NeurIPS 2026 论文列表

Building general-purpose role-playing agents that faithfully portray any character from a natural-language profile remains challenging. The dominant paradigm---supervised fine-tuning---encourages behavioral mimicry without deep, human-like internal thought processes, resulting in poor out-of-distribution generalization. Therefore, we propose Psy-CoT, a psychology-grounded chain-of-thought framework that decomposes pre-response reasoning into three role-specific steps---Interaction Perception, Psychological Empathy, and Logical Construction---so that the model thinks dynamically from the profile rather than merely mimicking surface patterns. While structured reasoning provides a foundation, it alone is insufficient; reinforcement learning is essential to further align the model with character fidelity. However, we observe that under LLM-based reward models, both generic phrases that hack the reward model and genuinely role-specific phrases receive identical gradient signals---this hacking accumulates over training, misleading the model into treating both as equally optimal choices. To address this, we propose Role-Aware Policy Optimization (RAPO), which uses profile--token mutual information to weight gradients asymmetrically---amplifying role-specific tokens under positive advantage while attenuating them under negative advantage. Experiments on CoSER, CharacterBench, and CharacterEval demonstrate that Psy-CoT outperforms existing role-playing CoT methods, and RAPO consistently surpasses GRPO across multiple model scales.

Hypergraphs record multi-way interactions among entities. Extracting information from the combinatorial structure underlying observed multi-way interactions is a central task in many real-world problems. Existing methods face several limitations. First, many deep architectures for hypergraphs do not explicitly exploit the potential low-rank structure, which can sacrifice parsimony and interpretability in the learned representations. Second, many low-rank-based methods operate on tensor representations, which typically require hyperlinks to have uniform sizes and thus limit their applicability to general hypergraphs with non-uniform hyperlink sizes. Third, many methods ignore the fact that hyperlinks often arise from heterogeneous mechanisms. For example, medical symptoms may co-occur in the profiles of patients with very different conditions, and such heterogeneity should be incorporated into the learning process. In this work, we develop a general framework for hypergraph representation learning using hyperlink random effects while exploiting the low-rank structure in hypergraphs. The proposed framework accommodates latent heterogeneity in hyperlink formation while preserving entity interaction patterns. We establish identifiability of the model parameters and leverage an expectation-maximization strategy for estimation. The framework allows flexible specifications for the hyperlink random effects; in this paper, we study three choices: categorical, Gaussian mixture, and score-based effects, and develop corresponding estimation algorithms. Through simulation studies, we demonstrate the effectiveness of the proposed method in recovering latent structure and capturing heterogeneous interaction patterns. Empirical studies on real-world hypergraph datasets further illustrate the practical utility of our approach.

Self-improving AI systems aim to reduce reliance on human engineering by learning to improve their own learning and problem-solving processes. Existing approaches to self-improvement rely on fixed, handcrafted meta-level mechanisms, fundamentally limiting how fast such systems can improve. The Darwin G\"odel Machine (DGM) achieves open-ended self-improvement in coding, but this relies on domain-specific alignment between task performance and self-modification skill. However, this alignment does not generally hold beyond coding domains. We introduce \textbf{hyperagents}, self-referential agents that integrate a task agent (which solves the target task) and a meta agent (which modifies itself and the task agent) into a single editable program. Crucially, the meta-level modification procedure is itself editable, enabling metacognitive self-modification, improving not only the task-solving behavior, but also the mechanism that generates future improvements. We instantiate this framework by extending DGM to create DGM-Hyperagents (DGM-H), eliminating the need for domain-specific alignment to potentially support self-accelerating progress on any computable task. Across diverse domains, the DGM-H improves performance over time and outperforms baselines without self-improvement or open-ended exploration, as well as prior self-improving systems. Furthermore, the DGM-H improves the process by which it generates new agents (e.g., persistent memory, performance tracking), and these meta-level improvements transfer across domains and accumulate across runs. All experiments were conducted with safety precautions (e.g., sandboxing, human oversight). We discuss what safety entails in this setting and the broader implications of self-improving systems. DGM-Hyperagents offer a glimpse of open-ended AI systems that do not merely search for better solutions, but continually improve their search for how to improve.

Real-world image datasets are deeply hierarchical, yet unsupervised hierarchical representation learning remains underdeveloped. Hyperbolic space is geometrically ideal for embedding trees with low distortion due to its exponential volume growth, but existing unsupervised hyperbolic methods either fail to scale beyond a few thousand points or depend heavily on class supervision. We bridge this gap with HyperTree, a fully unsupervised, scalable, and capacity-efficient framework for hierarchical representation learning in hyperbolic space, built on three contributions: (i) a geometrically consistent, closed-form definition of the hyperbolic Lowest Common Ancestor, valid across hyperbolic manifolds of arbitrary curvatures; (ii) a differentiable hierarchical objective based on Dasgupta's cost that is bounded and invariant to dataset scale; and (iii) a sample-complexity guarantee that enables scaling to datasets such as ImageNet-1k and iNaturalist-2018. Across four benchmarks, HyperTree outperforms Euclidean baselines by wide margins on hierarchical dendrogram purity, matches supervised hyperbolic baselines on CIFAR-100, and surpasses them on TinyImageNet at 16 dimensions — a highly capacity-efficient regime where competing methods collapse.

While large language models (LLMs) appear to be increasingly capable of solving compositional tasks, it is an open question whether they do so using compositional mechanisms. In this work, we investigate how feedforward LLMs solve two-hop factual recall tasks, which can be expressed compositionally as $g(f(x))$. We first confirm that modern LLMs continue to suffer from the "compositionality gap", i.e. their ability to compute both $z = f(x)$ and $y = g(z)$ does not entail their ability to compute the composition $y = g(f(x))$. We then decode residual stream representations and identify two processing mechanisms: one which solves tasks *compositionally*, computing $f(x)$ along the way to $g(f(x))$, and one which solves them *directly*, without any detectable signature of the intermediate variable $f(x)$. Finally, we find that embedding space geometry is strongly related to which mechanism is employed, where the idiomatic mechanism is dominant when tasks are represented by translations from $x$ to $g(f(x))$ in the embedding spaces.

We study signSGD in linear regression with diagonal features and power-law spectra, a toy but expressive setting where the effect of coordinate-wise sign normalization can be analyzed sharply. Specifically, we establish a functional scaling law (FSL) for signSGD under general learning rate schedules, which decomposes the loss dynamics into two components: learning of the target signal and noise accumulation. Compared with SGD, signSGD learns the signal faster but exhibits slower noise forgetting. This signal-noise tradeoff yields a regime-dependent comparison between signSGD and SGD: signSGD achieves better data-scaling efficiency in hard-task regimes, while they perform similarly in easy-task regimes. The key mechanism is that gradient noise is {\it curvature-aligned}: its directional variance is of the same order as the directional curvature. As a result, sign operation effectively acts as a \(\diag(\bH)^{-1/2}\) preconditioner in the noise-dominated regime, where $\bH$ is the Hessian matrix. Moreover, we demonstrate that the same square-root preconditioning mechanism can arise in RMSprop and Adam, where coordinate-wise normalizers essentially track the coordinate curvatures. Synthetic experiments validate the proposed mechanism and the predicted scaling behavior. Language model experiments further suggest that, beyond the controlled setting, the resulting FSL can serve as a surrogate model for fitting and predicting practical training loss curves.

Spatial transcriptomics (ST) provides spatially resolved gene expression measurements, but most 3D ST datasets are acquired as independently processed 2D sections. This sliced acquisition creates a difficult reconstruction problem: measurements are dense within each section but sparse along the axial direction, and inter-section technical variation can be confounded with true biological change. Existing coordinate-based implicit neural representations (INRs) provide a continuous modeling framework, but standard isotropic coordinate encodings do not explicitly reflect this axial-to-lateral imbalance. We present HoloGene, an anisotropic conditional INR framework for estimating continuous 3D gene-expression fields from pre-registered ST sections. HoloGene first compresses high-dimensional expression profiles into a graph-autoencoder latent manifold. It then uses pseudo-label-derived biological signature conditioning to provide coarse intra-section biological context, while depth-dependent FiLM modulation models axial variation. To reduce slice-specific covariance shifts, we introduce a signature-stratified correlation-alignment regularizer over latent features. Using held-out measured sections as cross-sectional references, HoloGene improves reconstruction accuracy and biological preservation metrics compared with general-purpose INRs and adapted spatial-transcriptomics baselines. Ablation analyses show that decoupled depth conditioning improves axial generalization, while correlation alignment provides the strongest reduction in slice-specific representation leakage. These results support anisotropic conditional modeling as a practical strategy for estimating continuous 3D expression fields from discontinuous ST sections.

Selective unlearning in large language models is increasingly necessary for safety and compliance, yet harmful or memorized content is often latent, entangled, and hard to specify as a clean ``forget set.'' We aim to close this gap by identifying what to forget from the model’s own failure modes rather than relying solely on explicit supervision. We propose Hallucination-Guided Unlearning (HGU), which treats hallucinations as diagnostic signals: it localizes unsupported spans via evidence and internal consensus, traces a sparse set of high-responsibility MLP neurons, and edits them with compact gates under a utility-preserving KL constraint. Across TOFU, HGU achieves the best MU--FE trade-off with Avg of 0.7221 (forget05) and 0.7699 (forget10), and it also attains the top SepS H-Avg (0.4382/0.4715) under mixed forget/retain prompts; on WMDP auxiliary, it yields the strongest forgetting (e.g., 14.2 on Economics) while improving Retain (52.7) and MMLU (57.2). These results suggest that reframing hallucinations as actionable traces enables targeted, post-hoc unlearning that better balances forgetting efficacy and general utility without full retraining.

Understanding how aligned LLMs internally represent safety is critical for diagnosing alignment vulnerabilities, as it explains why jailbreaks succeed and informs the design of robust alignment strategies. Prior work shows that aligned LLMs encode harmfulness and refusal as separable directions in the residual stream at prompt-side token positions. We show that jailbreaks succeed at prompt encoding by suppressing either the refusal or harmfulness direction before any token is generated, with distinct attack classes occupying separable regions of the harmfulness-refusal plane. Extending the analysis to response-token positions, we find that the model recognizes harmful content while it is generating that content, even when it failed to recognize the input as harmful at the prompt side. Motivated by our findings, we introduce HARC (Harmfulness-And-Refusal Coupling), a fine-tuning method that pairs the two directions across both prompt and response positions. Since the intervention is confined to the harmfulness-refusal subspace, it leaves the rest of the residual stream intact and does not degrade general capability or inflate over-refusal. Across extensive experiments, HARC achieves the strongest robustness-capability-usability trade-off among six baselines spanning the major training-time and inference-time safety methods. The harmfulness and refusal directions at prompt and response positions transfer across the five model families and two scales we tested without architecture-specific tuning.

Hallucinations in large language models (LLMs) that are plausible-sounding but factually incorrect or unsupported pose a major challenge for deploying these models in high-stakes applications such as medical diagnosis, legal reasoning, and knowledge-based question answering. Existing detection methods primarily rely on either generating multiple responses to check consistency or supervised training on labeled hallucinations. Multiple-response approaches tend to be computationally expensive and may be less practical for real-time scenarios, while supervised methods are limited to known hallucination types and cannot generalize to unseen cases. To address these limitations, we propose an efficient unsupervised hallucination detection method. Our approach measures the embedding discrepancy between the LLM’s response considered independently and the response contextualized with the question. A large discrepancy indicates that the question significantly influences the response, suggesting it is grounded, whereas a small discrepancy signals a higher likelihood of hallucination. This method is efficient, interpretable, and does not require labeled hallucinations. Extensive experiments across multiple tasks demonstrate that our approach consistently achieves superior or comparable detection performance while reducing the detection cost compared to the multiple-response methods.

Gradient descent dynamics on the deep matrix factorization problem is extensively studied as a simplified theoretical model for deep neural networks. Although the convergence theory for two-layer matrix factorization is well-established, no global convergence guarantee for general deep matrix factorization under random initialization has been established to date. To address this gap, we provide a polynomial-time global convergence guarantee for randomly initialized gradient descent on four-layer matrix factorization, given certain conditions on the target matrix and a standard balanced regularization term. Our analysis employs new techniques to show saddle-avoidance properties of gradient descent dynamics, and extends previous theories to characterize the change in eigenvalues of layer weights.

Packed high-token 3D fields are not just long-sequence modeling: once a volume is packed into tokens, sequence order is no longer the geometry of the sample. This creates a general failure mode for attention over packed geometric data: sequence-index positional mechanisms can learn the memory layout rather than the object. We introduce Pack3D, a geometry-aware local--global attention operator that separates memory layout from attention geometry. Pack3D keeps local neighborhood attention exact, compresses only distant context into contiguous 3D block summaries, and indexes attention with axial RoPE on true 3D token-grid coordinates rather than packed sequence positions. With RoPE applied before block compression, each query interacts with a geometry-aware mixture of block tokens, so compressed global context remains spatially interpretable after packing. A synthetic repacking benchmark makes the failure mode explicit: learned positional embeddings and packed-sequence RoPE fail under layout changes, whereas axial grid-coordinate RoPE remains invariant. On Allen WTC--11 hiPSC microscopy, our main claim-grade domain, Pack3D improves matched learned-position local--global attention by 0.0978 best macro Dice and 0.1248 final macro Dice at 262k tokens and depth 8. In an Allen-15 joint comparison, Pack3D also improves over a parameter-matched TokenGrid 3D U-Net on both best and final macro Dice. Inference-time ablations show that the trained model genuinely uses its compressed summaries, and the fused forward/backward implementation is $6.49\times$ lower latency than dense SDPA at 65k tokens.

Model merging aims to combine multiple models into one without additional training. Naïve parameter-space averaging can be fragile under architectural symmetries, as their geometry does not take them into account. In this work we show that not only the geometry, but also the averaging procedure itself, must be symmetry-invariant to achieve symmetry-aware merges. Consequently, we propose a general solution: merging as Fréchet averaging, i.e., selecting parameters that minimize a sum of geodesic distances on an appropriate manifold. In this view, the key design choice is the overall geometry, i.e., the choice of metric, manifold, and distance approximation, that determines what it means for two models to be “close.” We show that Fréchet averaging, combined with simplifying assumptions, contains Fisher merging. Building on this, we examine the particular case of low-rank adapters (LoRA), whose symmetries induce a distinct geometry: that of a quotient manifold. We outline the limitations of current LoRA merging methods, propose a practical algorithm for this setting, and show how they compare with other commonly used approaches.

General-purpose agents perform tasks in unfamiliar environments without domain-specific manual customization. Yet no study has systematically measured how agent architecture shapes performance across heterogeneous protocols and diverse unfamiliar environments. This is the first systematic study, comparing tool-calling, MCP, code-generation, and CLI agents on the same benchmarks with the same models. Two gaps blocked such a study: existing harnesses require per-benchmark wiring or fixed protocol classes (web for BrowserGym, CLI for Harbor), and benchmarks themselves expect human-authored prompts, context, and integration glue. To enable this study, we contribute (1) a unifying protocol that bridges existing benchmark and agent protocols; (2) an evaluation harness that surfaces any benchmark to any general-purpose agent and backbone model; and (3) the first Open General Agent Leaderboard of agent configurations, a full factorial over 5 agent architectures × 5 backbone LLMs (three closed-source, two open-weight) × 6 benchmarks spanning software engineering, customer service, deep research, and personal assistance. We find that (i) general agents adapt to every tested domain without per-domain customization; (ii) agent architecture choice swings results by up to 12pp within a single model, yet backbone model choice dominates overall performance; (iii) on 4 of 6 tested benchmarks, top general agents are indistinguishable from the leading heavily-customized domain-specific agents; (iv) open-weight models tested exhibit "generality sinks" absent from frontier closed-source models: they consistently collapse on specific agent architectures or benchmarks. Code, harness, leaderboard, and traces will be released upon acceptance.

Transformers excel at sequence modeling, yet Softmax attention incurs quadratic complexity and unbounded KV cache growth. While linear attention offers a promising alternative, existing approaches lack systematic functional comparison with Softmax attention, rigorous error analysis, and a theoretically grounded improvement roadmap. We address this gap by framing linearization as KV cache approximation and establishing a principled pathway from Softmax attention to linear models. Our analysis identifies five critical components—redundancy elimination, token-level quantization with positional separation, positional compression, inter-layer similarity, and multi-state decomposition—each accompanied by theoretical justification and error bounds, with explicit connections to existing mechanisms. Building upon this framework, we introduce GPA, a linearized attention model that inherits pretrained weights and achieves state-of-the-art results. GPA outperforms strong baselines including MVA and GSA across multiple benchmarks, while requiring less fine-tuning resources. Our work provides both theoretical clarity and practical guidance for advancing linear attention, charting a principled course toward efficient, scalable alternatives to Softmax attention.

The advancement of large language models~(LLMs) has significantly accelerated the development of search agents capable of autonomously gathering information through multi-turn web interactions. However, existing benchmarks often construct queries backward from answers, producing tasks misaligned with real-world needs. Moreover, these benchmarks tend to focus on either locating specific information or aggregating information from multiple sources, while relying on static answer sets prone to data contamination. To bridge these gaps, we introduce GISA, a benchmark for General Information-Seeking Assistants comprising 373 human-crafted queries reflecting authentic information-seeking scenarios. GISA features four structured answer formats (item, set, list, and table), enabling deterministic evaluation. It integrates both deep reasoning and broad aggregation within unified tasks, and includes a live subset with periodically updated answers to resist memorization. Notably, GISA provides complete human search trajectories for every query, serving as valuable references for analyzing and improving search agent behavior. Experiments on mainstream LLMs and commercial search products reveal that even the best model achieves only 19.30% exact match, with performance notably degrading on tasks requiring complex planning and comprehensive information gathering.

Learning expressive representations for Boolean logic circuits is a fundamental challenge at the intersection of graph learning and Electronic $\mbox{Design}$ Automation (EDA). Existing Graph Neural $\mbox{Networks}$ (GNNs) primarily rely on topological message passing, which often fails to capture the strict causal dependencies and discrete functional semantics of logic gates. In this paper, we $\mbox{propose} \textbf{FuncFormer}$, a Graph Transformer that incorporates $\textbf{functional simulation}$ not as a proxy task, but as a fundamental inductive bias directly into the representation learning process. Unlike standard GNNs which typically rely on isotropic aggregation, FuncFormer $\textbf{encodes the intrinsic flow of functional propagation}$ by analyzing randomized simulation traces as they evolve through the network. This approach effectively aligns the continuous embedding manifold with the discrete Boolean function space, effectively mitigating structural aliasing. By integrating these deterministic signal trajectories with a scalable dual-path attention mechanism, our model preserves functional consistency across long-range dependencies in both combinational and sequential circuits. Empirical results demonstrate that FuncFormer significantly outperforms state-of-the-art models (e.g., DeepGate4) in Quality-of-Results (QoR) prediction and formal verification tasks, exhibiting robust generalization to unseen circuit scales.

Large language models are increasingly used to estimate uncertain quantities from context, raising the question of whether their probabilistic outputs are internally consistent. A basic test of self-consistency is whether these estimates adhere to the law of total probability. We investigate this requirement through the lens of binary conditioning trees, which recursively partition a population into increasingly fine-grained subsets and expose multiple ways of estimating the same aggregate quantity. This construction yields a basic self-consistency check: marginal estimates must agree with prior-weighted aggregations of conditional estimates over any partition of the population. We find that current models systematically violate this requirement. In a case study on persona prompting, prior-weighted aggregates are consistently better aligned with human population statistics than direct estimates. Notably, this benefit of specificity persists even when the prior weights are themselves estimated by the LLM. Turning this discrepancy into a general evaluation criterion, we propose a family of self-consistency checks for LLMs grounded in the law of total probability. By evaluating these checks across frontier models, we show that failures of statistical consistency are widespread and not confined to persona prompting. Together with a benchmark dataset, our work provides a testbed for understanding when in-context learning can be treated as conditional inference.

Tools from random matrix theory have become central to deep learning theory, using spectral information to provide mechanisms for modeling generalization, robustness, scaling, and failure modes. While often capable of modeling empirical behavior, practical computations are limited by matrix size, often imposing a restriction to models that are too small to be realistic. This motivates the inference of properties of larger models from the behavior of smaller ones. Free decompression (FD) is a recently proposed method for extrapolating spectral information across matrix sizes, but its utility is currently limited by strong assumptions that preclude its implementation on more realistic machine learning (ML) models. We use algebraic spectral curve theory to provide a general FD methodology for spectral densities whose Stieltjes transform satisfies an algebraic relation, a modeling assumption that is more likely to hold in practice. This recasts FD as an evolution along spectral curves which can be readily integrated. Our framework enables the expansion of spectral densities that have multiple or multi-modal bulks, that exist at multiple scales, and that contain atoms, all characteristic of real-world data and popular ML models. We demonstrate the efficacy of our framework on models of interest in modern ML, including Hessian and activation matrices associated with neural networks and large-scale diffusion models.