Do NOT Think That Much for 2+3=? On the Overthinking of o1-Like LLMs

The remarkable performance of models like the OpenAI o1 can be attributed to their ability to emulate human-like long-time thinking during inference. These models employ extended chain-of-thought (CoT) processes, exploring multiple strategies to enhance problem-solving capabilities. However, a critical question remains: How to intelligently and efficiently scale computational resources during testing. This paper presents the first comprehensive study on the prevalent issue of overthinking in these models, where excessive computational resources are allocated for simple problems with minimal benefit. We introduce novel efficiency metrics from both outcome and process perspectives to evaluate the rational use of computational resources by o1-like models. Using a self-training paradigm, we propose strategies to mitigate overthinking, streamlining reasoning processes without compromising accuracy. Experimental results show that our approach successfully reduces computational overhead while preserving model performance across a range of testsets with varying difficulty levels, such as GSM8K, MATH500, GPQA, and AIME.

Teaching LLMs to Refine with Tools

Large language models (LLMs) can refine their responses based on feedback, enabling self-improvement through iterative training or test-time refinement. However, existing methods predominantly focus on refinement within the same reasoning format, which may lead to non-correcting behaviors. We propose CaP, a novel approach that uses external tools to refine chain-of-thought (CoT) responses generated by the same or other LLMs. CaP employs a two-stage training process: supervised fine-tuning followed by preference optimization with DPO variants. Our observations highlight the critical role of preference optimization in enabling effective refinement. Additionally, we compare several sampling strategies to leverage CoT and tools at inference time. Experimental results demonstrate CaP's potential for effective cross-reasoning refinement and efficient inference.

Findings of the WMT 2024 Shared Task on Discourse-Level Literary Translation

Following last year, we have continued to host the WMT translation shared task this year, the second edition of the Discourse-Level Literary Translation. We focus on three language directions: Chinese-English, Chinese-German, and Chinese-Russian, with the latter two ones newly added. This year, we totally received 10 submissions from 5 academia and industry teams. We employ both automatic and human evaluations to measure the performance of the submitted systems. The official ranking of the systems is based on the overall human judgments. We release data, system outputs, and leaderboard at https://www2.statmt.org/wmt24/literary-translation-task.html.

Critical Tokens Matter: Token-Level Contrastive Estimation Enhances LLM's Reasoning Capability

Large Language Models (LLMs) have exhibited remarkable performance on reasoning tasks. They utilize autoregressive token generation to construct reasoning trajectories, enabling the development of a coherent chain of thought. In this work, we explore the impact of individual tokens on the final outcomes of reasoning tasks. We identify the existence of ``critical tokens'' that lead to incorrect reasoning trajectories in LLMs. Specifically, we find that LLMs tend to produce positive outcomes when forced to decode other tokens instead of critical tokens. Motivated by this observation, we propose a novel approach - cDPO - designed to automatically recognize and conduct token-level rewards for the critical tokens during the alignment process. Specifically, we develop a contrastive estimation approach to automatically identify critical tokens. It is achieved by comparing the generation likelihood of positive and negative models. To achieve this, we separately fine-tune the positive and negative models on various reasoning trajectories, consequently, they are capable of identifying identify critical tokens within incorrect trajectories that contribute to erroneous outcomes. Moreover, to further align the model with the critical token information during the alignment process, we extend the conventional DPO algorithms to token-level DPO and utilize the differential likelihood from the aforementioned positive and negative model as important weight for token-level DPO learning.Experimental results on GSM8K and MATH500 benchmarks with two-widely used models Llama-3 (8B and 70B) and deepseek-math (7B) demonstrate the effectiveness of the propsoed approach cDPO.

Draft Model Knows When to Stop: A Self-Verification Length Policy for Speculative Decoding

Speculative Decoding (SD) has become an important technique in accelerating the inference speed of large language models. Conventional SD methods employ a fixed draft length, which ignores the token generation difficulty across tasks. Consequently, in this paper, we address such an issue and introduce SVIP - a difficulty-aware dynamic draft length policy for speculative decoding systems. Based on a theoretical lower bound of draft token acceptance rate and its inference-time approximation, SVIP adaptively determines the lengths of draft sequences based on the entropy of each draft token distribution. Experimental results on mainstream SD benchmarks and frameworks demonstrate the superior performance of SVIP, achieving up to 20\% walltime speedup on SpecBench over baseline SD methods and 60\% speedup on MT-Bench for long-form generation of up to 8K tokens. Moreover, SVIP is totally training-free and compatible with any existing SD methods that generate draft tokens autoregressively. Experimental results also show that SVIP yields consistent walltime improvement on top of GliDe & CaPE and EAGLE-2.

Identify Backdoored Model in Federated Learning via Individual Unlearning

Backdoor attacks present a significant threat to the robustness of Federated Learning (FL) due to their stealth and effectiveness. They maintain both the main task of the FL system and the backdoor task simultaneously, causing malicious models to appear statistically similar to benign ones, which enables them to evade detection by existing defense methods. We find that malicious parameters in backdoored models are inactive on the main task, resulting in a significantly large empirical loss during the machine unlearning process on clean inputs. Inspired by this, we propose MASA, a method that utilizes individual unlearning on local models to identify malicious models in FL. To improve the performance of MASA in challenging non-independent and identically distributed (non-IID) settings, we design pre-unlearning model fusion that integrates local models with knowledge learned from other datasets to mitigate the divergence in their unlearning behaviors caused by the non-IID data distributions of clients. Additionally, we propose a new anomaly detection metric with minimal hyperparameters to filter out malicious models efficiently. Extensive experiments on IID and non-IID datasets across six different attacks validate the effectiveness of MASA. To the best of our knowledge, this is the first work to leverage machine unlearning to identify malicious models in FL. Code is available at \url{https://github.com/JiiahaoXU/MASA}.

Fed-piLot: Optimizing LoRA Assignment for Efficient Federated Foundation Model Fine-Tuning

Foundation models (FMs) have shown remarkable advancements in enhancing the performance of intelligent applications. To address the need for data privacy in FM fine-tuning, federated learning has emerged as the de facto framework. Specifically, Federated FMs (FedFMs) fine-tuning using low-rank adaptation (LoRA) modules instead of the full model over multiple clients can achieve both parameter efficiency and data privacy. However, recent studies rarely address the challenges posed by clients with heterogeneous resources, particularly in GPU memory capacity. In this paper, we introduce Fed-piLot, an efficient FedFM fine-tuning framework with optimized local LoRA assignments for heterogeneous clients. By emphasizing the different memory consumption for training different LoRA layers, as well as the varying contributions of different layers to model performance, we formulate the LoRA assignment as a Knapsack Optimization Problem. We design a Local-Global Information Gain Score (IG-Score) based value function to optimize LoRA assignment under clients' memory constraints. To further mitigate the impact of heterogeneity in model updates, we propose a novel Spatial-Temporal model aggregation (STAgg) rule using the Dynamic Weight Adjustment (DWA) strategy. Experimental results on three datasets under both IID and non-IID conditions demonstrate the effectiveness and efficiency of Fed-piLot. The code will be publicly available.

Achieving Byzantine-Resilient Federated Learning via Layer-Adaptive Sparsified Model Aggregation

Federated Learning (FL) enables multiple clients to collaboratively train a model without sharing their local data. Yet the FL system is vulnerable to well-designed Byzantine attacks, which aim to disrupt the model training process by uploading malicious model updates. Existing robust aggregation rule-based defense methods overlook the diversity of magnitude and direction across different layers of the model updates, resulting in limited robustness performance, particularly in non-IID settings. To address these challenges, we propose the Layer-Adaptive Sparsified Model Aggregation (LASA) approach, which combines pre-aggregation sparsification with layer-wise adaptive aggregation to improve robustness. Specifically, LASA includes a pre-aggregation sparsification module that sparsifies updates from each client before aggregation, reducing the impact of malicious parameters and minimizing the interference from less important parameters for the subsequent filtering process. Based on sparsified updates, a layer-wise adaptive filter then adaptively selects benign layers using both magnitude and direction metrics across all clients for aggregation. We provide the detailed theoretical robustness analysis of LASA and the resilience analysis for the FL integrated with LASA. Extensive experiments are conducted on various IID and non-IID datasets. The numerical results demonstrate the effectiveness of LASA. Code is available at \url{https://github.com/JiiahaoXU/LASA}.

Refuse Whenever You Feel Unsafe: Improving Safety in LLMs via Decoupled Refusal Training

This study addresses a critical gap in safety tuning practices for Large Language Models (LLMs) by identifying and tackling a refusal position bias within safety tuning data, which compromises the models' ability to appropriately refuse generating unsafe content. We introduce a novel approach, Decoupled Refusal Training (DeRTa), designed to empower LLMs to refuse compliance to harmful prompts at any response position, significantly enhancing their safety capabilities. DeRTa incorporates two novel components: (1) Maximum Likelihood Estimation (MLE) with Harmful Response Prefix, which trains models to recognize and avoid unsafe content by appending a segment of harmful response to the beginning of a safe response, and (2) Reinforced Transition Optimization (RTO), which equips models with the ability to transition from potential harm to safety refusal consistently throughout the harmful response sequence. Our empirical evaluation, conducted using LLaMA3 and Mistral model families across six attack scenarios, demonstrates that our method not only improves model safety without compromising performance but also surpasses well-known models such as GPT-4 in defending against attacks. Importantly, our approach successfully defends recent advanced attack methods (e.g., CodeAttack) that have jailbroken GPT-4 and LLaMA3-70B-Instruct. Our code and data can be found at https://github.com/RobustNLP/DeRTa.

Deep Learning for Detecting and Early Predicting Chronic Obstructive Pulmonary Disease from Spirogram Time Series: A UK Biobank Study

Chronic Obstructive Pulmonary Disease (COPD) is a chronic inflammatory lung condition that causes airflow obstruction. The existing methods can only detect patients who already have COPD based on obvious features shown in the spirogram (In this article, the spirogram specifically involves measuring Volume-Flow curve time series). Early prediction of COPD risk is vital for monitoring COPD disease progression, slowing it down, or even preventing its onset. However, these methods fail to early predict an individual's probability of COPD in the future based on subtle features in the spirogram. To address this gap, for the first time, we propose DeepSpiro, a method based on deep learning for early prediction of future COPD risk. DeepSpiro consists of four parts. First, we construct Volume-Flow curves guided by Time-Volume instability smoothing (SpiroSmoother) to enhance the stability of the original Volume-Flow curves precisely. Second, we extract critical features from the evolution of varied-length key patches (SpiroEncoder) to capture the key temporal evolution from original high-dimensional dynamic sequences to a unified low-dimensional temporal representation. Third, we explain the model based on temporal attention and heterogeneous feature fusion (SpiroExplainer), which integrates information from heterogeneous data such as spirogram and demographic information. Fourth, we predict the risk of COPD based on the evolution of key patch concavity (SpiroPredictor), enabling accurate prediction of the risk of disease in high-risk patients who are not yet diagnosed, for up to 1, 2, 3, 4, 5 years, and beyond. We conduct experiments on the UK Biobank dataset. Results show that DeepSpiro achieves an AUC value of 0.8328 in the task of detecting COPD. In early prediction tasks, high-risk and low-risk groups show significant differences in the future, with a p-value of <0.001.