ICLR 2025 Papers

This work is about estimating when a conditional generative model (CGM) can solve an in-context learning (ICL) problem. An in-context learning (ICL) problem comprises a CGM, a dataset, and a prediction task. The CGM could be a multi-modal foundation model; the dataset, a collection of patient histories, test results, and recorded diagnoses; and the prediction task to communicate a diagnosis to a new patient. A Bayesian interpretation of ICL assumes that the CGM computes a posterior predictive distribution over an unknown Bayesian model defining a joint distribution over latent explanations and observable data. From this perspective, Bayesian model criticism is a reasonable approach to assess the suitability of a given CGM for an ICL problem. However, such approaches---like posterior predictive checks (PPCs)---often assume that we can sample from the likelihood and posterior defined by the Bayesian model, which are not explicitly given for contemporary CGMs. To address this, we show when ancestral sampling from the predictive distribution of a CGM is equivalent to sampling datasets from the posterior predictive of the assumed Bayesian model. Then we develop the generative predictive $p$-value, which enables PPCs and their cousins for contemporary CGMs. The generative predictive $p$-value can be used in a statistical decision procedure to determine when the model is appropriate for an ICL problem. Our method only requires generating queries and responses from a CGM and evaluating its response log probability. Using large language models, we empirically evaluate our method on tasks involving tabular data, imaging data, and natural language data.

Efficiently modeling sequences with infinite context length has long been a challenging problem. Previous approaches have either suffered from quadratic computational complexity or limited extrapolation ability in length generalization. In this work, we present Samba, a simple hybrid architecture that layer-wise combines Mamba, a selective State Space Model (SSM), with Sliding Window Attention (SWA). Samba selectively compresses a given sequence into recurrent hidden states while still maintaining the ability to precisely recall recent memories with the attention mechanism. We scale Samba up to 3.8B parameters with 3.2T training tokens and demonstrate that it significantly outperforms state-of-the-art models across a variety of benchmarks. Pretrained on sequences of 4K length, Samba shows improved perplexity in context lengths of up to 1M in zero-shot. When finetuned on 4K-length sequences, Samba efficiently extrapolates to a 256K context length with perfect memory recall on the Passkey Retrieval task, and exhibits superior retrieval extrapolation on the challenging Phonebook task compared to full-attention models. As a linear-time sequence model, Samba achieves a 3.73× higher throughput compared to Transformers with grouped-query attention for user prompts of 128K length, and a 3.64× speedup when generating 64K tokens with unlimited streaming.

Extremely low-resource (XLR) languages lack substantial corpora for training NLP models, motivating the use of all available resources such as dictionaries and grammar books. Machine Translation from One Book (Tanzer et al., 2024) suggests that prompting long-context LLMs with one grammar book enables English–Kalamang translation, an XLR language unseen by LLMs—a noteworthy case of linguistics helping an NLP task. We investigate the source of this translation ability, finding almost all improvements stem from the book’s parallel examples rather than its grammatical explanations. We find similar results for Nepali and Guarani, seen low-resource languages, and we achieve performance comparable to an LLM with a grammar book by simply fine-tuning an encoder-decoder translation model. We then investigate where grammar books help by testing two linguistic tasks, grammaticality judgment and gloss prediction, and we explore what kind of grammatical knowledge helps by introducing a typological feature prompt that achieves leading results on these more relevant tasks. We thus emphasise the importance of task-appropriate data for XLR languages: parallel examples for translation, and grammatical data for linguistic tasks. As we find no evidence that long-context LLMs can make effective use of grammatical explanations for XLR translation, we conclude data collection for multilingual XLR tasks such as translation is best focused on parallel data over linguistic description.

Despite significant advancements in large language models (LLMs), the rapid and frequent integration of small-scale experiences, such as interactions with sur- rounding objects, remains a substantial challenge. Two critical factors in assimilating these experiences are (1) **Efficacy**: the ability to accurately remember recent events; (2) **Retention**: the capacity to recall long-past experiences. Current methods either embed experiences within model parameters using continual learning, model editing, or knowledge distillation techniques, which often struggle with rapid updates and complex interactions, or rely on external storage to achieve long-term retention, thereby increasing storage requirements. In this paper, we propose **SELF-PARAM** (Self-Updatable Large Language Models with Parameter Integration). SELF-PARAM requires no extra parameters while ensuring near-optimal efficacy and long-term retention. Our method employs a training objective that minimizes the Kullback-Leibler (KL) divergence between the predictions of an original model (with access to contextual information) and a target model (without such access). By generating diverse question-answer pairs related to the knowledge and minimizing the KL divergence across this dataset, we update the target model to internalize the knowledge seamlessly within its parameters. Evaluations on question-answering and conversational recommendation tasks demonstrate that SELF-PARAM significantly outperforms existing methods, even when accounting for non-zero storage requirements. This advancement paves the way for more efficient and scalable integration of experiences in large language models by embedding knowledge directly into model parameters.

Transformer models have demonstrated exceptional performance across a wide range of applications. Though forming the foundation of Transformer models, the dot-product attention does not scale well to long-context data since its time requirement grows quadratically with context length. In this work, we propose Radar, a training-free approach that accelerates inference by dynamically searching for the most important context tokens. For any pre-trained Transformer, Radar can reduce the decoding time complexity without training or heuristically evicting tokens. Moreover, we provide theoretical justification for our approach, demonstrating that Radar can reliably identify the most important tokens with high probability. We conduct extensive comparisons with the previous methods on a wide range of tasks. The results demonstrate that Radar achieves the state-of-the-art performance across different architectures with reduced time complexity, offering a practical solution for efficient long-context processing of Transformers. The code is publicly available at https://github.com/BorealisAI/radar-decoding.

In-Context Learning (ICL) has significantly expanded the general-purpose nature of large language models, allowing them to adapt to novel tasks using merely the inputted context. This has motivated a series of papers that analyze tractable synthetic domains and postulate precise mechanisms that may underlie ICL. However, the use of relatively distinct setups that often lack a sequence modeling nature to them makes it unclear how general the reported insights from such studies are. Motivated by this, we propose a synthetic sequence modeling task that involves learning to simulate a finite mixture of Markov chains. As we show, models trained on this task reproduce most well-known results on ICL, hence offering a unified setting for studying the concept. Building on this setup, we demonstrate we can explain a model’s behavior by decomposing it into four broad algorithms that combine a fuzzy retrieval vs. inference approach with either unigram or bigram statistics of the context. These algorithms engage in a competitive dynamics to dominate model behavior, with the precise experimental conditions dictating which algorithm ends up superseding others: e.g., we find merely varying context size or amount of training yields (at times sharp) transitions between which algorithm dictates the model behavior, revealing a mechanism that explains the transient nature of ICL. In this sense, we argue ICL is best thought of as a mixture of different algorithms, each with its own peculiarities, instead of a monolithic capability. This also implies that making general claims about ICL that hold universally across all settings may be infeasible.

Stein Variational Gradient Descent (SVGD) is a widely used sampling algorithm that has been successfully applied in several areas of Machine Learning. SVGD operates by iteratively moving a set of $n$ interacting particles (which represent the samples) to approximate the target distribution. Despite recent studies on the complexity of SVGD and its variants, their long-time asymptotic behavior (i.e., after numerous iterations $k$) is still not understood in the finite number of particles regime. We study the long-time asymptotic behavior of a noisy variant of SVGD. First, we establish that the limit set of noisy SVGD for large $k$ is well-defined. We then characterize this limit set, showing that it approaches the target distribution as $n$ increases. In particular, noisy SVGD avoids the variance collapse observed for SVGD. Our approach involves demonstrating that the trajectories of noisy SVGD closely resemble those described by a McKean-Vlasov process.

Randomization is a powerful tool that endows algorithms with remarkable properties. For instance, randomized algorithms excel in adversarial settings, often surpassing the worst-case performance of deterministic algorithms with large margins. Furthermore, their success probability can be amplified by simple strategies such as repetition and majority voting. In this paper, we enhance deep neural networks, in particular transformer models, with randomization. We demonstrate for the first time that randomized algorithms can be instilled in transformers through learning, in a purely data- and objective-driven manner. First, we analyze known adversarial objectives for which randomized algorithms offer a distinct advantage over deterministic ones. We then show that common optimization techniques, such as gradient descent or evolutionary strategies, can effectively learn transformer parameters that make use of the randomness provided to the model. To illustrate the broad applicability of randomization in empowering neural networks, we study three conceptual tasks: associative recall, graph coloring, and agents that explore grid worlds. In addition to demonstrating increased robustness against oblivious adversaries through learned randomization, our experiments reveal remarkable performance improvements due to the inherently random nature of the neural networks' computation and predictions.

We present MELODI, a novel memory architecture designed to efficiently process long documents using short context windows. The key principle behind MELODI is to represent short-term and long-term memory as a hierarchical compression scheme across both transformer layers and context windows. Specifically, the short-term memory is achieved through recurrent compression of context windows across multiple layers, ensuring smooth transitions between windows. In contrast, the long-term memory performs further compression within a single middle layer and aggregates information across context windows, effectively consolidating crucial information from the entire history. Compared to a strong baseline - the Memorizing Transformer employing dense attention over a large long-term memory (64K key-value pairs) - our method demonstrates superior performance on various long-context datasets while remarkably reducing the memory footprint by a factor of 8.

Training and serving long-context large language models (LLMs) incurs substantial overhead. To address this, two critical steps are often required: a pretrained LLM typically undergoes a separate stage for context length extension by training on long-context data, followed by architectural modifications to reduce the overhead of KV cache during serving. This paper argues that integrating length extension with a GPU-friendly KV cache reduction architecture not only reduces training overhead during length extension, but also achieves better long-context performance. This leads to our proposed LongGen, which finetunes a pretrained LLM into an efficient architecture during length extension. LongGen builds on three key insights: (1) Sparse attention patterns, such as window attention (attending to recent tokens), attention sink (initial ones), and blockwise sparse attention (strided token blocks) are well-suited for building efficient long-context models, primarily due to their GPU-friendly memory access patterns, enabling efficiency gains not just theoretically but in practice as well. (2) It is essential for the model to have direct access to all tokens. A hybrid architecture with 1/3 full attention layers and 2/3 efficient ones achieves a balanced trade-off between efficiency and long-context performance. (3) Lightweight training on 5B long-context data is sufficient to extend the hybrid model's context length from 4K to 128K. We evaluate LongGen on both Llama-2 7B and Llama-2 70B, demonstrating its effectiveness across different scales. During training with 128K-long contexts, LongGen achieves 1.55x training speedup and reduces wall-clock time by 36%, compared to a full-attention baseline. During inference, LongGen reduces KV cache memory by 62%, achieving 1.67x prefilling speedup and 1.41x decoding speedup. Compared to baselines that apply KV-cache reduction techniques to full-attention long-context LLMs, LongGen achieves substantially stronger performance not only on the Needle-in-a-Haystack retrieval task, but also on more challenging long-context reasoning tasks, including BABILong and RULER.

Predicting the biophysical and functional properties of proteins is essential for in silico protein design. Machine learning has emerged as a promising technique for such prediction tasks. However, the relative scarcity of in vitro annotations means that these models often have little, or no, specific data on the desired fitness prediction task. As a result of limited data, protein language models (PLMs) are typically trained on general protein sequence modeling tasks, and then fine-tuned, or applied zero-shot, to protein fitness prediction. When no task data is available, the models make strong assumptions about the correlation between the protein sequence likelihood and fitness scores. In contrast, we propose meta-learning over a distribution of standard fitness prediction tasks, and demonstrate positive transfer to unseen fitness prediction tasks. Our method, called Metalic (Meta-Learning In-Context), uses in-context learning and fine-tuning, when data is available, to adapt to new tasks. Crucially, fine-tuning enables considerable generalization, even though it is not accounted for during meta-training. Our fine-tuned models achieve strong results with 18 times fewer parameters than state-of-the-art models. Moreover, our method sets a new state-of-the-art in low-data settings on ProteinGym, an established fitness-prediction benchmark. Due to data scarcity, we believe meta-learning will play a pivotal role in advancing protein engineering.

Spiking neural networks (SNNs) are posited as a computationally efficient and biologically plausible alternative to conventional neural architectures, with their core computational framework primarily using the leaky integrate-and-fire (LIF) neuron model. However, the limited hidden state representation of LIF neurons, characterized by a scalar membrane potential, and sequential spike generation process, poses challenges for effectively developing scalable spiking models to address long-range dependencies in sequence learning tasks. In this study, we develop a scalable probabilistic spiking learning framework for long-range dependency tasks leveraging the fundamentals of state space models. Unlike LIF neurons that rely on the deterministic Heaviside function for a sequential process of spike generation, we introduce a SpikeSampler layer that samples spikes stochastically based on an SSM-based neuronal model while allowing parallel computations. To address non-differentiability of the spiking operation and enable effective training, we also propose a surrogate function tailored for the stochastic nature of the SpikeSampler layer. To enhance inter-neuron communication, we introduce the SpikeMixer block, which integrates spikes from neuron populations in each layer. This is followed by a ClampFuse layer, incorporating a residual connection to capture complex dependencies, enabling scalability of the model. Our models attain state-of-the-art performance among SNN models across diverse long-range dependency tasks, encompassing the Long Range Arena benchmark, permuted sequential MNIST, and the Speech Command dataset and demonstrate sparse spiking pattern highlighting its computational efficiency.

In-context learning (ICL) allows LLMs to learn from examples without changing their weights: this is a particularly promising capability for long-context LLMs that can potentially learn from many examples. Recently, Lin et al. (2024) proposed URIAL, a method using only three in-context examples to align base LLMs, achieving non-trivial instruction following performance. In this work, we show that, while effective, ICL alignment with URIAL still underperforms compared to instruction fine-tuning on established benchmarks such as MT-Bench and AlpacaEval 2.0 (LC), especially with more capable base LLMs. We then uncover the most relevant elements for successful in-context alignment, finding the crucial role of the decoding parameters. Based on these insights, we show that the approach of URIAL can indeed be improved by adding more, potentially carefully selected, high-quality demonstrations in context, getting closer to the performance of instruct models. Finally, we provide the first, to our knowledge, systematic comparison of ICL and instruction fine-tuning (IFT) for instruction following in the low data regime. Overall, our work advances the understanding of ICL as an alignment technique and its relationship to IFT.

Large Language Model's are instruction-finetuned to enhance their ability to follow user instructions and better comprehend input context. Still, they often struggle to follow the input context, especially when it contradicts model's parametric knowledge. This manifests as various failures, such as hallucinations where a model inserts outdated or unwarranted facts into its response. In this work, we observe an intriguing phenomenon: the context reliance of the model decreases as instruction finetuning progresses, $\textit{despite an initial expected increase}$. We call this phenomenon as the $\textbf{context-parametric inversion}$. This is surprising, as one would expect instruction tuning to improve the model's ability to follow input instructions. We observe this behavior on multiple general purpose instruction tuning datasets such as TULU, Alpaca and Ultrachat, across multiple model families like Llama, Mistral and Pythia. We perform various controlled studies to eliminate some simple hypothesis for this observed behavior and isolate what datapoints cause this counter-intuitive behavior. We then analyze the phenomenon theoretically, to explain why context reliance varies across the trajectory of finetuning. We tie the observed context-parametric inversion to the properties of the finetuning data, which provides us with some potential mitigation strategies that provide limited but insightful gains.

State-space models (SSMs) that utilize linear, time-invariant (LTI) systems are known for their effectiveness in learning long sequences. To achieve state-of-the-art performance, an SSM often needs a specifically designed initialization, and the training of state matrices is on a logarithmic scale with a very small learning rate. To understand these choices from a unified perspective, we view SSMs through the lens of Hankel operator theory. Building upon it, we develop a new parameterization scheme, called HOPE, for LTI systems that utilizes Markov parameters within Hankel operators. Our approach helps improve the initialization and training stability, leading to a more robust parameterization. We efficiently implement these innovations by nonuniformly sampling the transfer functions of LTI systems, and they require fewer parameters compared to canonical SSMs. When benchmarked against HiPPO-initialized models such as S4 and S4D, an SSM parameterized by Hankel operators demonstrates improved performance on Long-Range Arena (LRA) tasks. Moreover, our new parameterization endows the SSM with non-decaying memory within a fixed time window, which is empirically corroborated by a sequential CIFAR-10 task with padded noise.

We explore the capability of transformers to address endogeneity in in-context linear regression. Our main finding is that transformers inherently possess a mechanism to handle endogeneity effectively using instrumental variables (IV). First, we demonstrate that the transformer architecture can emulate a gradient-based bi-level optimization procedure that converges to the widely used two-stage least squares (2SLS) solution at an exponential rate. Next, we propose an in-context pretraining scheme and provide theoretical guarantees showing that the global minimizer of the pre-training loss achieves a small excess loss. Our extensive experiments validate these theoretical findings, showing that the trained transformer provides more robust and reliable in-context predictions and coefficient estimates than the 2SLS method, in the presence of endogeneity.

In modern large language models (LLMs), increasing the context length is crucial for improving comprehension and coherence in long-context, multi-modal, and retrieval-augmented language generation. While many recent transformer models attempt to extend their context length over a million tokens, they remain impractical due to the quadratic time and space complexities. Although recent works on linear and sparse attention mechanisms can achieve this goal, their real-world applicability is often limited by the need to re-train from scratch and significantly worse performance. In response, we propose a novel approach, Hierarchically Pruned Attention (HiP), which reduces the time complexity of the attention mechanism to $O(T \log T)$ and the space complexity to $O(T)$, where $T$ is the sequence length. We notice a pattern in the attention scores of pretrained LLMs where tokens close together tend to have similar scores, which we call "attention locality". Based on this observation, we utilize a novel tree-search-like algorithm that estimates the top-$k$ key tokens for a given query on the fly, which is mathematically guaranteed to have better performance than random attention pruning. In addition to improving the time complexity of the attention mechanism, we further optimize GPU memory usage by implementing KV cache offloading, which stores only $O(\log T)$ tokens on the GPU while maintaining similar decoding throughput. Experiments on benchmarks show that HiP, with its training-free nature, significantly reduces both prefill and decoding latencies, as well as memory usage, while maintaining high-quality generation with minimal degradation. HiP enables pretrained LLMs to scale up to millions of tokens on commodity GPUs, potentially unlocking long-context LLM applications previously deemed infeasible.

In the realm of large-scale language models, a significant challenge arises when extrapolating sequences beyond the maximum allowable length. This is because the model's position embedding mechanisms are limited to positions encountered during training, thus preventing effective representation of positions in longer sequences. We analyzed conventional position encoding methods for long contexts and found the following characteristics. (1) When the representation dimension is regarded as the time axis, Rotary Position Embedding (RoPE) can be interpreted as a restricted wavelet transform using Haar-like wavelets. However, because it uses only a fixed scale parameter, it does not fully exploit the advantages of wavelet transforms, which capture the fine movements of non-stationary signals using multiple scales (window sizes). This limitation could explain why RoPE performs poorly in extrapolation. (2) Previous research as well as our own analysis indicates that Attention with Linear Biases (ALiBi) functions similarly to windowed attention, using windows of varying sizes. However, it has limitations in capturing deep dependencies because it restricts the receptive field of the model. From these insights, we propose a new position representation method that captures multiple scales (i.e., window sizes) by leveraging wavelet transforms without limiting the model's attention field. Experimental results show that this new method improves the performance of the model in both short and long contexts. In particular, our method allows extrapolation of position information without limiting the model's attention field.

Large language models (LLMs) encounter computational challenges during long-sequence inference, especially in the attention pre-filling phase, where the complexity grows quadratically with the prompt length. Previous efforts to mitigate these challenges have relied on fixed sparse attention patterns or identifying sparse attention patterns based on limited cases. However, these methods lacked the flexibility to efficiently adapt to varying input demands. In this paper, we introduce FlexPrefill, a Flexible sparse Pre-filling mechanism that dynamically adjusts sparse attention patterns and computational budget in real-time to meet the specific requirements of each input and attention head. The flexibility of our method is demonstrated through two key innovations: 1) Query-Aware Sparse Pattern Determination: By measuring Jensen-Shannon divergence, this component adaptively switches between query-specific diverse attention patterns and predefined attention patterns. 2) Cumulative-Attention Based Index Selection: This component dynamically selects query-key indexes to be computed based on different attention patterns, ensuring the sum of attention scores meets a predefined threshold. FlexPrefill adaptively optimizes the sparse pattern and sparse ratio of each attention head based on the prompt, enhancing efficiency in long-sequence inference tasks. Experimental results show significant improvements in both speed and accuracy over prior methods, providing a more flexible and efficient solution for LLM inference.