NeurIPS 2026 Papers

Watermarking is an important tool for promoting the responsible use of large language models (LLMs). Existing watermarks insert a signal into generated tokens that either flags LLM-generated text (zero-bit watermarking) or encodes more complex messages (multi-bit watermarking). Though a number of recent approaches insert multiple bits into text without perturbing average next-token predictions, they largely extend design principles from the zero-bit setting, such as encoding a single bit per token. In contrast, a watermarker capable of embedding multiple bytes into the text would dramatically increase the potential applications, by embedding information such as the ID of the user who submitted the prompt, the precise model version that was used, or even the prompt itself. We address this problem by introducing ArcMark: a new watermark construction based on coding and information-theoretic principles that is capable of reliably embedding multiple bytes of information into just a few hundred tokens, without any distortion of the underlying LLM next-token distribution. We derive ArcMark by formulating the distortion-free watermarking problem as a channel coding problem, and deriving an information-theoretic channel capacity that establishes the fundamental limit of embedding information in LLM output in a distortion-free manner. This capacity formulation informs the design of ArcMark. In practice, ArcMark outperforms competing multi-bit distortion-free watermarks in terms of reconstruction accuracy, including in the face of attacks that alter a subset of the LLM text. ArcMark output is also shown to be indistinguishable from unwatermarked text in terms of perplexity, and in downstream task quality.

Formulaic alpha discovery is a core challenge in quantitative trading, as identifying alphas that work well together remains difficult. Recent reinforcement learning (RL) methods formulate this task as a Markov decision process (MDP), but two important issues remain unresolved. First, as the alpha pool evolves, the reward function changes accordingly, making the MDP inherently non-stationary. Second, most existing methods optimize a single objective, typically predictive power, while ignoring other important properties of a high-quality alpha pool. Motivated by these challenges, we propose AlphaPareto, an RL method for formulaic alpha discovery. To address non-stationarity, AlphaPareto augments the state to include both the alpha under construction and the current alpha pool, and applies a large language model (LLM) to encode the pool. This design allows the agent to adapt to the evolving search environment. To overcome the limitation of single-objective reward design, AlphaPareto replaces the scalar reward with a multi-objective vector-valued reward that simultaneously captures predictive power, temporal stability, perturbation robustness, and diversity, and optimizes these objectives through a Pareto-regularized learning procedure. Empirical applications to real-world datasets show that our AlphaPareto method outperforms its competitors.

Modern agricultural data is sourced from diverse platforms and spans multiple spatial scales, ranging from ground-level close-up photography to Unmanned Aerial Vehicle (UAV) aerial observation and satellite remote sensing imagery. Accordingly, agricultural multimodal reasoning demands robust cross-scale spatial understanding. However, due to the lack of multi-view agricultural benchmark datasets, existing multimodal large language models (MLLMs) exhibit severe ground-level bias, which leads to scale confusion then semantic collapse in agricultural perception tasks, such as misinterpreting farmland imagery as walls or floors. To address this, we introduce \textbf{AgroOmni}, a large-scale multi-view training corpus with 288K Visual Question Answering pairs covering 56 specialized task categories across 14 task types, designed to capture diverse scales in modern precision agriculture. Built on this dataset, we propose AgroNVILA, which achieves a new state-of-the-art of 62.32\% on the AgroMind benchmark ($+15.03\%$ over GPT-5.2), effectively mitigating the multi-view cross-scale gap for holistic agricultural understanding. Diagnostic evaluations on AgMMU further reveal an inherent heterogeneity between macro-priors and micro-diagnostics through constrained zero-shot performance. Meanwhile, even minimal fine-tuning leads to a dramatic performance gain of AgroNVILA on AgMMU, strongly demonstrating its generalization capability empowered by AgroOmni. Full training scripts are publicly available at \url{https://anonymous.4open.science/r/AgroOmni-6510}.

Agentic tasks require a large language model to interact with the world, navigating information and gathering evidence across multiple steps with restricted resources. Due to this complexity, agentic task failures arise from various sources, and pinpointing these failure causes is essential to diagnose and improve agentic systems. Existing benchmarks, however, tend to focus on a single leaderboard score, leaving the underlying failure modes opaque. To fill this gap, we introduce AgentHop, a diagnostic benchmark of 1,011 multiple-choice questions paired with a controlled seven-tool sandbox under fixed token, turn, and tool-call constraints. AgentHop reveals model vulnerabilities by dissecting a single accuracy score along four axes of agent operation: retrieval, synthesis, tool-call, and resource management. Across 19 models, we find that behavior clusters by model family, with tool-call signatures revealing distinct family fingerprints: GPT models commit early, Anthropic and GLM checkpoints verify before committing, DeepSeek and Kimi over-search, and Gemini-3 Pro stays balanced. Decomposed axes further expose within-family structure: Claude Opus 4.6 and Sonnet 4.6 land within one accuracy point yet diverge on retrieval-versus-synthesis emphasis, with Opus retrieving more and Sonnet synthesizing better. We release the full benchmark set and the harness to support diagnostic agent benchmarking.

Multi-agent systems achieve state-of-the-art outcomes through peer collaboration. However, when an agent in the pipeline silently drops a constraint, the system's final output may look correct even though the reasoning chain was quietly corrupted, and existing outcome-based evaluations are blind to such multi-hop process failures. To make these vulnerabilities measurable before deployment, we introduce AgentCollabBench, a diagnostic benchmark of 900 human-validated tasks spanning software engineering, DevOps, and data engineering. Each task isolates one of four behavioral risks: instruction decay (does a constraint survive peer pressure?), false-belief contagion (does a falsehood spread through consensus?), context leakage (does information bleed between tasks?), and tracer durability (does marked data reach the final agent?). Evaluating four modern LLMs (GPT 4.1 mini, Gemini 2.5 Flash Lite, Qwen-3.5-35B-A3B, and Llama 3.1 8B Instruct), we expose model-specific vulnerability profiles invisible to outcome-only evaluation; Qwen-3.5-35B-A3B, for example, leads on tracer durability and instruction stability, while GPT 4.1 mini leads on leakage containment and false-belief resistance. Beyond per-model differences, communication topology emerges as a primary risk factor that explains 7-40% of the variance in multi-hop information survival. The effect traces to a synthesis bottleneck specific to converging-DAG nodes: an agent weighing competing parent inputs discards constraints carried by a minority branch, a bottleneck structurally absent from linear chains. AgentCollabBench demonstrates that suboptimal topology can silently erase the safeguards of highly capable models, arguing that multi-agent reliability is fundamentally a structural problem and that scaling model intelligence alone is no substitute for architecture.

While large language model (LLM) multi-agent systems achieve superior reasoning performance through iterative debate, practical deployment is limited by their high computational cost and error propagation. This paper proposes AgentArk, a novel framework to distill multi-agent dynamics into the weights of a single model, effectively transforming explicit test-time interactions into implicit model capabilities. This equips a single agent with the intelligence of multi-agent systems while remaining computationally efficient. Specifically, we investigate three hierarchical distillation strategies across various models, tasks, scaling, and scenarios: reasoning-enhanced fine-tuning; trajectory-based augmentation; and process-aware distillation. By shifting the burden of computation from inference to training, the distilled models preserve the efficiency of one agent while exhibiting strong reasoning and self-correction performance of multiple agents. They further demonstrate enhanced robustness and generalization across diverse reasoning tasks. We hope this work can shed light on future research on efficient and robust multi-agent development. Our code is at https://anonymous.4open.science/r/AgentArk.

Multi-turn conversation is the predominant form of interaction with large language models (LLMs), where user queries continuously evolve. However, existing LLM routing methods are primarily designed for single-turn interactions or fixed queries settings, overlooking the dynamic nature of multi-turn conversations and the challenge of delayed rewards, thereby limiting their ability to optimize cumulative performance. To address this challenge, we move from myopic, single-turn selection to long-horizon routing for multi-turn conversation. Accordingly, we propose ConvoRouter, which first performs MCTS to explore conversation branches induced by different LLM selections and collect trajectories with high cumulative rewards. ConvoRouter then learns a lightweight routing policy from search-derived data, augmented with retrieval-based future state approximation, enabling multi-turn routing without online search. Experiments on both open-domain and domain-specific conversation tasks across diverse candidate sets of both open-source and closed-source LLMs demonstrate that ConvoRouter significantly outperforms single LLMs and existing routing baselines in task success rate, while achieving a superior performance-cost trade-off when combined with a cost-aware reward.

Large agentic models can now operate in complex digital games by pretraining on web-scale data, but gameplay trajectories often contain highly redundant actions. Single-action agents can exploit this redundancy spuriously by copying recent actions, leading to causal confusion. Fixed-length action chunks reduce this ambiguity by predicting multiple future actions at once, but their open-loop execution accumulates errors because the agent cannot adapt to new observations during the chunk. The agent should therefore learn when to act for longer without feedback and when to stop early for a new observation. We introduce AdaMAP (Adaptive Multi-Action Prediction), an adaptive multi-action prediction framework that directly generates variable-horizon action chunks, interleaves actions with grounded visual dreams that provide verifiable future constraints, and optimizes the resulting policy with offline imitation learning followed by online MAP-PPO. Across Minecraft and VizDoom, AdaMAP attains the best or tied-best result in five of six task categories and improves average Minecraft success by 3.56 percentage points over the fixed-horizon RL baseline. The learned horizons vary systematically across tasks and games, suggesting that adaptive multi-action prediction discovers useful temporal abstractions.

Reinforcement learning (RL) with verifiable rewards has proven effective at post-training LLMs for coding, yet deploying separate task-specific specialists incurs costs that scale with the number of tasks, motivating a unified multi-task RL (MTRL) approach. However, existing MTRL methods treat all coding tasks uniformly, relying on fixed data curricula under a shared optimization strategy, ultimately limiting the effectiveness of multi-task training. To address these limitations, we propose \textbf{{\tool}}, a multi-t\textbf{AS}k code reinforcement learning framework via u\textbf{T}ility-driven co\textbf{OR}dination. Centered on \textit{task utility}, a signal capturing each task's learning potential and cross-task synergy, {\tool} comprises two coupled modules: \textit{1) \moduleone} module hierarchically allocates training budget and prioritizes informative prompts, steering training toward the most valuable data; and \textit{2) \moduletwo} module dynamically scales per-task KL regularization, matching update constraints to each task's current training state. Experiments on two widely-used LLMs across four representative coding tasks demonstrate that {\tool} consistently improves a single model across all tasks, outperforming the best task-specific specialist by 9.0\%--9.5\% and surpassing the strongest MTRL baseline by 7.5\%--12.8\%.

Optimizing nonlinear preferences in multi-objective reinforcement learning (MORL) is essential for capturing complex trade-offs like risk aversion or fairness. However, such non-linearity has historically bifurcated nonlinear MORL objectives into two distinct paradigms: Scalarized Expected Return (SER) and Expected Scalarized Return (ESR). While SER requires global-level optimization and ESR requires non-Markovian policies, leading to fragmented optimization strategies, we bridge this divide through the Aggregation–Expectation–Transformation (AET) framework. By unifying both criteria through a tripartite decomposition of scalarization, AET provides a principled foundation for general nonlinear MORL. Building on this framework, we propose AETDICE, a tractable offline RL algorithm for AET objectives. By utilizing DICE-style density-ratio estimation in an augmented state space, AETDICE enables sample-based optimization from static datasets. Our framework resolves long-standing barriers and captures respective trade-offs induced by AET framework, which existing methods fail to address.

Leveraging Graph Neural Networks (GNNs) as graph encoders and aligning the resulting representations with Large Language Models (LLMs) through alignment instruction tuning has become a mainstream paradigm for constructing Graph Language Models (GLMs), combining the generalization ability of LLMs with the structural modeling capacity of GNNs. However, existing GLMs that adopt GNNs as graph encoders largely overlook the problem of aligning GNN-encoded representations across domains and tasks with the LLM token space to obtain unified graph tokens, thereby limiting their ability to generalize across diverse graph data. To bridge this gap, we aim to incorporate a multi-domain, multi-task GNN encoder into GLMs and align its representations with LLMs to enable multi-domain, multi-task graph alignment instruction tuning. This alignment problem remains underexplored and poses two key challenges: 1) learning GNN-encoded representations that are simultaneously generalizable across domains and tasks and well aligned with textual semantics is difficult, due to substantial variations in graph structures, feature distributions, and supervision signals, together with the lack of textual-semantic alignment guidance in task-specific GNN training; 2) diverse graph data and task-specific instructions can exhibit different degrees of compatibility with the LLM token space during instruction tuning, leading to varying alignment difficulty and rendering a fixed alignment strategy suboptimal. To tackle these challenges, we propose UniGraphLM, a Unified Graph Language Model that incorporates a multi-domain, multi-task GNN encoder to learn generalizable graph representations aligned with textual semantics, and then adaptively aligns these representations with the LLM. Specifically, we first develop a graph-text pair pretraining strategy with a tailored GNN encoder, trained on large-scale graph-text data spanning multiple domains and tasks to obtain generalizable representations naturally aligned with textual semantics. We further design a curriculum alignment tuning strategy that adaptively adjusts the alignment process by accounting for varying alignment difficulty across diverse graph data. Extensive experiments demonstrate that UniGraphLM consistently outperforms state-of-the-art baselines across graph datasets from different domains and tasks.

Virtual sensors replace expensive physical sensors in critical applications through machine learning by predicting target signals from available measurements. Existing virtual sensor approaches require application-specific models with hand-selected inputs for each sensor, cannot leverage task synergies, and lack consistent benchmarks. While emerging time series foundation models offer general-purpose, pretrained solutions in other domains, they are computationally expensive and limited to predicting their input signals, making them incompatible with virtual sensors. We introduce the first multi-task model for virtual sensors addressing both limitations. Our unified model can simultaneously predict diverse virtual sensors exploiting synergies while maintaining computational efficiency. It learns relevant input signals for each virtual sensor, eliminating expert knowledge requirements while adding explainability. In our large-scale evaluation on three standard benchmarks and an application-specific dataset with over 18 billion samples, our architecture reduces computation time by up to 415x and memory requirements by 951x, while maintaining or even improving predictive quality compared to unified baselines. Compared to existing isolated models for a single virtual sensor, our unified approach generates superior predictions at similar inference speed while scaling gracefully to hundreds of virtual sensors with nearly constant parameter count, enabling practical deployment in large-scale sensor networks.

Class-incremental learning (class-IL) faces a structural obstacle: discriminative sequential training optimises only the diagonal blocks of a pairwise loss matrix, leaving the inter-task blocks---task confusion (TC)---unminimised. A recent Infeasibility Theorem makes this precise, proving that no discriminative class-IL learner can attain the joint-training optimum even when catastrophic forgetting (CF) is perfectly addressed. Yet the state of the art consists almost entirely of discriminative methods built on pretrained foundation models that approach joint-training performance. We explain this regime by showing that pretraining \emph{exponentially attenuates} the obstacle. For any pretrained backbone $\phi$, the TC gap of any CF-optimal prototype-based discriminative learner---a class that captures the test-time classifier of SLDA, RanPAC, FeCAM, and the prototype branches of EASE and InfLoRA---is bounded by $\exp(-\gamma_{\min}(\phi)/8)$, where $\gamma_{\min}(\phi)$ is the worst-case Mahalanobis margin between task centroids in feature space: a single, computable scalar that determines how much TC remains. Combined with a finite-sample convergence analysis under heavy-tailed gradient noise and a representation drift bound governing backbone fine-tuning, this yields a diagnostic decomposition of the class-IL excess risk into $E_{\mathrm{CF}} + E_{\mathrm{TC}}$, with disjoint direct dependencies on optimisation budget and backbone quality. Across 12 (backbone, dataset) configurations spanning ViT-IN21K, DINOv2, and CLIP plus four random-backbone anchor points, the pooled empirical slope of $\log(\mathrm{TC\ gap})$ versus $\widehat\gamma_{\min}$ on softmax cross-entropy heads is $-0.119$ (bootstrap 95\% CI $[-0.134, -0.105]$), consistent with the predicted $-1/8$ and excluding the conservative fall-back $-1/16$. The slope fit extends the rate beyond the formal LDA/QDA scope. $E_{\mathrm{CF}}$ scales as $K^{-(\alpha-1)/\alpha}$ as predicted, with the rate steepening by more than $3\times$ under gradient clipping; representation drift scales linearly in the backbone learning rate ($r = +0.998$). A theory-derived training procedure---whose schedule is determined by $\widehat\gamma_{\min}$, the tail index $\widehat\alpha$, and warm-up estimates of smoothness and curvature---is competitive with the strongest pretrained-CIL baselines across the four standard benchmarks.

Online Conformal Prediction (OCP) sequentially constructs prediction sets whose average empirical coverage converges to a target level over time, for arbitrary, potentially adversarial datasets. In the multi-level setting of OCP, prediction sets are produced at $d$ coverage levels at once. Prior work showed that online gradient descent with lazy isotonic projections produces sets properly nested across levels, but with a $\sqrt{d}$ dependence in the per-level coverage convergence rate. Here, we both generalize and improve prior work. First, we show that lazy isotonic projections can be safely used for the construction of multi-level online conformal prediction sets from \emph{any} online learning algorithm. Then, through a black-box reduction, we show that the linearized regret of the online algorithm itself automatically gives control over both per-level calibration and distributional consistency of multi-level predictions via Fenchel duality. Finally, by using the Weighted Interval Score, the standard scoring rule for assessing multi-level interval forecasts, we prove that these projections reduce both the size of the sets and their under-coverage. We show two instantiations of our theory, $p$-norm Online Mirror Descent and coordinate-wise Universal-Portfolio, both improving the dependence on the number of levels to $\sqrt{\log d}$ for the multi-OCP problem. Experiments on stock prices, electricity demand, and synthetic streams confirm consistent gains in coverage-width trade-offs over existing baselines.