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.
Recent efforts to combine low-rank adaptation (LoRA) with mixture-of-experts (MoE) for adapting large language models (LLMs) to multiple tasks still exhibit prevailing limitations: they either swap entire attention/feed-forward layers for switch experts or bolt on parallel expert branches, diluting parameter efficiency and task fidelity. We propose the LoRA-Mixer, a modular and lightweight MoE framework that integrates LoRA experts. Our core innovation lies in replacing the projection matrices of the attention module's input/output linear layers with dynamically routed, task-specific LoRA experts. This design ensures seamless compatibility with diverse foundation models, including transformers and state space models (SSMs), by leveraging their inherent linear projection structures. The framework supports two operational paradigms: (1) joint optimization of LoRA experts and routing mechanisms via a novel hard-soft routing strategy, or (2) direct deployment of pre-trained, frozen LoRA modules sourced from external repositories. To enable robust router training with limited data while ensuring stable routing decisions and maximizing expert reuse, we introduce an adaptive Specialization Balance Loss (SBL) that jointly optimizes expert balance and task-specific alignment. Extensive experiments on seven benchmark datasets, including MedQA, CoLA, SST-2, GSM8K, ARC-E, ARC-C, and HumanEval, demonstrate the effectiveness of LoRA-Mixer. On datasets such as GSM8K, HumanEval, and MedQA, LoRA-Mixer achieves significant improvements of 7.61%, 4.88%, and 3.08% over the base models, respectively. Compared with state-of-the-art methods, LoRA-Mixer achieves additional improvements of 1.09%, 1.45%, and 1.68%, respectively, using only 48% of the parameters, demonstrating its efficiency and strong performance.
Multilingual Large Language Models (LLMs) develop cross-lingual abilities despite being trained on limited parallel data. However, they often struggle to generate responses in the intended language, favoring high-resource languages such as English. In this work, we introduce CoCo-CoLa (Correct Concept - Correct Language), a novel metric to evaluate language adherence in multilingual LLMs. Using fine-tuning experiments on a closed-book QA task across seven languages, we analyze how training in one language affects others' performance. Our findings reveal that multilingual models share task knowledge across languages but exhibit biases in the selection of output language. We identify language-specific layers, showing that final layers play a crucial role in determining output language. Accordingly, we propose a partial training strategy that selectively fine-tunes key layers, improving language adherence while significantly reducing computational cost. Our method achieves comparable or superior performance to full fine-tuning, particularly for low-resource languages, offering a more efficient multilingual adaptation.
The scaling law of Large Language Models (LLMs) reveals a power-law relationship, showing diminishing return on performance as model scale increases. While training LLMs from scratch is resource-intensive, fine-tuning a pre-trained model for specific tasks has become a practical alternative. Full fine-tuning (FFT) achieves strong performance; however, it is computationally expensive and inefficient. Parameter-efficient fine-tuning (PEFT) methods, like LoRA, have been proposed to address these challenges by freezing the pre-trained model and adding lightweight task-specific modules. LoRA, in particular, has proven effective, but its application to multi-task scenarios is limited by interference between tasks. Recent approaches, such as Mixture-of-Experts (MOE) and asymmetric LoRA, have aimed to mitigate these issues but still struggle with sample scarcity and noise interference due to their fixed structure. In response, we propose CoLA, a more flexible LoRA architecture with an efficient initialization scheme, and introduces three collaborative strategies to enhance performance by better utilizing the quantitative relationships between matrices $A$ and $B$. Our experiments demonstrate the effectiveness and robustness of CoLA, outperforming existing PEFT methods, especially in low-sample scenarios. Our data and code are fully publicly available at https://github.com/zyy-2001/CoLA.
LLMs' remarkable ability to tackle a wide range of language tasks opened new opportunities for collaborative human-AI problem solving. LLMs can amplify human capabilities by applying their intuitions and reasoning strategies at scale. We explore whether human guides can be simulated, by generalizing from human demonstrations of guiding an AI system to solve complex language problems. We introduce CoLa, a novel self-guided learning paradigm for training automated $\textit{guides}$ and evaluate it on two QA datasets, a puzzle-solving task, and a constrained text generation task. Our empirical results show that CoLa consistently outperforms competitive approaches across all domains. Moreover, a small-sized trained guide outperforms a strong model like GPT-4 when acting as a guide. We compare the strategies employed by humans and automated guides by conducting a human study on a QA dataset. We show that automated guides outperform humans by adapting their strategies to reasoners' capabilities and conduct qualitative analyses highlighting distinct differences in guiding strategies.
Large Language Models (LLMs) are highly vulnerable to input perturbations, as even a small prompt change may result in a substantially different output. Existing methods to enhance LLM robustness are primarily focused on perturbed data samples, whereas improving resiliency to perturbations of task-level instructions has remained relatively underexplored. In this work, we focus on character- and word-level edits of task-specific instructions, which substantially degrade downstream performance. We experiment with a variety of techniques to enhance the robustness of LLMs, including self-denoising and representation alignment, testing different models (Llama 3 and Flan-T5), datasets (CoLA, QNLI, SST-2) and instructions (both task-oriented and role-oriented). We find that, on average, self-denoising -- whether performed by a frozen LLM or a fine-tuned model -- achieves substantially higher performance gains than alternative strategies, including more complex baselines such as ensembling and supervised methods.
Large Language Models (LLMs) have significantly advanced Natural Language Processing (NLP), particularly in Natural Language Understanding (NLU) tasks. As we progress toward an agentic world where LLM-based agents autonomously handle specialized tasks, it becomes crucial for these models to adapt to new tasks without forgetting previously learned information - a challenge known as catastrophic forgetting. This study evaluates the continual fine-tuning of various open-source LLMs with different parameter sizes (specifically models under 10 billion parameters) on key NLU tasks from the GLUE benchmark, including SST-2, MRPC, CoLA, and MNLI. By employing prompt engineering and task-specific adjustments, we assess and compare the models' abilities to retain prior knowledge while learning new tasks. Our results indicate that models such as Phi-3.5-mini exhibit minimal forgetting while maintaining strong learning capabilities, making them well-suited for continual learning environments. Additionally, models like Orca-2-7b and Qwen2.5-7B demonstrate impressive learning abilities and overall performance after fine-tuning. This work contributes to understanding catastrophic forgetting in LLMs and highlights prompting engineering to optimize model performance for continual learning scenarios.
Adapting Large Language Models (LLMs) to downstream tasks using Reinforcement Learning (RL) has proven to be an effective approach. However, LLMs do not inherently define the structure of an agent for RL training, particularly in terms of defining the action space. This paper studies learning a compact latent action space to enhance the controllability and exploration of RL for LLMs. We propose Controlling Large Language Models with Latent Actions (CoLA), a framework that integrates a latent action space into pre-trained LLMs. We apply CoLA to the Llama-3.1-8B model. Our experiments demonstrate that, compared to RL with token-level actions, CoLA's latent action enables greater semantic diversity in text generation. For enhancing downstream tasks, we show that CoLA with RL achieves a score of 42.4 on the math500 benchmark, surpassing the baseline score of 38.2, and reaches 68.2 when augmented with a Monte Carlo Tree Search variant. Furthermore, CoLA with RL consistently improves performance on agent-based tasks without degrading the pre-trained LLM's capabilities, unlike the baseline. Finally, CoLA reduces computation time by half in tasks involving enhanced thinking prompts for LLMs by RL. These results highlight CoLA's potential to advance RL-based adaptation of LLMs for downstream applications.
The full-size MLPs and the projection layers in attention introduce tremendous model sizes of large language models (LLMs), imposing extremely demanding needs of computational resources in the pre-training stage. However, we empirically observe that the activations of pre-trained LLMs exhibit low-rank property. Motivated by such observations, we propose CoLA and its memory-efficient implementation, CoLA-M, to replace these full-size layers with compute-efficient auto-encoders that naturally enforce low-rank activations throughout training. This fundamental architectural change eliminates the activation redundancy and significantly boosts model capacity and training efficiency. Experiments on LLaMA models with 60 million to 7 billion parameters show that CoLA reduces the computing cost by $\bf 2\pmb{\times}$ and improves training throughput by $\bf 1.86\pmb{\times}$ while maintaining full-rank level performance. CoLA-M further squeezes memory cost without sacrificing throughput, offering a pre-training approach with collectively superior parameter, computing, and memory efficiency. The LLMs produced are also $\bf 2\pmb{\times}$ smaller, enabling faster inference with lower memory cost on resource-constrained platforms.
This study explores the fine-tuning (FT) of the Open Pre-trained Transformer (OPT-125M) for grammatical acceptability tasks using the CoLA dataset. By comparing Vanilla-Fine-Tuning (VFT), Pattern-Based-Fine-Tuning (PBFT), and Parameter-Efficient Fine-Tuning techniques (PEFT) like Low-Rank Adaptation (LoRA), we demonstrate significant improvements in computational efficiency while maintaining high accuracy. Our experiments reveal that while VFT achieves the highest accuracy (81.2%), LoRA enhancing FT by reducing memory usage and iteration time by more than 50%, and increases accuracy in PBFT case. Context Distillation (CD), though computationally efficient, underperformed with accuracy around 31%. Our findings contribute to democratizing access to large language models (LLM) by reducing computational barriers.
A fundamental open challenge in modern LLM scaling is the lack of understanding around emergent capabilities. In particular, language model pretraining loss is known to be highly predictable as a function of compute. However, downstream capabilities are far less predictable -- sometimes even exhibiting emergent jumps -- which makes it challenging to anticipate the capabilities of future models. In this work, we first pose the task of emergence prediction: given access to current LLMs that have random few-shot accuracy on a task, can we predict whether future models (GPT-N+1) will have non-trivial accuracy on that task? We then discover a simple insight for this problem: finetuning LLMs on a given task can shift the point in scaling at which emergence occurs towards less capable models. To operationalize this insight, we can finetune LLMs with varying amounts of data and fit a parametric function that predicts when emergence will occur (i.e., "emergence laws"). We validate this approach using four standard NLP benchmarks where large-scale open-source LLMs already demonstrate emergence (MMLU, GSM8K, CommonsenseQA, and CoLA). Using only small-scale LLMs, we find that, in some cases, we can accurately predict whether models trained with up to 4x more compute have emerged. Finally, we present a case study of two realistic uses for emergence prediction.
Compositional reasoning is a hallmark of human visual intelligence. Yet, despite the size of large vision-language models, they struggle to represent simple compositions by combining objects with their attributes. To measure this lack of compositional capability, we design Cola, a text-to-image retrieval benchmark to Compose Objects Localized with Attributes. To solve Cola, a model must retrieve images with the correct configuration of attributes and objects and avoid choosing a distractor image with the same objects and attributes but in the wrong configuration. Cola contains about 1.2k composed queries of 168 objects and 197 attributes on around 30K images. Our human evaluation finds that Cola is 83.33% accurate, similar to contemporary compositionality benchmarks. Using Cola as a testbed, we explore empirical modeling designs to adapt pre-trained vision-language models to reason compositionally. We explore 6 adaptation strategies on 2 seminal vision-language models, using compositionality-centric test benchmarks - Cola and CREPE. We find the optimal adaptation strategy is to train a multi-modal attention layer that jointly attends over the frozen pre-trained image and language features. Surprisingly, training multimodal layers on CLIP performs better than tuning a larger FLAVA model with already pre-trained multimodal layers. Furthermore, our adaptation strategy improves CLIP and FLAVA to comparable levels, suggesting that training multimodal layers using contrastive attribute-object data is key, as opposed to using them pre-trained. Lastly, we show that Cola is harder than a closely related contemporary benchmark, CREPE, since simpler fine-tuning strategies without multimodal layers suffice on CREPE but not on Cola. However, we still see a significant gap between our best adaptation and human accuracy, suggesting considerable room for further research.
Storing open-source fine-tuned models separately introduces redundancy and increases response times in applications utilizing multiple models. Delta-parameter pruning (DPP), particularly the random drop and rescale (DARE) method proposed by Yu et al., addresses this by pruning the majority of delta parameters--the differences between fine-tuned and pre-trained model weights--while typically maintaining minimal performance loss. However, DARE fails when either the pruning rate or the magnitude of the delta parameters is large. We highlight two key reasons for this failure: (1) an excessively large rescaling factor as pruning rates increase, and (2) high mean and variance in the delta parameters. To push DARE's limits, we introduce DAREx (DARE the eXtreme), which features two algorithmic improvements: (1) DAREx-q, a rescaling factor modification that significantly boosts performance at high pruning rates (e.g., >30 % on COLA and SST2 for encoder models, with even greater gains in decoder models), and (2) DAREx-L2, which combines DARE with AdamR, an in-training method that applies appropriate delta regularization before DPP. We also demonstrate that DAREx-q can be seamlessly combined with vanilla parameter-efficient fine-tuning techniques like LoRA and can facilitate structural DPP. Additionally, we revisit the application of importance-based pruning techniques within DPP, demonstrating that they outperform random-based methods when delta parameters are large. Through this comprehensive study, we develop a pipeline for selecting the most appropriate DPP method under various practical scenarios.
While adversarial training is an effective defense method against adversarial attacks, it notably increases the training cost. To this end, fast adversarial training (FAT) is presented for efficient training and has become a hot research topic. However, FAT suffers from catastrophic overfitting, which leads to a performance drop compared with multi-step adversarial training. However, the cause of catastrophic overfitting remains unclear and lacks exploration. In this paper, we present an example taxonomy in FAT, which identifies that catastrophic overfitting is caused by the imbalance between the inner and outer optimization in FAT. Furthermore, we investigated the impact of varying degrees of training loss, revealing a correlation between training loss and catastrophic overfitting. Based on these observations, we redesign the loss function in FAT with the proposed dynamic label relaxation to concentrate the loss range and reduce the impact of misclassified examples. Meanwhile, we introduce batch momentum initialization to enhance the diversity to prevent catastrophic overfitting in an efficient manner. Furthermore, we also propose Catastrophic Overfitting aware Loss Adaptation (COLA), which employs a separate training strategy for examples based on their loss degree. Our proposed method, named example taxonomy aware FAT (ETA), establishes an improved paradigm for FAT. Experiment results demonstrate our ETA achieves state-of-the-art performance. Comprehensive experiments on four standard datasets demonstrate the competitiveness of our proposed method.
Due to the high computation overhead of Vision Transformers (ViTs), In-memory Computing architectures are being researched towards energy-efficient deployment in edge-computing scenarios. Prior works have proposed efficient algorithm-hardware co-design and IMC-architectural improvements to improve the energy-efficiency of IMC-implemented ViTs. However, all prior works have neglected the overhead and co-depencence of attention blocks on the accuracy-energy-delay-area of IMC-implemented ViTs. To this end, we propose TReX- an attention-reuse-driven ViT optimization framework that effectively performs attention reuse in ViT models to achieve optimal accuracy-energy-delay-area tradeoffs. TReX optimally chooses the transformer encoders for attention reuse to achieve near iso-accuracy performance while meeting the user-specified delay requirement. Based on our analysis on the Imagenet-1k dataset, we find that TReX achieves 2.3x (2.19x) EDAP reduction and 1.86x (1.79x) TOPS/mm2 improvement with ~1% accuracy drop in case of DeiT-S (LV-ViT-S) ViT models. Additionally, TReX achieves high accuracy at high EDAP reduction compared to state-of-the-art token pruning and weight sharing approaches. On NLP tasks such as CoLA, TReX leads to 2% higher non-ideal accuracy compared to baseline at 1.6x lower EDAP.
Recent advancements have introduced multiple vision-language models (VLMs) demonstrating impressive commonsense reasoning across various domains. Despite their individual capabilities, the potential of synergizing these complementary VLMs remains underexplored. The Cola Framework addresses this by showcasing how a large language model (LLM) can efficiently coordinate multiple VLMs through natural language communication, leveraging their distinct strengths. We have verified this claim on the challenging A-OKVQA dataset, confirming the effectiveness of such coordination. Building on this, our study investigates whether the same methodology can be applied to surveillance videos for action recognition. Specifically, we explore if leveraging the combined knowledge base of VLMs and LLM can effectively deduce actions from a video when presented with only a few selectively important frames and minimal temporal information. Our experiments demonstrate that LLM, when coordinating different VLMs, can successfully recognize patterns and deduce actions in various scenarios despite the weak temporal signals. However, our findings suggest that to enhance this approach as a viable alternative solution, integrating a stronger temporal signal and exposing the models to slightly more frames would be beneficial.
Instruction-tuned large language models have demonstrated remarkable capabilities in following human instructions across various domains. However, their proficiency remains notably deficient in many low-resource languages. To address this challenge, we begin by introducing FarsInstruct a comprehensive instruction dataset designed to enhance the instruction following ability of large language models specifically for the Persian language a significant yet underrepresented language globally. FarsInstruct encompasses a wide range of task types and datasets, each containing a mix of straightforward to complex manual written instructions, as well as translations from the Public Pool of Prompts, ensuring a rich linguistic and cultural representation. Furthermore, we introduce Co-CoLA, a framework designed to enhance the multi-task adaptability of LoRA-tuned models. Through extensive experimental analyses, our study showcases the effectiveness of the FarsInstruct dataset coupled with training by the Co-CoLA framework, in improving the performance of large language models within the Persian context. As of the current writing, FarsInstruct comprises 197 templates across 21 distinct datasets, and we intend to update it consistently, thus augmenting its applicability.
NLP(natural language processsing) has achieved great success through the transformer model.However, the model has hundreds of millions or billions parameters,which is huge burden for its deployment on personal computer or small scale of server.To deal with it, we either make the model's weight matrix relatively sparser, or compress attention layer. Pattern pruning ,one of the most important pruning methods, permits selecting fixed number of parameters in each divided pattern block and prunes it. However, the effect of pattern pruning is strictly limited by the sparsity within a region of weights in each layer. In this paper,we first introduced Alternating Direction Method of Multipliers(ADMM) based pattern pruning framework to reshape the distribution of activation map. Specifically, we propose to formulate the pattern pruning on transformer as a constrained optimization and use ADMM to optimize the problem. In this way, the initial dense feature maps is transformed to rather regionally sparsified ones.Therefore, we can then achieve higher compression ratio with better performance based on pattern pruning method. Additionally, this paper provides a theoretical derivations of the ADMM with local sparsity. Finally, we also extend the proposed ADMM based framework with SR-STE to demonstrate its generalization and to avoid gradient vanishing problem. We conduct extensive experiments on classification tasks over GLUE datasets. Significantly, we achieve 50% percent compression ratio while maintaining overall score 80.1 on GLUE dataset.
The depth/thermal information is beneficial for detecting salient object with conventional RGB images. However, in dual-modal salient object detection (SOD) model, the robustness against noisy inputs and modality missing is crucial but rarely studied. To tackle this problem, we introduce \textbf{Co}nditional Dropout and \textbf{LA}nguage-driven(\textbf{CoLA}) framework comprising two core components. 1) Language-driven Quality Assessment (LQA): Leveraging a pretrained vision-language model with a prompt learner, the LQA recalibrates image contributions without requiring additional quality annotations. This approach effectively mitigates the impact of noisy inputs. 2) Conditional Dropout (CD): A learning method to strengthen the model's adaptability in scenarios with missing modalities, while preserving its performance under complete modalities. The CD serves as a plug-in training scheme that treats modality-missing as conditions, strengthening the overall robustness of various dual-modal SOD models. Extensive experiments demonstrate that the proposed method outperforms state-of-the-art dual-modal SOD models, under both modality-complete and modality-missing conditions. We will release source code upon acceptance.
Recent advancements in generative AI have introduced novel prospects and practical implementations. Especially diffusion models show their strength in generating diverse and, at the same time, realistic features, positioning them well for generating counterfactual explanations for computer vision models. Answering "what if" questions of what needs to change to make an image classifier change its prediction, counterfactual explanations align well with human understanding and consequently help in making model behavior more comprehensible. Current methods succeed in generating authentic counterfactuals, but lack transparency as feature changes are not directly perceivable. To address this limitation, we introduce Concept-guided Latent Diffusion Counterfactual Explanations (CoLa-DCE). CoLa-DCE generates concept-guided counterfactuals for any classifier with a high degree of control regarding concept selection and spatial conditioning. The counterfactuals comprise an increased granularity through minimal feature changes. The reference feature visualization ensures better comprehensibility, while the feature localization provides increased transparency of "where" changed "what". We demonstrate the advantages of our approach in minimality and comprehensibility across multiple image classification models and datasets and provide insights into how our CoLa-DCE explanations help comprehend model errors like misclassification cases.