NeurIPS 2026 论文列表

Tiny object detection (TOD) remains challenging due to the extremely limited spatial details of small objects. Existing methods heavily rely on multi-scale feature pyramids with discrete feature maps, which suffer from insufficient spatial resolution, quantization errors, and poor localization accuracy, while imposing heavy computational and memory overhead. We present the implicit DEtection TRansformer (iDETR), an efficient detector that operates exclusively on low-resolution single-scale features. It incorporates two novel components: (1) implicit Attention (iAttn), which leverages implicit neural representations to model continuous features and enables precise sub-pixel querying beyond discrete grid limitations; and (2) Centroid-Guided Query Initialization (CGQI) for robust query initialization under single-scale constraints. We further propose head-conditional sampling in iAttn, which reduces querying computational cost by 4× and memory footprint by 3× without sacrificing performance. Extensive experiments demonstrate the effectiveness of iDETR. Especially, on AI-TODv2, iDETR outperforms state-of-the-art methods by 0.2% AP overall, with particularly strong gains of 1.1% AP on very tiny objects, while reducing computational cost by 37%. The code is available at https://anonymous.4open.science/r/iDETR-FC25/.

Deep Equilibrium Models (DEQs) are an established framework for image restoration that learn a problem-adapted regularization by solving a fixed-point (i.e. equilibrium) problem. While flexible and expressive, DEQs are often hindered by high computational cost and training instability. We propose an inertial DEQ (i-DEQ) that learn an explicit nonconvex regularization within the DEQ formulation. By using momentum within the fixed-point iterations, i-DEQ has convergence guarantees and accelerated rates. Moreover, we observe that i-DEQ is significantly more stable during the training and robust to rough initialization than DEQs. Numerical experiments on various linear and nonlinear inverse problems demonstrate that i-DEQ achieves reconstruction quality comparable to state-of-the-art methods, while reducing DEQ's inference time by a factor of two.

Benchmark quality is critical for meaningful evaluation and sustained progress in time series forecasting, particularly with the rise of time series foundation models (TSFMs). Existing benchmarks often have limited domain coverage or overlook real-world settings such as tasks with covariates. Their aggregation procedures frequently lack statistical rigor, making it unclear whether performance differences reflect true improvements or random variation. Many benchmarks lack consistent evaluation infrastructure or are too rigid for integration into existing pipelines. To address these gaps, we propose fev-bench, a benchmark of 100 forecasting tasks across seven domains, including 46 with covariates. Supporting the benchmark, we introduce fev, a lightweight Python library for forecasting evaluation emphasizing reproducibility and integration with existing workflows. Using fev, fev-bench employs principled aggregation with bootstrapped confidence intervals to report performance along two dimensions: win rates and skill scores. We evaluate various models on fev-bench and find that existing TSFMs often miss accuracy by ignoring covariates, pointing to a clear direction for future work.

Influence functions are a popular tool for measuring the impact of a training point on a model's prediction; in particular, self-influence, which quantifies the influence of a training point on itself, has found many uses such as data selection and outlier detection. However, the use of influence functions has been severely limited in modern models such as LLMs due to their low accuracy or high computational cost: most existing algorithms are either highly inaccurate, or require computing gradients or approximations to inverse Hessians, which can be prohibitive for large models. In this work, we introduce a highly efficient zeroth-order approximation to data influence that requires only a fraction of the time and memory footprint of previous methods and is applicable to both differentiable and non-differentiable loss functions. We demonstrate that in addition to its computational advantage, our method delivers superior accuracy in estimating self-influence and comparable or better accuracy in estimating train-test influence for fine-tuned large language models, paving the way for broader and more practical application of influence-function techniques in state-of-the-art AI systems.

Generating realistic human trajectories is essential for mobility simulation but remains challenging when they must generalize across cities and respect road network constraints. In practice, existing methods often suffer from three key limitations: geographic overfitting, length-rigid generation, and off-road drift when generating GPS coordinates directly. To address these issues, we propose \textbf{XTraj}, a transferable coarse-to-fine autoregressive framework for road-consistent trajectory generation. Specifically, XTraj first learns transferable road-segment representations by integrating road geometry, POI context, historical traffic intensity, and graph topology. Based on these representations, it then autoregressively generates variable-length road-segment routes under road-connectivity constraints. Finally, instead of directly regressing latitude-longitude coordinates, XTraj further refines each generated route in a route-progress space by predicting monotonic progress increments along valid road geometry, from which GPS points are recovered by interpolation. This route-aligned design naturally supports variable-length GPS generation and guarantees road-geometry consistency by construction. Experiments on two real-world vehicle trajectory datasets show that XTraj improves spatial fidelity over competitive baselines, transfers effectively to unseen cities in a zero-shot setting, and ensures road-geometry consistency by construction without any post-hoc map matching. The implementation is provided in \textbf{https://anonymous.4open.science/r/XTraj-EBE7/}.

Heterogeneous multi-agent LLM systems, where agents are powered by different model families, consistently outperform homogeneous configurations. Yet existing communication protocols either operate through text, discarding the sender’s internal representations, or require architectural homogeneity for latent-level transfer. We identify the entity grounding problem in cross-architecture communication: cross-attention bridges that transfer continuous representations across different LLM families suffer from rare-token compression collapse, where entity identity is lost in the continuous bottleneck (bridge-only F1 ∼30%). We propose XBRIDGE, a decode-free communication protocol that resolves this through two mechanisms. Lexical Anchor Mapping (LAM) maps the sender’s original context tokens to the receiver’s vocabulary, providing discrete entity anchors. A Latent Enrichment Bridge (LEB) lets the receiver query the sender’s hidden states for contextual enrichment. The entity anchors ground the bridge’s relational signals to specific entities through the receiver’s own self-attention. Across three model families (Llama, Qwen, Mistral), seven benchmarks, and both communication directions, XBRIDGE outperforms text-based communication (NLComm) on all 7 tasks per model pair with 11×lower latency, and exceeds the same-architecture KVComm ceiling on 6 of 7 tasks. LEB requires only 264M trainable parameters (3.8% of the receiver), trains on 587 samples in under 10 minutes, and adds negligible inference overhead.

Learning robust robot policies in real-world environments requires diverse data augmentation, yet scaling real-world data collection is costly due to the need for acquiring physical assets and reconfiguring environments. Therefore, augmenting real-world scenes into simulation has become a practical augmentation for efficient learning and evaluation. We present WorldComposer, an automated generative Real2Sim2Real framework that maps real-world panoramas into high-fidelity, task-ready simulation environments and further synthesizes diverse Digital Cousins through scene and object variations. Combined with high-quality physics engines and realistic assets, WorldComposer supports interactive manipulation across rigid, articulated, and deformable objects. Additionally, we incorporate multi-room stitching to construct consistent large-scale environments for long-horizon tasks. Experiments demonstrate a strong sim-to-real correlation validating our platform's fidelity, and show that extensively scaling up data generation leads to significantly better generalization to unseen scene and object variations, demonstrating the effectiveness of Digital Cousins for generalizable robot learning and evaluation.

Mixture-of-Experts (MoE) architectures are central to large-scale deep learning, relying on sparse execution and expert specialisation. Despite their success, their large-scale training dynamics remain poorly understood, and training is often unstable. While scaling limits such as infinite-width theory have clarified the behavior of dense models, the scaling behavior of MoEs remains largely unexplored. To this end, we derive the infinite-width limits of MoE architectures under fixed number of experts with both soft and Top-$K$ routing, trained with SGD and Adam, using Tensor Programs. Our results show that under the Standard Parameterisation (SP), router updates freeze after one step of training in both soft and Top-$K$ routing MoEs. This provides, at least in part, a principled explanation for the widespread use of auxiliary losses such as load-balancing or $z$-losses in MoE training. In dense networks, Maximal Update heuristics allow for stable and non-vanishing feature updates. However, in MoEs the Maximal Update heuristic with soft-routing and softmax gating causes the experts to collapse to the same distribution and router gradients to vanish. While sigmoid allows non-trivial router evolution, they still fail to prevent expert collapse. Thus fixed expert soft-routing MoEs do not admit a parameterisation that simultaneously ensures stability, feature learning, and expert specialisation in the infinite-width limit. We then show that, beyond computational efficiency, Top-\(K\) also acts as a symmetry breaking mechanism allowing for both feature learning and specialisation. We empirically validate these predictions, observing close agreement between finite-width training dynamics and the predicted scaling behaviour.

Deciding which member of a group produced a vocalization and where that vocalization was produced are central to studying social communication but are rarely evaluated with physically verified ground truth. Existing resources either localize sounds without identifying the caller, or recognize individuals from isolated vocalizations without reliance on the 3d candidate geometry. We close this gap with V33DA, a multimodal benchmark for spatial vocal attribution in social zebra finches: it comprises 33,625 verified events coupling 5-channel audio with three camera views, calibrated 3D pose for every visible candidate, and per-bird radio telemetry across 10 individuals and three experiments. Caller identity is verified from an on-body accelerometer-derived vibration signal; this channel is withheld from benchmark models and used only by an oracle ceiling. We also provide V33DA++ as an auxiliary extension with verified overlapping-call events and $\pm 2$s context windows, intended to support future source-separation and vocal-activity-detection studies. V33DA separates familiar-individual recognition from transferable spatial attribution. We evaluate candidate-conditioned caller attribution, with 3D source localization as a secondary diagnostic, under three regimes: session-disjoint testing, held-out-experiment transfer, and leave-one-individual-out evaluation. A broad set of reference methods exposes a consistent gap between in-domain accuracy and transfer under identity shift: methods that can exploit familiar vocal identity perform well on known callers but degrade sharply on unseen ones, whereas candidate-aware methods preserving explicit spatial reasoning transfer substantially better. The accelerometer oracle stays near-perfect across regimes, confirming the internal consistency of the accelerometer-verified labels and quantifying the ceiling available when the withheld verification channel is observed. We release V33DA with fixed attribution/localization protocols and V33DA++ with overlap and long-context tasks, together with code for the evaluated methods.

Direct Preference Optimization (DPO) has become a standard method for preference-based fine-tuning in RLHF pipelines. A practical but largely heuristic step in DPO is dataset curation: given a fixed budget of preference labels, which candidate pairs should be compared to produce the most informative training set and the best downstream policy? This paper studies DPO dataset curation as a sampling-design problem. We model the curated dataset by a sampling design over candidate pairs and analyze how this design propagates through DPO training to the KL-regularized RLHF objective. Our main results provide matching upper and lower bounds on the RLHF optimality gap of the policy learned from a curated DPO dataset. The bounds reveal a simple structural message: the effect of pair selection enters only through a single design-dependent matrix that summarizes how informative the collected comparisons are for learning the policy parameters. This leads to a trace-form criterion that characterizes both an achievable performance guarantee for the DPO estimator and an information-theoretic lower bound for any induced policy estimator, thereby identifying a canonical objective for comparison curation under budget constraints.

Mechanistic studies of grokking often begin after an endpoint circuit has been identified. We study the preceding retrospective triage problem: given a solved final checkpoint and a long saved trajectory, which path and window should be inspected first? We propose solution-frame path triage. The final checkpoint fixes a solution frame; earlier checkpoints are then scored path by path. The Geometric Coherence Score (GCS) measures coherence in this procedure: it asks whether neighboring inputs in the final frame undergo similar local Jacobian transformations through the attention, MLP, or full block measurement maps. Centered Kernel Alignment (CKA) and $L_2$ distance to the final path state provide the complementary proximity scores. Across 104 modular arithmetic Transformer runs, GCS reveals a repeatable path schedule that simple norm and dimensionality summaries miss: attention coherence rises near grokking and often turns over, while MLP maps diversify and later reconverge. After selection, held-out checks show that the resulting windows are enriched for accuracy and algebraic changes, align with Fourier progress on modular addition, and identify attention-to-MLP interface states with high replacement cost. The method returns a short list of path/window hypotheses for endpoint analysis. It is retrospective by design: the final checkpoint is part of the procedure.

Offline RL can learn effective policies from fixed datasets, but deployment objectives often shift after training, while the trained actor may be frozen due to data, cost, or governance constraints. We study deployment-time adaptation of frozen offline actors using Product-of-Experts (PoE) composition with a goal-conditioned prior. Our main practical finding is graceful degradation rather than universal improvement: with degraded or random priors, precision-weighted PoE remains anchored to the frozen actor, whereas additive and prior-only adaptation often collapse, and a KL-budget selector typically reaches a near-oracle operating point. We also clarify that, for diagonal-Gaussian actors and priors, PoE with coefficient $\alpha$ induces the same deterministic policy as KL-regularized adaptation with $\beta=\alpha/(1-\alpha)$, with posterior covariance differences cancelling under deterministic action selection. Across D4RL MuJoCo and AntMaze diagnostics, results reveal an actor-competence ceiling: composition helps or preserves performance when the frozen actor is sufficiently competent, but cannot recover when the actor itself fails. Swept CQL/IQL-guided and SF/GPI baselines are competitive in some settings, but do not eliminate this regime boundary. Thus, PoE/KL-Reg is best viewed as an actor-anchored safety layer for deployment-time objective shift, not as a general return-maximization method, with effectiveness governed by actor competence rather than knob choice.

Long-horizon LLM agents solve interactive tasks by repeatedly choosing the next action from the current state. At many steps, a compact decision is sufficient to specify the action, yet the base agent still emits a full realization in its native output format (e.g., code or action text). Once the compact decision determines the executable content of the step, subsequent generation mostly adds formatting and elaboration, yielding redundant realization. This separation is reflected inside the decoder: compact decisions become locally recoverable in middle layer representations before full realizations, and attention over its token span becomes concentrated when the decision is sufficiently supported as the current step output. Building on this separation, we propose MIRA, a training-free inference method for Model-Internal Reasoning Amortization. MIRA amortizes realization through two complementary stages: Consensus-Guided Decision retrieves successful reference steps in the middle layer representation space to propose a compact decision candidate. Attention-Guided Commitment evaluates its token span using token confidence and attention concentration. The candidate is used as the step output only when both signals support it for the current state; otherwise, the agent falls back to the base agent's full realization path. Experiments on long-horizon interactive agents show that MIRA improves task performance while reducing redundant realization, yielding a 47.5\% reduction in output tokens and a 24.7\% reduction in inference time across tasks and agent formats.

Conversational large language models can respond with warmth and reassurance while failing to advance the user's stated goal. We conceptualize this failure as $\textbf{goal displacement}$, the substitution of persona-maintenance content for goal-relevant substance. To measure it, we introduce $\textbf{GDB}$ (Goal Displacement Benchmark), comprising 230 scenarios across nine domains and $4{,}134$ human-scored responses from 18 models. A $2 \times 2$ taxonomy crosses hard constraints with emotional distress, distinguishing failures to confront infeasible goals or combine empathy with practical guidance from reverse displacement, where unsolicited problem-solving replaces requested emotional support. A deletion-test rubric assesses whether substantive help remains after purely supportive sentences are removed. Independent raters achieve weighted $\kappa = 0.82$ against author scores for displacement severity. Trap displacement ranges from 0.0\% to 37.7\% and is strongly negatively correlated with Chatbot Arena Elo ($\rho = -0.706$, $n = 15$). Reverse displacement reaches 41.9\%, including substantial rates in models with negligible trap displacement. A factorial probe supports a distinction from social sycophancy. Human-scored responses in the high-complexity, low-pressure condition exhibit 35.0\% sycophancy but 7.5\% displacement, with little association between the constructs. Matched OLMo-2 checkpoints show that mixed empathy-with-constraints displacement remains at 52.3\% after supervised fine-tuning and falls to 13.8\% after direct preference optimization. GDB provides a reproducible diagnostic of whether responses advance the operationalized user goal, supporting model comparison, post-training evaluation, and regression testing. We release the dataset, annotations, and rubric under CC-BY-4.0 and the evaluation code under MIT.

Precisely reproducing a target span verbatim within a long context is a fundamental capability test for LLMs: even when the target string is present in context, models often still drift within the span or fail to stop correctly, revealing limitations in source grounding and boundary control. This capability has broad practical relevance. For instance, code generation often requires embedding sensitive strings such as API keys exactly as they appear in context. In this work, we construct a benchmark to test exact span reproduction ability across contracts, scientific literature, code, and random sequences, and find that current models still make surprisingly frequent errors. To address this, we propose CopyGen, a lightweight copy-constrained plugin for frozen LLMs that turns copying from a byproduct of next-token prediction into a controlled process. Our key idea is to decompose copying into three distinct decisions: when to copy, how far to copy, and when to stop. Concretely, CopyGen first determines whether to enter copy mode, then predicts a tentative copy length, and finally verifies the generated span token by token. During verification, we adaptively adjust the logits of copy continuation tokens to incentivize or suppress continuation, and emit only the longest valid prefix. Across three LLM backbones and four benchmark tasks, CopyGen improves average exact match by 35% relative without model finetuning, and integrates seamlessly with SFT to further improve it by 4%, while substantially accelerating generation. These results suggest that reliable exact copying is not something LLMs obtain for free, but a decoding-time control problem that requires explicit modeling. Code is available at https://anonymous.4open.science/r/CopyGen-A65D.

Practitioners often combine independently trained models as a fixed pool of frozen experts. Standard combination methods, such as voting, averaging, or stacking, rely on final predictions, amounting to a few scalars per expert, and discard potentially richer cross-expert information found in the penultimate representations. An alternative is to combine these models in feature space. Whether this yields a meaningful improvement depends heavily on the evaluation design. Informal assessments on the same data can support materially different conclusions depending on baseline strength, overlap control, reporting granularity, and pool construction. We argue that auditing the sensitivity of evaluation outcomes to these design choices is critical for the effective composition of expert models. To address these sensitivities, we propose a four-part evaluation protocol: (1) a baseline ladder for prediction-space (PS) models; (2) a deterministic feature-space (FS) anchor plus a higher-capacity cross-check; (3) axis-isolated pools to disentangle performance gains; and (4) grouped evaluation with overlap-aware controls. Applying our protocol to 21 frozen deepfake detectors from DF40 (Yan et al., 2024), we find that a 10K-parameter logistic regression on raw concatenated features outperforms the strongest PS baseline by +0.123 AUROC on an architecture-diverse pool (generator-level bootstrap 95% CI [+0.078, +0.174]). Crucially, the higher-capacity cross-check (a 3.5M-parameter feature-space MLP) adds only +0.004 AUROC, suggesting that the gain arises from the richer feature interface rather than from combiner complexity. Prediction-space models suffice when individual expert performance is high but fall short in architecture-diverse, high-bottleneck pools. Our primary contribution is an evaluation-design audit showing that, on this benchmark and expert pool, relaxing any of these controls risks overstating the apparent FS advantage or creating the illusion of a gain that fails to hold across data subgroups. Our findings demonstrate that, without rigorous controls, the same data can support multiple, sometimes spurious, conclusions.

We study learning-augmented algorithms for scheduling on multiple parallel machines under the objectives of minimizing the makespan or the total weighted completion time. We assume that the algorithm has access to noisy information (the prediction) about an optimal solution, expressed through a solution encoding that implicitly quantifies both the amount of information provided and the nature of noise (errors in the prediction). Our central question is to determine the maximum level of noise that still allows for efficiently recovering this optimal solution. We observe that information guiding greedy scheduling to optimality or revealing the optimal assignment of tasks to machines is very sensitive to noise, and prove that recovering an optimal solution from solution encodings capturing either is computationally hard, even for a small (non-constant) number of errors. On the positive side, we show that an order-based encoding, which leads to errors that are dispersed locally, is robust to noise. We present a dynamic programming framework that utilizes this encoding and solves optimally several classical scheduling problems, including makespan and total weighted completion time minimization on identical or unrelated machines.

Recent work has shown that flow matching models can be trained without explicit time conditioning, challenging the standard view that the interpolation time is needed to disambiguate velocity targets. But why should a time-blind model work at all? Decomposing the time-blind flow matching loss, we identify two sources of irreducible error: a *coupling variance*, which arises from ambiguous velocity targets induced by how noise and data points are paired, and the *time-blindness gap*, which is the additional error caused by ignoring time. This gap shows that time-blind training is strictly harder than conventional training, reinforcing the puzzle that time-blind models work so well in practice. We resolve this tension by showing that the geometry of high-dimensional data makes time identifiable directly from noisy observations. When data concentrates near a $k$-dimensional subspace, time can be recovered from the statistical structure of noisy interpolants in directions orthogonal to the data; under a spiked-covariance model, this yields a closed-form estimator that recovers $t$ from a single observation $z$ at rate $O(1/\sqrt{d-k})$ for ambient dimension $d$. As a consequence, we prove that the time-blindness gap is asymptotically negligible relative to the coupling variance. We empirically demonstrate our identifiability result on real-world data and show that changing the coupling has a much larger effect on loss and sample quality than removing time conditioning across CIFAR-10, CelebA-HQ, and FFHQ. These results explain why time-blind flow matching works and show that the main practical lever is the choice of coupling, not explicit time conditioning.

Room acoustics modeling requires capturing intricate wave phenomena beyond direct sound propagation, including reflections, refractions, and diffractions. Recent neural acoustic synthesis methods have progressively incorporated richer scene representations and physics-informed priors, yet typically learn acoustic behavior without considering wave propagation dynamics, which inherently capture diffraction and interference phenomena. We propose WavNAF, a framework that bridges wave dynamics and neural acoustic modeling by leveraging Finite-Difference Time-Domain (FDTD) simulations, which solve the wave equation on discrete spatiotemporal grids to capture these phenomena. However, directly applying FDTD at full scale is computationally prohibitive under CFL stability constraints. Inspired by traditional acoustic scale models, WavNAF constructs a geometry-preserving, scaled digital replica from NeRF-derived scene geometry and acoustically probes it with temporally compressed FDTD simulation. The resulting pressure maps are encoded not as direct full-scale Room Impulse Response (RIR) estimates, but as compressed-domain wave signatures that provide the neural acoustic field with a structured inductive bias beyond static geometry. To bridge the mismatch between these scaled simulation features and full-scale acoustic responses, we introduce a Neural Acoustic Scaling Module that adaptively recalibrates pressure-map features at the feature level for full-scale RIR prediction. Experiments show that WavNAF yields consistent gains over prior neural baselines across standard acoustic metrics.