ICLR 2026 Papers

Leveraging the intrinsic importance differences among experts, recent research has explored expert-level compression techniques to enhance the efficiency of Mixture-of-Experts (MoE) large language models (LLMs). However, existing approaches often rely on empirical heuristics to identify critical experts, while lacking a deeper understanding into the heterogeneous importance of experts and the inner workings of MoE LLMs. In this study, we report, for the first time, the discovery and systematic investigation of a distinct subset of experts that play a pivotal role in the model's forward inference. These experts are prevalent in open-source MoE LLMs, and despite their extremely limited number, pruning them results in a substantial decline in model performance (e.g., prune just three out of 6,144 causes Qwen3-30B-A3B to generate repetitive and uninformative outputs). We refer to these experts as Super Experts (SEs). Our comprehensive analysis provides progressively deeper insights into SEs: (i) SEs are characterized by rare but extreme activation outliers in the output of the down\_proj, which give rise to massive activations in the hidden states between decoder layers. Moreover, the distribution of SEs is model-specific, data-agnostic, and remains unaffected by post-training processes. (ii) By pruning SEs, we assess their significance across a variety of tasks, revealing their considerable impact on the model's overall performance, particularly in mathematical reasoning. (iii) We further investigate why compressing SEs exerts such a pronounced impact. We show that, in MoE LLMs, SEs serve as the primary source of the systematic outlier mechanism in Transformers, and that compressing them profoundly disrupts this process, ultimately causing the collapse of attention sinks. These findings advance the understanding of the internal dynamics of MoE LLMs, filling an important gap in the current knowledge. In addition, we developed an automated tool for rapid and accurate SE profiling. The code is provided in the supplementary materials.

Compressing neural networks without retraining is vital for deployment at scale. We study calibration-free compression through the lens of projection geometry: structured pruning is an axis-aligned projection, whereas model folding performs a low-rank projection via weight clustering. We formalize both as orthogonal operators and show that, within a rank distance of one, folding provably yields smaller parameter reconstruction error, and under mild smoothness assumptions, smaller functional perturbations than pruning. At scale, we evaluate >1'000 checkpoints spanning ResNet18, PreActResNet18, ViT-B/32, and CLIP ViT-B/32 on CIFAR-10 and ImageNet-1K, covering diverse training hyperparameters (optimizers, learning rates, augmentations, regularization, sharpness-aware training). We show that folding typically achieves higher post-compression accuracy, with the largest gains at moderate–high compression. The gap narrows and occasionally reverses at specific training setups. Our results position folding as a geometry-aware, calibration-free alternative to pruning that is often superior in practice and principled in theory.

We prove that a minimal Transformer with frozen weights emulates a broad class of algorithms by in-context prompting. We formalize two modes of in-context algorithm emulation. In the *task-specific mode*, for any continuous function $f: \mathbb{R} \to \mathbb{R}$, we show the existence of a single-head softmax attention layer whose forward pass reproduces reproduces functions of the form $f(w^\top x - y)$ to arbitrary precision. This general template subsumes many popular machine learning algorithms (e.g., gradient descent, linear regression, ridge regression). In the *prompt-programmable mode*, we prove universality: a single fixed-weight two-layer softmax attention module emulates all algorithms from the task-specific class (i.e., each implementable by a single softmax attention) via only prompting. Our key idea is to construct prompts that encode an algorithm’s parameters into token representations, creating sharp dot-product gaps that force the softmax attention to follow the intended computation. This construction requires no feed-forward layers and no parameter updates. All adaptation happens through the prompt alone. Numerical results corroborate our theory. These findings forge a direct link between in-context learning and algorithmic emulation, and offer a simple mechanism for large Transformers to serve as prompt-programmable interpreters of algorithms. They illuminate how GPT-style foundation models may swap algorithms via prompts alone, and establish a form of algorithmic universality in modern Transformer models.

Context compression presents a promising approach for accelerating large language model (LLM) inference by compressing long contexts into compact representations.Current context compression methods predominantly rely on autoencoding tasks to train context-agnostic compression tokens to compress contextual semantics.While autoencoding tasks enable compression tokens to acquire compression capabilities, compression via autoencoding tasks creates a fundamental mismatch: the models are optimized for reconstruction that diverge from actual downstream tasks, thereby weakening the features more beneficial for real-world usage.We propose Semantic-Anchor Compression (SAC), a novel method that shifts from autoencoding task based compression to an architecture that is equipped with this compression capability $\textit{a priori}$. Instead of training models to compress contexts through autoencoding tasks, SAC directly selects so-called anchor tokens from the original context and aggregates contextual information into their key-value (KV) representations. By deriving representations directly from the contextual tokens, SAC eliminates the need for autoencoding training. To ensure compression performance while directly leveraging anchor tokens directly, SAC incorporates two key designs: (1) anchor embeddings that enable the compressor to identify critical tokens, and (2) bidirectional attention modification that allows anchor tokens to capture information from the entire context.Experimental results demonstrate that SAC consistently outperforms existing context compression methods across various compression ratios. On out-of-distribution evaluation using MRQA, SAC achieves 1 EM improvement at 5x compression over strong baselines,with increasing advantages at higher compression ratios.

Large language models (LLMs) present significant deployment challenges due to their immense computational and memory requirements. While semi-structured pruning, particularly 2:4 sparsity, offers a path to practical hardware acceleration, existing methods often incur substantial performance degradation. To bridge this gap, we introduce ARMOR: (Adaptive Representation with Matrix- factORization), a novel one-shot post-training pruning algorithm. Instead of directly pruning weights, ARMOR factorizes each weight matrix into a 2:4 sparse core wrapped by two low-overhead, block diagonal matrices. These wrappers act as efficient pre- and post-transformation error correctors, offering greater flexibility to preserve model quality compared to conventional 2:4 pruning techniques. The sparse core and block diagonal wrappers are chosen through a block coordinate descent algorithm that minimizes a layer-wise proxy loss. We prove this optimization is guaranteed to converge to a solution with a proxy loss less than or equal to state-of-the-art pruning algorithms. Experiments on Llama (Touvron et al., 2023; Dubey et al., 2024) and Qwen (Yang et al., 2025) model families demonstrate that ARMOR consistently and significantly outperforms state-of-the-art 2:4 pruning methods across a wide range of downstream tasks and perplexity evaluations, and generalizes to provide improvements for general N:M patterns and unstructured sparsity. ARMOR achieves this superior performance while retaining the inference speedups and substantial memory usage reductions of 2:4 pruning, establishing a more effective trade-off between model compression and task accuracy.

The success of Transformers lies in their ability to improve inference through two complementary strategies: the permanent refinement of model parameters via _in-weight learning_ (IWL), and the ephemeral modulation of inferences via _in-context learning_ (ICL), which leverages contextual information maintained in the model's activations. Evolutionary biology tells us that the predictability of the environment across timescales predicts the extent to which analogous strategies should be preferred. Genetic _evolution_ adapts to stable environmental features by gradually modifying the genotype over generations. Conversely, environmental volatility favors _plasticity_, which enables a single genotype to express different traits within a lifetime, provided there are reliable cues to guide the adaptation. We operationalize these dimensions (environmental stability and cue reliability) in controlled task settings (sinusoid regression and Omniglot classification) to characterize their influence on learning in Transformers. We find that stable environments favor IWL, often exhibiting a sharp transition when conditions are static. Conversely, reliable cues favor ICL, particularly when the environment is volatile. Furthermore, an analysis of learning dynamics reveals task-dependent transitions between strategies (ICL $\to$ IWL and vice versa). We demonstrate that these transitions are governed by (1) the asymptotic optimality of the strategy with respect to the environment, and (2) the optimization cost of acquiring that strategy, which depends on the task structure and the learner's inductive bias.

Optimizing inference for long-context large language models (LLMs) is increasingly important due to the quadratic compute and linear memory cost of Transformers. Existing approximate inference methods, including key-value (KV) cache dropping, sparse attention, and prompt compression, typically rely on coarse predictions of token or KV pair importance. We unify and extend recent work by introducing a framework for approximate LLM inference that leverages small draft models to more accurately predict token and KV pair importance. We provide novel theoretical and empirical analyses justifying lookahead-based importance estimation techniques. Within this framework, we present: (i) **SpecKV**, the first method to use lookahead with a small draft model to enable precise KV cache dropping; (ii) **SpecPC**, which leverages draft model attention activations to identify and discard less important prompt tokens; and (iii) **SpecKV-PC**, a cascaded compression strategy combining both techniques. Extensive experiments on long-context benchmarks demonstrate that our methods consistently achieve higher accuracy than existing baselines while retaining the same efficiency gains in memory usage, latency, and throughput.