ICLR 2026 Papers

Determinism is indispensable for reproducibility in large language model (LLM) training, yet it often exacts a steep performance cost. In widely used attention implementations such as FlashAttention-3, the deterministic backward pass can incur up to a 37.9% throughput reduction relative to its non‑deterministic counterpart, primarily because gradient accumulation operations must be serialized to guarantee numerical consistency. This performance loss stems from suboptimal scheduling of compute and gradient‑reduction phases, leading to significant hardware underutilization. To address this challenge, we formulate the backward pass of deterministic attention as a scheduling problem on a Directed Acyclic Graph (DAG) and derive schedules that minimize the critical path length. Building on this formulation, we present DASH (Deterministic Attention Scheduling for High-Throughput), which encapsulates two complementary scheduling strategies: (i) Descending Q‑Tile Iteration, a reversed query‑block traversal that shrinks pipeline stalls in causal attention, and (ii) Shift Scheduling, a theoretically optimal schedule within our DAG model that reduces pipeline stalls for both full and causal masks. Our empirical evaluations on NVIDIA H800 GPUs demonstrate that DASH narrows the performance gap of deterministic attention. The proposed strategies improve the throughput of the attention backward pass by up to 1.28$\times$ compared to the baseline, significantly advancing the efficiency of reproducible LLM training.

Serving large language models (LLMs) at scale necessitates efficient key-value (KV) cache management. KV caches can be reused across conversation turns via shared-prefix prompts that are common in iterative code editing and chat. However, stale caches consume scarce GPU memory, require offloading, or force recomputation. We present KVTC, a lightweight transform coder that compresses KV caches for compact on-GPU and off-GPU storage. Drawing on classical media compression, KVTC combines PCA-based feature decorrelation, adaptive quantization, and entropy coding. It requires only a brief initial calibration and leaves model parameters unchanged. By exploiting redundancies in KV caches, KVTC achieves up to 20x compression, while maintaining reasoning and long-context accuracy in Llama 3.1, Mistral-NeMo, and R1-Qwen 2.5. Across benchmarks including AIME25, LiveCodeBench, GSM8K, MMLU, Qasper, RULER, and MATH500, KVTC consistently outperforms inference-time baselines such as token eviction, quantization, and SVD-based methods, delivering substantially higher compression ratios. These results support KVTC as a practical building block for memory-efficient LLM serving with reusable KV caches.

Large Language Models (LLMs) exhibit impressive capabilities yet suffer from sensitivity to slight input context variations, hampering reliability. Conventional metrics like accuracy and perplexity fail to assess local prediction robustness, as normalized output probabilities can obscure the underlying resilience of an LLM's internal state to perturbations. We introduce the Token Constraint Bound , a novel metric that quantifies the maximum internal state perturbation an LLM can withstand before its dominant next-token prediction significantly changes. Intrinsically linked to output embedding space geometry, provides insights into the stability of the model's internal predictive commitment. Our experiments show correlates with effective prompt engineering and uncovers critical prediction instabilities missed by perplexity during in-context learning and text generation. offers a principled, complementary approach to analyze and potentially improve the contextual stability of LLM predictions.

While Large Language Models (LLMs) hold promise to become autonomous agents, they often explore suboptimally in sequential decision-making. Recent work has sought to enhance this capability via supervised fine-tuning (SFT) or reinforcement learning (RL), improving regret on the classic multi-armed bandit task. However, it remains unclear how these learning methods shape exploration strategies and how well they generalize. We investigate both paradigms by training LLMs with SFT on expert trajectories and RL with a range of tailored reward signals including a strategic, regret-shaped reward to reduce variance, and an algorithmic reward that enables oracle imitation. The resulting agents outperform pre-trained models and achieve performance comparable to Upper Confidence Bound (UCB) and Thompson Sampling, with robust generalization to 6$\times$ longer horizons and across bandit families. Behavioral analysis reveals that gains often stem from more sophisticated but greedier exploitation: RL/SFT agents are more prone to early catastrophic failure than pre-trained models, prematurely abandoning exploration. Furthermore, agents trained to imitate UCB learn to outperform their teacher by adopting more exploitative variants. Our findings clarify when each training paradigm is preferable and advocate tailored reward design and evaluation beyond average regret to promote robust exploratory behavior.

While scaling laws for Large Language Models (LLMs) traditionally focus on proxy metrics like pretraining loss, predicting downstream task performance has been considered unreliable. This paper challenges that view by proposing a direct framework to model the scaling of downstream accuracy from the training budget. We demonstrate that for a fixed token-to-parameter ratio, a simple two-parameter scaling law accurately describes this relationship. Our findings are validated by experiments on models with up to 17B parameters trained on up to 350B tokens, showing that downstream capabilities scaling can be described using a scaling law. Furthermore, we extend this framework to extrapolate and predict accuracy of target model with up to 6.7x larger training budget based on a set of smaller experiments. We will release a complete list of model losses and downstream evaluation results at various different scales to support reproducibility and encourage future research.

We present Contamination Detection via Context (CoDeC), a practical and accurate method to detect and quantify training data contamination in large language models. CoDeC distinguishes between data memorized during training and data outside the training distribution by measuring how in-context learning affects model performance. We find that in‑context examples typically boost confidence for unseen datasets but may reduce it when the dataset was part of training, due to disrupted memorization patterns. Experiments show that CoDeC produces interpretable contamination scores that clearly separate seen and unseen datasets, and reveals strong evidence of memorization in open-weight models with undisclosed training corpora. The method is simple, automated, and both model- and dataset-agnostic, making it easy to integrate with benchmark evaluations.

The proliferation of long-context large language models (LLMs) exposes a key bottleneck: the rapidly expanding key-value cache during decoding, which imposes heavy memory and latency costs. While recent approaches attempt to alleviate this by sharing a single set of crucial tokens across layers, such coarse-grained sharing undermines model performance by neglecting the functional diversity of attention heads. To address this, we propose LycheeDecode, an efficient decoding method centered on a fine-grained hybrid-head attention mechanism that employs a hardware-efficient top-$k$ selection strategy. Specifically, the novel HardKuma-based mechanism partitions attention heads into a small subset of retrieval heads that dynamically identify crucial tokens and a majority of sparse heads that reuse them for efficient computation. Through extensive experiments on leading models like Llama3 and Qwen3 across diverse benchmarks for long-context understanding (e.g., LongBench, RULER) and complex reasoning (e.g., AIME24, OlympiadBench), we demonstrate that LycheeDecode achieves generative quality comparable to, and at times surpassing even the full-attention baseline. Crucially, this is accomplished with up to a 2.7x speedup at a 128K context length. By preserving the functional diversity of attention heads, our fine-grained strategy overcomes the performance bottlenecks of existing methods, providing a powerful and validated pathway to both efficient and high-quality long-context LLM inference. The implementation code, kernels, and models will be publicly available.

Large language models (LLMs) sometimes memorize undesirable knowledge, which must be removed after deployment. Prior work on machine unlearning has focused largely on optimization methods that adjust parameters to enforce forgetting while preserving retention. However, these approaches assume that the forget and retain sets are readily available, which rarely holds in practice. Unlearning is typically triggered by an undesired generation at inference time, making the retrieval of relevant data the central challenge. We introduce the notion of \emph{data Pareto improvement} for LLM unlearning, which formalizes how retrieval can expand the achievable trade-off frontier between forgetting and retention. To realize this principle, we propose Randomized Antipodal Search on Linearized Influence Kernel (RASLIK), a retrieval algorithm that combines permutation–projection hashing with randomized antipodal search. RASLIK reduces selection variance, achieves sublinear complexity, and yields a double gain in both quality and efficiency. Across multiple models, datasets, and unlearning algorithms, RASLIK consistently outperforms deterministic baselines and even oracle sampling, establishing randomized search as a principled and scalable solution for data-centric unlearning.

High-quality visual representation is a long-standing pursuit in computer vision. In the context of multimodal LLMs (MLLMs), feeding higher-resolution images can produce more fine-grained visual tokens. However, it introduces additional computational and design complexity, due to multiple forward passes and post-processing of increased tokens. Before simply adopting a higher resolution, have we truly unlocked the model's full perception capability at a standard resolution? Therefore, we study an interesting problem: how to achieve fine visual perception under lower cost without larger images. We present SigLIP-HD in this work. The core is a highly simple fine-to-coarse supervision design. We enforce the coarse feature of a mid-resolution image to mimic the fine-grained feature of its high-resolution version. We build this framework on the advanced SigLIP 2 model. Our final model produces better visual tokens at exactly the same inference budget. It is validated on extensive MLLM benchmarks and consistently delivers stronger results than our baseline model, especially on OCR-related tasks.

The scaling of Generative AI (GenAI) models into the hundreds of billions of parameters makes low-precision computation indispensable for efficient deployment. We argue that the fundamental solution lies in developing low-precision \emph{floating-point} formats, which inherently provide numerical stability, memory savings, and hardware efficiency without dequantization overhead. In this paper, we present a theoretical and empirical study of an \emph{exponent concentration} phenomenon in GenAI weights: exponents consistently exhibit low entropy across architectures and modalities. We show that this arises naturally from $\alpha$-stable distributions induced by stochastic gradient descent, and we prove tight bounds on the entropy of exponents. Our analysis establishes a theoretical compression limit near FP4.67, which motivates the design of a practical FP8 format. Building on these insights, we propose \textbf{Exponent-Concentrated FP8 (ECF8)}, a lossless compression framework with entropy-aware encoding and GPU-optimized decoding. Experiments on LLMs and DiTs up to 671B parameters demonstrate up to 26.9\% memory savings and 177.1\% throughput acceleration, with perfectly lossless computations, i.e., no deviation in model outputs. Our results establish exponent concentration as a statistical law of trained models and open a principled path for lossless low-precision floating-point design in the FP8 era.

Contemporary coding-agent benchmarks applaud “first correct answer,” silently assuming infinite tokens, container minutes, and developer patience. In production, every LLM call, test re-run, and rollback incurs hard cost; agents that cannot budget these resources are dead on arrival. We close the gap with USACOArena, an ICPC-inspired arena where agents pay deterministic credits for every prompt, compilation, test, or rollback. A task becomes a cost–benefit negotiation under uncertainty: is a second sample worth 15\% of the remaining budget, or should the agent pivot to a cheaper heuristic? Real-time deduction exposes decision profiles hidden from static leaderboards: the tax of over-specialized generators, the ROI of early-exit heuristics, and the compound interest of lightweight scaffolding. Even identically seeded agents diverge in self-play, revealing a rich policy space where the same model oscillates between spendthrift submission sprees and parsimonious exploration. Released as a reproducible benchmark and zero-shot curriculum, USACOArena provides the traces, credit engine, and six state-of-the-art decision logs to catalyze research on coding agents that know when to stop.

Cybersecurity operations demand assistant LLMs that support diverse workflows without exposing sensitive data. Existing solutions either rely on proprietary APIs with privacy risks or on open models lacking domain adaptation. To bridge this gap, we curate 11.8B tokens of cybersecurity-focused continual pretraining data via large-scale web filtering and manual collection of high-quality resources, spanning 28.6K documents across frameworks, offensive techniques, and security tools. Building on this, we design an agentic augmentation pipeline that simulates expert workflows to generate 266K multi-turn cybersecurity samples for supervised fine-tuning. Combined with general open-source LLM data, these resources enable the training of RedSage, an open-source, locally deployable cybersecurity assistant with domain-aware pretraining and post-training. To rigorously evaluate the models, we introduce RedSage-Bench, a benchmark with 30K multiple-choice and 240 open-ended Q\&A items covering cybersecurity knowledge, skills, and tool expertise. RedSage is further evaluated on established cybersecurity benchmarks (e.g., CTI-Bench, CyberMetric, SECURE) and general LLM benchmarks to assess broader generalization. At the 8B scale, RedSage achieves consistently better results, surpassing the baseline models by up to +5.59 points on cybersecurity benchmarks and +5.05 points on Open LLM Leaderboard tasks. These findings demonstrate that domain-aware agentic augmentation and pre/post-training can not only enhance cybersecurity-specific expertise but also help to improve general reasoning and instruction-following. All models, datasets, and code will be released to advance reproducibility and open cybersecurity LLM research.

With the rapid progress of large language models (LLMs) trained on every available piece of data, it becomes increasingly challenging to reliably evaluate their intelligence due to potential data contamination and benchmark overfitting. To overcome these challenges, we investigate a new angle of benchmarking LLMs' intelligence by evaluating their capability in forecasting real-world future events, a paradigm we call "LLM-as-a-Prophet". Such forecasting tasks require combination of sophisticated capabilities while remaining free from data contamination or overfitting. To systematically evaluate such predictive intelligence of LLMs, we introduce $\texttt{Prophet Arena}$, a general evaluation benchmark that continuously collects live forecasting tasks and decomposes each task into distinct pipeline stages, supporting our controlled and large-scale experimentation. Our comprehensive evaluation reveals that many LLMs already exhibit impressive forecasting capabilities, reflected in, e.g., their small calibration errors, consistent prediction confidence and promising market returns. However, we also uncover key bottlenecks even in frontier models, such as inaccurate event recalls, misunderstanding of data sources and slower information aggregation compared to markets when resolution nears.

Model merging in weight space is an efficient way to reuse pretrained models, but existing methods typically assume matching architectures or sizes, making heterogeneous merges brittle or infeasible. We address this limitation by encoding model weights into a smooth latent space, enabling cross-architecture operations, and performing the merge in the latent space before decoding back to weights. This approach faces two major challenges. First, LLMs contain billions of parameters, which makes latent encoding computationally demanding. Second, using high compression ratios often hinders the encoder’s ability to generalize to unseen weights. We tackle these issues with a transformer-based variational autoencoder (VAE) trained in a two-stage compression curriculum with structured layer-aware chunking: the model first learns a high-capacity latent representation and then distills to a compact code, improving both stability and out-of-distribution generalization. To align heterogeneous models, we introduce a dimensionality-matching projection that allows interpolation between models of different sizes. Empirically, latent-space interpolation is consistently more robust than direct weight-space averaging and yields stronger downstream performance when merging models of different sizes. Together, these components provide a scalable, architecture-agnostic recipe for model merging.

Although large language models (LLMs) are becoming increasingly capable of solving challenging real-world tasks, accurately quantifying their uncertainty remains a critical open problem—one that limits their applicability in high-stakes domains. This challenge is further compounded by the closed-source, black-box nature of many state-of-the-art LLMs. Moreover, LLM-based systems can be highly sensitive to the prompts that bind them together, which often require significant manual tuning (i.e., prompt engineering). In this work, we address these challenges by viewing LLM-based systems through a Bayesian lens. We interpret prompts as textual parameters in a statistical model, allowing us to use a small training dataset to perform Bayesian inference over these prompts. This novel perspective enables principled uncertainty quantification over both the model’s textual parameters and its downstream predictions, while also incorporating prior beliefs about these parameters expressed in free-form text. To perform Bayesian inference— a difficult problem even for well-studied data modalities—we introduce Metropolis-Hastings through LLM Proposals (MHLP), a novel Markov chain Monte Carlo (MCMC) algorithm that combines prompt optimization techniques with standard MCMC methods. MHLP is a turnkey modification to existing LLM pipelines, including those that rely exclusively on closed-source models. Empirically, we demonstrate that our method yields improvements in both predictive accuracy and uncertainty quantification (UQ) on a range of LLM benchmarks and UQ tasks. More broadly, our work demonstrates a viable path for incorporating methods from the rich Bayesian literature into the era of LLMs, paving the way for more reliable and calibrated LLM-based systems.

Using Reinforcement Learning with Verifiable Rewards (RLVR) to optimize Large Language Models (LLMs) can be conceptualized as progressively editing a query's 'Reasoning Tree'. This process involves exploring nodes (tokens) and dynamically modifying the model's policy at each node. When combined with data scheduling, this process yields further gains in data efficiency and accuracy. However, existing RLVR data scheduling methods typically rely on path-based metrics to rank queries, overlooking the reasoning tree structures of these queries. In this paper, we introduce a novel metric, namely **Reasoning Score (r-score)**, which measures the query's learning difficulty based on the structure of its reasoning tree. Based on the r-score, we propose the **Reasoning Tree Schedule (Re-Schedule)**, a scheduling algorithm that constructs a curriculum progressing from structurally simple (high r-score) to complex (low r-score) queries. Experiments on six math-reasoning benchmarks show that Re-Schedule significantly improves average accuracy, achieving gains of up to 3.2\%. These strong results validate our approach and demonstrate that a structural understanding of the reasoning tree provides a more powerful and principled foundation for RLVR data scheduling.

Growing concerns over the theft and misuse of Large Language Models (LLMs) underscore the need for effective fingerprinting to link a model to its original version and detect misuse. We define five essential properties for a successful fingerprint: Transparency, Efficiency, Persistence, Robustness, and Unforgeability. We present a novel fingerprinting framework that provides verifiable proof of ownership while preserving fingerprint integrity. Our approach makes two main contributions. First, a "chain and hash" technique that cryptographically binds fingerprint prompts to their responses, preventing collisions and enabling irrefutable ownership claims. Second, we address a realistic threat model in which instruction-tuned models' output distribution can be significantly altered through meta-prompts. By incorporating random padding and varied meta-prompt configurations during training, our method maintains robustness even under significant output style changes. Experiments show that our framework securely proves ownership, resists both benign transformations (e.g., fine-tuning) and adversarial fingerprint removal, and extends to fingerprinting LoRA adapters.

High-quality time series (TS) data are essential for ensuring TS model performance, rendering research on rating TS data quality indispensable. Existing methods have shown promising rating accuracy within individual domains, primarily by extending data quality rating techniques such as influence functions and Shapley values to account for temporal characteristics. However, they neglect the fact that real-world TS data can span vastly different domains and exhibit distinct properties, hampering the accurate and efficient rating of diverse TS data. In this paper, we propose TSRating, a novel and unified framework for rating the quality of time series data crawled from diverse domains. TSRating leverages LLMs' inherent ample knowledge, acquired during their extensive pretraining, to comprehend and discern quality differences in diverse TS data. We verify this by devising a series of prompts to elicit quality comparisons from LLMs for pairs of TS samples. We then fit a dedicated rating model, termed TSRater, to convert the LLMs' judgments into efficient quality predictions by inferring future TS samples through TSRater's inference. To ensure cross-domain adaptability, we develop a meta-learning scheme to train TSRater on quality comparisons collected from nine distinct domains. To improve training efficiency, we employ signSGD for inner-loop updates, thus circumventing the demanding computation of hypergradients. Extensive experimental results on eleven benchmark datasets across three time series tasks, each using both conventional TS models and TS foundation models, demonstrate that TSRating outperforms baselines in terms of estimation accuracy, efficiency, and domain adaptability.

Large language model (LLM)-based agents increasingly rely on tool use to complete real-world tasks. While existing works evaluate the LLMs' tool use capability, they largely focus on the final answers yet overlook the detailed tool usage trajectory, i.e., whether tools are selected, parameterized, and ordered correctly. We introduce TRAJECT-Bench, a trajectory-aware benchmark to comprehensively evaluate LLMs' tool use capability through diverse tasks with fine-grained evaluation metrics. TRAJECT-Bench pairs high-fidelity, executable tools across practical domains with tasks grounded in production-style APIs, and synthesizes trajectories that vary in breadth (parallel calls) and depth (interdependent chains). Besides final accuracy, TRAJECT-Bench also reports trajectory-level diagnostics, including tool selection and argument correctness, and dependency/order satisfaction. Analyses reveal failure modes such as similar tool confusion and parameter-blind selection, and scaling behavior with tool diversity and trajectory length where the bottleneck of transiting from short to mid-length trajectories is revealed, offering actionable guidance for LLMs' tool use.