustness against model poisoning attacks, particularly outperforming existing methods by up to 50% under adaptive attacks. Additionally, SLVR demonstrates effective adaptability and stable convergence under various distribution shift scenarios.

BounTCHA: A CAPTCHA Utilizing Boundary Identification in AI-extended Videos

In recent years, the rapid development of artificial intelligence (AI) especially multi-modal Large Language Models (MLLMs), has enabled it to understand text, images, videos, and other multimedia data, allowing AI systems to execute various tasks based on human-provided prompts. However, AI-powered bots have increasingly been able to bypass most existing CAPTCHA systems, posing significant security threats to web applications. This makes the design of new CAPTCHA mechanisms an urgent priority. We observe that humans are highly sensitive to shifts and abrupt changes in videos, while current AI systems still struggle to comprehend and respond to such situations effectively. Based on this observation, we design and implement BounTCHA, a CAPTCHA mechanism that leverages human perception of boundaries in video transitions and disruptions. By utilizing AI's capability to expand original videos with prompts, we introduce unexpected twists and changes to create a pipeline for generating short videos for CAPTCHA purposes. We develop a prototype and conduct experiments to collect data on humans' time biases in boundary identification. This data serves as a basis for distinguishing between human users and bots. Additionally, we perform a detailed security analysis of BounTCHA, demonstrating its resilience against various types of attacks. We hope that BounTCHA will act as a robust defense, safeguarding millions of web applications in the AI-driven era.

SDPO: Segment-Level Direct Preference Optimization for Social Agents

Social agents powered by large language models (LLMs) can simulate human social behaviors but fall short in handling complex goal-oriented social dialogues. Direct Preference Optimization (DPO) has proven effective in aligning LLM behavior with human preferences across a variety of agent tasks. Existing DPO-based approaches for multi-turn interactions are divided into turn-level and session-level methods. The turn-level method is overly fine-grained, focusing exclusively on individual turns, while session-level methods are too coarse-grained, often introducing training noise. To address these limitations, we propose Segment-Level Direct Preference Optimization (SDPO), which focuses on specific key segments within interactions to optimize multi-turn agent behavior while minimizing training noise. Evaluations on the SOTOPIA benchmark demonstrate that SDPO-tuned agents consistently outperform both existing DPO-based methods and proprietary LLMs like GPT-4o, underscoring SDPO's potential to advance the social intelligence of LLM-based agents. We release our code and data at https://github.com/AlibabaResearch/DAMO-ConvAI/tree/main/SDPO.

Algorithmic Strategies for Sustainable Reuse of Neural Network Accelerators with Permanent Faults

Hardware failures are a growing challenge for machine learning accelerators, many of which are based on systolic arrays. When a permanent hardware failure occurs in a systolic array, existing solutions include localizing and isolating the faulty processing element (PE), using a redundant PE for re-execution, or in some extreme cases decommissioning the entire accelerator for further investigation. In this paper, we propose novel algorithmic approaches that mitigate permanent hardware faults in neural network (NN) accelerators by uniquely integrating the behavior of the faulty component instead of bypassing it. In doing so, we aim for a more sustainable use of the accelerator where faulty hardware is neither bypassed nor discarded, instead being given a second life. We first introduce a CUDA-accelerated systolic array simulator in PyTorch, which enabled us to quantify the impact of permanent faults appearing on links connecting two PEs or in weight registers, where one bit is stuck at 0 or 1 in the float32, float16, or bfloat16 representation. We then propose several algorithmic mitigation techniques for a subset of stuck-at faults, such as Invertible Scaling or Shifting of activations and weights, or fine tuning with the faulty behavior. Notably, the proposed techniques do not require any hardware modification, instead relying on existing components of widely used systolic array based accelerators, such as normalization, activation, and storage units. Extensive experimental evaluations using fully connected and convolutional NNs trained on MNIST, CIFAR-10 and ImageNet show that the proposed fault-tolerant approach matches or gets very close to the original fault-free accuracy.

LLaVA-Zip: Adaptive Visual Token Compression with Intrinsic Image Information

Multi-modal large language models (MLLMs) utilizing instruction-following data, such as LLaVA, have achieved great progress in the industry. A major limitation in these models is that visual tokens consume a substantial portion of the maximum token limit in large language models (LLMs), leading to increased computational demands and decreased performance when prompts include multiple images or videos. Industry solutions often mitigate this issue by increasing computational power, but this approach is less feasible in academic environments with limited resources. In this study, we propose Dynamic Feature Map Reduction (DFMR) based on LLaVA-1.5 to address the challenge of visual token overload. DFMR dynamically compresses the visual tokens, freeing up token capacity. Our experimental results demonstrate that integrating DFMR into LLaVA-1.5 significantly improves the performance of LLaVA in varied visual token lengths, offering a promising solution for extending LLaVA to handle multi-image and video scenarios in resource-constrained academic environments and it can also be applied in industry settings for data augmentation to help mitigate the scarcity of open-domain image-text pair datasets in the continued pretraining stage.

SAM Decoding: Speculative Decoding via Suffix Automaton

Large Language Models (LLMs) have revolutionized natural language processing by unifying tasks into text generation, yet their large parameter sizes and autoregressive nature limit inference speed. SAM-Decoding addresses this by introducing a novel retrieval-based speculative decoding method that uses a suffix automaton for efficient and accurate draft generation. Unlike n-gram matching used by the existing method, SAM-Decoding finds the longest suffix match in generating text and text corpuss, achieving an average time complexity of $O(1)$ per generation step. SAM-Decoding constructs static and dynamic suffix automatons for the text corpus and input prompts, respectively, enabling fast and precise draft generation. Meanwhile, it is designed as an approach that can be combined with existing methods, allowing SAM-Decoding to adaptively select a draft generation strategy based on the matching length, thus increasing the inference speed of the LLM. When combined with Token Recycling, evaluations show SAM-Decoding outperforms existing model-free methods, achieving a speedup of $2.27\times$ over autoregressive decoding on Spec-Bench. When combined with EAGLE2, it reaches a speedup of $2.49\times$, surpassing all current approaches. Our code is available at https://github.com/hyx1999/SAM-Decoding.

Wireless Resource Allocation with Collaborative Distributed and Centralized DRL under Control Channel Attacks

In this paper, we consider a wireless resource allocation problem in a cyber-physical system (CPS) where the control channel, carrying resource allocation commands, is subjected to denial-of-service (DoS) attacks. We propose a novel concept of collaborative distributed and centralized (CDC) resource allocation to effectively mitigate the impact of these attacks. To optimize the CDC resource allocation policy, we develop a new CDC-deep reinforcement learning (DRL) algorithm, whereas existing DRL frameworks only formulate either centralized or distributed decision-making problems. Simulation results demonstrate that the CDC-DRL algorithm significantly outperforms state-of-the-art DRL benchmarks, showcasing its ability to address resource allocation problems in large-scale CPSs under control channel attacks.

LiNeS: Post-training Layer Scaling Prevents Forgetting and Enhances Model Merging

Large pre-trained models exhibit impressive zero-shot performance across diverse tasks, but fine-tuning often leads to catastrophic forgetting, where improvements on a target domain degrade generalization on other tasks. To address this challenge, we introduce LiNeS, Layer-increasing Network Scaling, a post-training editing technique designed to preserve pre-trained generalization while enhancing fine-tuned task performance. LiNeS scales parameter updates linearly based on their layer depth within the network, maintaining shallow layers close to their pre-trained values to preserve general features while allowing deeper layers to retain task-specific representations. We further extend this approach to multi-task model merging scenarios, where layer-wise scaling of merged parameters reduces negative task interference. LiNeS demonstrates significant improvements in both single-task and multi-task settings across various benchmarks in vision and natural language processing. It mitigates forgetting, enhances out-of-distribution generalization, integrates seamlessly with existing multi-task model merging baselines improving their performance across benchmarks and model sizes, and can boost generalization when merging LLM policies aligned with different rewards via RLHF. Importantly, our method is simple to implement and complementary to many existing techniques.