Simeng Sun

Simeng Sun, Yang Liu, Dan Iter et al. · stanford

Fine-tuning large language models is becoming ever more impractical due to their rapidly-growing scale. This motivates the use of parameter-efficient adaptation methods such as prompt tuning (PT), which adds a small number of tunable embeddings to an otherwise frozen model, and in-context learning (ICL), in which demonstrations of the task are provided to the model in natural language without any additional training. Recently, Singhal et al. (2022) propose ``instruction prompt tuning'' (IPT), which combines PT with ICL by concatenating a natural language demonstration with learned prompt embeddings. While all of these methods have proven effective on different tasks, how they interact with each other remains unexplored. In this paper, we empirically study when and how in-context examples improve prompt tuning by measuring the effectiveness of ICL, PT, and IPT on five text generation tasks with multiple base language models. We observe that (1) IPT does \emph{not} always outperform PT, and in fact requires the in-context demonstration to be semantically similar to the test input to yield improvements; (2) PT is unstable and exhibits high variance, but combining PT and ICL (into IPT) consistently reduces variance across all five tasks; and (3) prompts learned for a specific source task via PT exhibit positive transfer when paired with in-context examples of a different target task. Our results offer actionable insights on choosing a suitable parameter-efficient adaptation method for a given task.

Simeng Sun, Dhawal Gupta, Mohit Iyyer

During the last stage of RLHF, a large language model is aligned to human intents via PPO training, a process that generally requires large-scale computational resources. In this technical report, we empirically investigate an efficient implementation of RLHF using low-rank adaptation (LoRA), which allows us to align the LLaMA 7B checkpoint on the Alpaca dataset using only two A100 GPUs instead of the eight required for full model fine-tuning. Despite tuning only 0.2% of LLaMA 7B's parameters, our implementation achieves better performance than the publicly-released AlpacaFarm checkpoint with full model fine-tuning. Next, we analyze several configurations of our LoRA-based PPO implementation, varying the form of the KL regularization term in the training objective. We find that (1) removing this penalty term does not harm performance on the AlpacaFarm evaluation set under our LoRA setup; (2) other regularizers, such as Jensen-Shannon divergence, lead to improved performance; and (3) while PPO training negatively impacts the factuality of model-generated responses, training with LoRA largely mitigates this effect. We release our code and pretrained checkpoints to facilitate future research on more efficient RLHF.

Simeng Sun, Katherine Thai, Mohit Iyyer

While numerous architectures for long-range language models (LRLMs) have recently been proposed, a meaningful evaluation of their discourse-level language understanding capabilities has not yet followed. To this end, we introduce ChapterBreak, a challenge dataset that provides an LRLM with a long segment from a narrative that ends at a chapter boundary and asks it to distinguish the beginning of the ground-truth next chapter from a set of negative segments sampled from the same narrative. A fine-grained human annotation reveals that our dataset contains many complex types of chapter transitions (e.g., parallel narratives, cliffhanger endings) that require processing global context to comprehend. Experiments on ChapterBreak show that existing LRLMs fail to effectively leverage long-range context, substantially underperforming a segment-level model trained directly for this task. We publicly release our ChapterBreak dataset to spur more principled future research into LRLMs.

Cheng-Ping Hsieh, Simeng Sun, Samuel Kriman et al. · nvidia

The needle-in-a-haystack (NIAH) test, which examines the ability to retrieve a piece of information (the "needle") from long distractor texts (the "haystack"), has been widely adopted to evaluate long-context language models (LMs). However, this simple retrieval-based test is indicative of only a superficial form of long-context understanding. To provide a more comprehensive evaluation of long-context LMs, we create a new synthetic benchmark RULER with flexible configurations for customized sequence length and task complexity. RULER expands upon the vanilla NIAH test to encompass variations with diverse types and quantities of needles. Moreover, RULER introduces new task categories multi-hop tracing and aggregation to test behaviors beyond searching from context. We evaluate 17 long-context LMs with 13 representative tasks in RULER. Despite achieving nearly perfect accuracy in the vanilla NIAH test, almost all models exhibit large performance drops as the context length increases. While these models all claim context sizes of 32K tokens or greater, only half of them can maintain satisfactory performance at the length of 32K. Our analysis of Yi-34B, which supports context length of 200K, reveals large room for improvement as we increase input length and task complexity. We open source RULER to spur comprehensive evaluation of long-context LMs.

Simeng Sun, Maha Elbayad, Anna Sun et al.

With multilingual machine translation (MMT) models continuing to grow in size and number of supported languages, it is natural to reuse and upgrade existing models to save computation as data becomes available in more languages. However, adding new languages requires updating the vocabulary, which complicates the reuse of embeddings. The question of how to reuse existing models while also making architectural changes to provide capacity for both old and new languages has also not been closely studied. In this work, we introduce three techniques that help speed up effective learning of the new languages and alleviate catastrophic forgetting despite vocabulary and architecture mismatches. Our results show that by (1) carefully initializing the network, (2) applying learning rate scaling, and (3) performing data up-sampling, it is possible to exceed the performance of a same-sized baseline model with 30% computation and recover the performance of a larger model trained from scratch with over 50% reduction in computation. Furthermore, our analysis reveals that the introduced techniques help learn the new directions more effectively and alleviate catastrophic forgetting at the same time. We hope our work will guide research into more efficient approaches to growing languages for these MMT models and ultimately maximize the reuse of existing models.

Krishna C. Puvvada, Faisal Ladhak, Santiago Akle Serrano et al. · nvidia

We present a decoder-only Transformer architecture that robustly generalizes to sequence lengths substantially longer than those seen during training. Our model, SWAN-GPT, interleaves layers without positional encodings (NoPE) and sliding-window attention layers equipped with rotary positional encodings (SWA-RoPE). Experiments demonstrate strong performance on sequence lengths significantly longer than the training length without the need for additional long-context training. This robust length extrapolation is achieved through our novel architecture, enhanced by a straightforward dynamic scaling of attention scores during inference. In addition, SWAN-GPT is more computationally efficient than standard GPT architectures, resulting in cheaper training and higher throughput. Further, we demonstrate that existing pre-trained decoder-only models can be efficiently converted to the SWAN architecture with minimal continued training, enabling longer contexts. Overall, our work presents an effective approach for scaling language models to longer contexts in a robust and efficient manner.

Simeng Sun, Cheng-Ping Hsieh, Faisal Ladhak et al.

Complex reasoning tasks often rely on the ability to consistently and accurately apply simple rules across incremental steps, a foundational capability which we term "level-0" reasoning. To systematically evaluate this capability, we introduce L0-Bench, a language model benchmark for testing procedural correctness -- the ability to generate correct reasoning processes, complementing existing benchmarks that primarily focus on outcome correctness. Given synthetic Python functions with simple operations, L0-Bench grades models on their ability to generate step-by-step, error-free execution traces. The synthetic nature of L0-Bench enables systematic and scalable generation of test programs along various axes (e.g., number of trace steps). We evaluate a diverse array of recent closed-source and open-weight models on a baseline test set. All models exhibit degradation as the number of target trace steps increases, while larger models and reasoning-enhanced models better maintain correctness over multiple steps. Additionally, we use L0-Bench to explore test-time scaling along three dimensions: input context length, number of solutions for majority voting, and inference steps. Our results suggest substantial room to improve "level-0" reasoning and potential directions to build more reliable reasoning systems.

Simeng Sun, Cheng-Ping Hsieh

We analyze contextual representations in neural autoregressive language models, emphasizing long-range contexts that span several thousand tokens. Our methodology employs a perturbation setup and the metric \emph{Anisotropy-Calibrated Cosine Similarity}, to capture the degree of contextualization of long-range patterns from the perspective of representation geometry. We begin the analysis with a case study on standard decoder-only Transformers, demonstrating that similar perplexity can exhibit markedly different downstream task performance, which can be explained by the difference in contextualization of long-range content. Next, we extend the analysis to other models, covering recent novel architectural designs and various training configurations. The representation-level results illustrate a reduced capacity for high-complexity (i.e., less compressible) sequences across architectures, and that fully recurrent models rely heavily on local context, whereas hybrid models more effectively encode the entire sequence structure. Finally, preliminary analysis of model size and training configurations on the encoding of long-range context suggest potential directions for improving existing language models.

Xin Jin, Ruoyu Feng, Simeng Sun et al.

Traditional media coding schemes typically encode image/video into a semantic-unknown binary stream, which fails to directly support downstream intelligent tasks at the bitstream level. Semantically Structured Image Coding (SSIC) framework makes the first attempt to enable decoding-free or partial-decoding image intelligent task analysis via a Semantically Structured Bitstream (SSB). However, the SSIC only considers image coding and its generated SSB only contains the static object information. In this paper, we extend the idea of semantically structured coding from video coding perspective and propose an advanced Semantically Structured Video Coding (SSVC) framework to support heterogeneous intelligent applications. Video signals contain more rich dynamic motion information and exist more redundancy due to the similarity between adjacent frames. Thus, we present a reformulation of semantically structured bitstream (SSB) in SSVC which contains both static object characteristics and dynamic motion clues. Specifically, we introduce optical flow to encode continuous motion information and reduce cross-frame redundancy via a predictive coding architecture, then the optical flow and residual information are reorganized into SSB, which enables the proposed SSVC could better adaptively support video-based downstream intelligent applications. Extensive experiments demonstrate that the proposed SSVC framework could directly support multiple intelligent tasks just depending on a partially decoded bitstream. This avoids the full bitstream decompression and thus significantly saves bitrate/bandwidth consumption for intelligent analytics. We verify this point on the tasks of image object detection, pose estimation, video action recognition, video object segmentation, etc.

Simeng Sun, Angela Fan, James Cross et al.

Recent work in multilingual machine translation (MMT) has focused on the potential of positive transfer between languages, particularly cases where higher-resourced languages can benefit lower-resourced ones. While training an MMT model, the supervision signals learned from one language pair can be transferred to the other via the tokens shared by multiple source languages. However, the transfer is inhibited when the token overlap among source languages is small, which manifests naturally when languages use different writing systems. In this paper, we tackle inhibited transfer by augmenting the training data with alternative signals that unify different writing systems, such as phonetic, romanized, and transliterated input. We test these signals on Indic and Turkic languages, two language families where the writing systems differ but languages still share common features. Our results indicate that a straightforward multi-source self-ensemble -- training a model on a mixture of various signals and ensembling the outputs of the same model fed with different signals during inference, outperforms strong ensemble baselines by 1.3 BLEU points on both language families. Further, we find that incorporating alternative inputs via self-ensemble can be particularly effective when training set is small, leading to +5 BLEU when only 5% of the total training data is accessible. Finally, our analysis demonstrates that including alternative signals yields more consistency and translates named entities more accurately, which is crucial for increased factuality of automated systems.

Simeng Sun, Wenlong Zhao, Varun Manjunatha et al.

While large-scale pretrained language models have significantly improved writing assistance functionalities such as autocomplete, more complex and controllable writing assistants have yet to be explored. We leverage advances in language modeling to build an interactive writing assistant that generates and rephrases text according to fine-grained author specifications. Users provide input to our Intent-Guided Assistant (IGA) in the form of text interspersed with tags that correspond to specific rhetorical directives (e.g., adding description or contrast, or rephrasing a particular sentence). We fine-tune a language model on a dataset heuristically-labeled with author intent, which allows IGA to fill in these tags with generated text that users can subsequently edit to their liking. A series of automatic and crowdsourced evaluations confirm the quality of IGA's generated outputs, while a small-scale user study demonstrates author preference for IGA over baseline methods in a creative writing task. We release our dataset, code, and demo to spur further research into AI-assisted writing.

Simeng Sun, Tao Yu, Jiahua Xu et al.

A good distortion representation is crucial for the success of deep blind image quality assessment (BIQA). However, most previous methods do not effectively model the relationship between distortions or the distribution of samples with the same distortion type but different distortion levels. In this work, we start from the analysis of the relationship between perceptual image quality and distortion-related factors, such as distortion types and levels. Then, we propose a Distortion Graph Representation (DGR) learning framework for IQA, named GraphIQA, in which each distortion is represented as a graph, i.e., DGR. One can distinguish distortion types by learning the contrast relationship between these different DGRs, and infer the ranking distribution of samples from different levels in a DGR. Specifically, we develop two sub-networks to learn the DGRs: a) Type Discrimination Network (TDN) that aims to embed DGR into a compact code for better discriminating distortion types and learning the relationship between types; b) Fuzzy Prediction Network (FPN) that aims to extract the distributional characteristics of the samples in a DGR and predicts fuzzy degrees based on a Gaussian prior. Experiments show that our GraphIQA achieves the state-of-the-art performance on many benchmark datasets of both synthetic and authentic distortions.