NeurIPS 2026 Papers

Large Reasoning Models (LRMs) rely on extensive Chain-of-Thought generation but suffer from memory bottlenecks due to linear Key-Value (KV) cache growth. While low-bit quantization offers a potential solution, we identify that current static quantization strategies suffer from severe degradation on complex reasoning tasks at extreme bit-widths. Notably, this accuracy drop is accompanied by increased generation length, which ironically limits the intended efficiency gains. We attribute this performance collapse to the fact that static quantization methods are insufficiently flexible to meed dynamic demands of precise long-range retrieval. Utilizing Average Attention Distance (AAD) to quantify retrieval patterns, we find that the attention patterns in shallow layers exhibit strong locality, whereas that in deep layers demonstrate intensive long-range dependencies. Furthermore, deep layers have fluctuating AAD values across token positions. Based on above findings, we propose AdaKVQ, an adaptive mixed-precision quantization framework that modulates the bit-width allocation guided by the dynamic demands of long-range context retrieval in each layer. Extensive experiments on complex reasoning benchmarks with Llama-3.1 and Qwen2.5 demonstrate that AdaKVQ consistently outperforms state-of-the-art baselines, achieving performance comparable to full-precision models.

Large language models face substantial computational challenges in long-sequence inference due to the quadratic cost of self-attention, with the prefill stage being particularly bottlenecked. Static sparse attention reduces compute via predefined masks but can miss task-critical evidence, while dynamic sparsification is more flexible yet can be brittle across tasks, leading to either insufficient coverage or distractor-induced noise. We propose AdaCal, an Adaptive Calibration framework for robust sparse attention that synergizes coarse, experience-driven task dispatch with head-wise physical profiling. By deriving entropy, concentration, and mid-zone coverage metrics from a lightweight probe pattern, AdaCal dynamically governs head activation and augmentation policies, allocating additional computation strictly on demand. This design is realized through task-conditional policies that effectively suppress unnecessary expansion for structure-dominant inputs while enhancing long-range coverage for retrieval-intensive tasks. Experiments show that AdaCal improves robustness over strong static and dynamic baselines, with clearer gains in long-context settings beyond 32K tokens.

In training Mixture-of-Experts models, sharded data parallelism splits each expert’s parameters across GPUs. Before each layer runs, GPUs must use Allgather to rebuild the full weight matrix. This communication can take a large part of each training iteration. Prior work reduces this cost with lossy compression, but lossy methods can affect accuracy. We propose Adaptive Exponent Reuse Allgather, or AdERA, a lossless Allgather compression method for sharded MoE training. Our key observation is that, after a short warmup phase, most weight exponents stay the same across iterations. AdERA stores these exponents locally and sends only the sign and mantissa when the exponent has not changed. The receiver then rebuilds the exact original weight by combining the cached exponent with the received sign and mantissa. Because parameter matrices have different shapes across layers, AdERA compresses only when compression saves time. It also overlaps compression with Allgather communication and computation. Our real experiments and large-scale trace-driven simulator show that, compared with the lossless baseline, AdERA achieves a 3.70× speedup on 16 GPUs and is projected to reach 4.28× on 128 GPUs, while preserving bitwise-exact parameter reconstruction. Compared with the lossy baseline, AdERA achieves a 3.93× speedup on 8 GPUs and is projected to reach 4.42× on 128 GPUs. The source code is publicly available.

Modern neural-network optimizers occupy a three-way design space: online adaptivity, architecture-aware geometry, and acceleration. We introduce AcceleGrad#, an accelerated linear-coupling template that couples an adaptive anchor with a norm-induced sharp, mirror, or linear-minimization geometry branch. The scalar-anchor specialization recovers the accelerated $O(1/K^2)$ rate for smooth convex objectives, while the diagonal-AdaGrad clipped-LMO training variant admits local $\mu$-Kurdyka-Lojasiewicz certificates: a generic conservative $O(K^{-1/4})$ rate and, in a SCION cross-entropy regime, a $O(K^{-1/3})$ last-iterate rate to a self-bounded mini-batch noise level. On image classification and language modeling, the training variant improves over Euclidean AcceleGrad and remains competitive with recent non-Euclidean optimizers. Using stochastic gradients directly in spectral oracles also reduces polar-computation time and optimizer overhead in practice.

Self-distillation enables language models to learn in an on-policy fashion from their own trajectories by using the same model as both the teacher and student, with the teacher being conditioned on privileged information that is unavailable to the student. This information can take multiple forms or views, including solutions, demonstrations, or feedback; for example, for a math problem, one view might be a ground truth reasoning trace. Supervising the student via the privileged teacher enables fine-grained token-level feedback and removes misalignment caused by distillation from external policies with different semantic biases. However, self-distillation also creates a fundamental asymmetry: the teacher may encode view-specific information tied to the privileged information provided during training, which the student cannot access at inference time. Moreover, the best form of privileged information is often task-dependent, making it difficult to choose a single teacher view. In this work, we address both these challenges jointly by introducing AVSD (Adaptive-View Self-Distillation), a novel method of self-distillation with multiple privileged-information views, which reconstructs token-level supervision by separating stable cross-view consensus from view-specific residual support. The consensus signal provides a robust update direction, while the residual signal is selectively used to adjust the update magnitude. Together, these components form a gated aggregation mechanism that balances a conservative target (geometric mean) with a more permissive target (arithmetic mean) for aggregating teacher signals at the token level. Experiments on math competition benchmarks (AIME24, AIME25, and HMMT25) show that AVSD consistently outperforms single-view self-distillation baselines and GRPO, achieving 3.1\% Avg@8 gain over the strongest baseline on Qwen3-8B on average. Moreover, on code-generation benchmarks (Codeforces, LiveCodeBench) using Qwen3-8B, AVSD outperforms single-view self-distillation baselines by 2.9\% on average. More broadly, our analysis shows that AVSD provides better token-level learning signals than any single-view method by reconstructing supervision from multiple privileged teachers.

Federated Learning (FL) is vulnerable to model poisoning attacks, where malicious clients manipulate uploaded model updates to corrupt the global training process. Existing attacks are typically evaluated by their eventual damage, such as final accuracy degradation or convergence to random-guess performance, where the learned model becomes unusable and may ultimately lead to denial-of-service (DoS). However, such metrics overlook attack round complexity, defined as the number of communication rounds required for the poisoned global model to reach and remain near a desired degradation objective. In realistic FL systems, requiring more communication rounds increases the adversary's exposure to client participation, robust aggregation, filtering, and detection. We propose ASAP (Adaptive Sliding Agnostic Poisoning), a fast aggregation-rule-agnostic (AGR-agnostic) attack that treats the whole FL training process as an uncertain dynamical system and formulates poisoning as a feedback-control problem for steering the model toward a desired attack objective. ASAP combines adaptive sliding mode control (ASMC) with finite Fourier-basis uncertainty estimation, treating the unknown effects of benign training and aggregation as a time-varying uncertainty and using the resulting estimate to design the malicious updates. We theoretically prove that the sliding variable reaches the sliding surface in finite time and the tracking error subsequently converges exponentially. Experiments across multiple datasets, models, and aggregation rules demonstrate that ASAP reaches specified degradation objectives in fewer communication rounds while maintaining smaller target deviation than existing model poisoning attacks.

Pre-training neural operators on diverse PDE datasets has emerged as a promising paradigm for building general-purpose surrogate models in scientific machine learning. However, the **inherent complexity** and **structural diversity** of PDE solution operators make multi-PDE pre-training fundamentally difficult. Existing approaches address this mainly by enlarging model capacity, while the target solution operators themselves remain unchanged. Inspired by classical numerical analysis---where a problem-tailored *operator transformation* reformulates a complex PDE into a numerically easier equivalent---we propose to similarly reformulate the target solution operators, turning a heterogeneous set of complex, structurally divergent operators into a simpler and better-aligned equivalents. Since different PDEs admit different simplifications, the transformation must be *adaptive and input-dependent*, so that a single neural operator can approximate the entire family jointly. We instantiate this idea as AOT-POT (**A**daptive **O**perator-**T**ransformation for **P**re-training **O**perator **T**ransformer), which realizes such a transformation by expanding the hidden representation into multiple parallel streams, aggregating and redistributing them with input-dependent weights before and after each sub-layer, and mixing streams through Sinkhorn-projected doubly stochastic matrices for stable training. Together, these mechanisms reformulate the diverse, complex solution operators into a simpler equivalents that a single architecture can approximate jointly. Empirically, AOT-POT achieves state-of-the-art results on 12 PDE benchmarks with only 3% additional parameters, reducing the relative L2 error by up to 77.6% (40.9% on average). Fine-tuning further reduces L2 error by up to 92% on in-domain PDEs and 89% on out-of-domain PDEs, confirming that adaptive operator transformation is an orthogonal and effective axis for advancing PDE foundation models, beyond merely scaling model capacity.

Scaling network depth has been a central driver behind the success of modern foundation models, yet recent investigations suggest that deep layers are often underutilized. This paper revisits the default mechanism for deepening neural networks, namely residual connections, from an optimization perspective. On a motivating example, rigorous analysis proves that the layout of residual connections can fundamentally shape convergence behavior, and even induces an *exponential* gap in convergence rates. Prompted by this insight, we introduce adaptive neural connection reassignment (ANCRe), a principled and lightweight framework that parameterizes and learns residual connectivities from the data. ANCRe adaptively reassigns residual connections with negligible computational and memory overhead (<1%), while enabling more effective utilization of network depth. Extensive numerical tests across pre-training of large language models, diffusion models, and deep ResNets demonstrate consistently accelerated convergence, boosted performance, and enhanced depth efficiency over conventional residual connections.

Multimodal Large Language Models (MLLMs) often exhibit significant modality preference, which is a tendency to favor one modality over another. Depending on the input, they may over-rely on linguistic priors relative to visual evidence, or conversely over-attend to visually salient cues over textual facts. Prior work has applied a uniform steering intensity to adjust the modality preference of MLLMs. However, strong steering can impair standard inference and increase error rates, whereas weak steering is often ineffective. Since steering sensitivity varies substantially across instances, a single global strength is difficult to calibrate. To address this, we introduce an instance-aware diagnostic metric Modality Contribution Ratio (MCR) that quantifies each modality’s information contribution and reveals sample-specific susceptibility to steering. Building on this signal, we propose a scaling strategy and a learnable module for instance-aware control of modality preference. Experiments show that AMPS outperforms conventional steering on $MC^2$ by improving preference control with lower generation collapse, and improves MLLM performance on general multimodal QA dataset such as MM-Vet v2.

Domain specialization of large language models (LLMs) often improves target-domain performance but degrades broad capabilities. Model merging is an attractive training-free alternative, yet common parameter-space strategies (for example, interpolation and sparsified averaging) provide limited control over task-specific interference. We introduce Adaptive Model Alignment via Ranked Inherited Subspaces (AMARIS), a covariance-driven framework that transfers specialist updates only along selected activation directions. AMARIS formulates merging as budgeted subspace inheritance: it maximizes specialist activation capture while constraining base-activation deviation, then builds a low-rank projection gate from ranked generalized eigendirections under an explicit preservation budget. This yields a single static merged checkpoint that concentrates transfer in high-utility, low-interference directions. We evaluate AMARIS on Qwen3-based merges (8B and 32B) across finance, instruction-following, and multilingual SEA specialization. In our evaluated settings, AMARIS reaches Pareto operating points that are competitive with strong training-free baselines, including 98.37\% specialist retention with 95.76\% core retention (finance), 103.49\% core retention with 90.28\% specialist retention (instruction following), and 93.69\% specialist retention with 99.77\% core retention (SEA). These results support subspace-constrained merging as a practical route to compositional capability integration without additional training.

Representation alignment has recently emerged as an effective paradigm for accelerating Diffusion Transformer training. Despite their success, existing alignment methods typically impose a fixed supervision target or a fixed alignment granularity throughout the entire denoising trajectory, whether the guidance is provided by external vision encoders, internal self-representations, or VAE-derived features. We argue that such timestep-agnostic alignment is suboptimal because the useful granularity of representation supervision changes systematically with the signal-to-noise ratio. In high-noise regimes, diffusion models benefit more from coarse semantic and layout-level anchoring, whereas in low-noise regimes, the training signal should emphasize spatially detailed and structurally faithful refinement. This non-stationary alignment behavior creates a representational mismatch for static single-level supervisors. To address this issue, we propose Adaptive Hierarchical Prior Alignment (AHPA), a lightweight alignment framework that exploits the hierarchical representations naturally embedded in the frozen VAE encoder. Instead of using only a single compressed latent as the alignment target, AHPA extracts multi-level VAE features that provide complementary priors ranging from local geometry and spatial topology to coarse semantic layout. A timestep-conditioned Dynamic Router adaptively selects and weights these hierarchical priors along the denoising trajectory, thereby synchronizing the alignment granularity with the model's evolving training needs. Extensive experiments show that AHPA improves convergence and generation quality over baselines and incurs no additional inference cost while avoiding external encoder supervision during training.

Reinforcement learning (RL) has substantially improved the ability of large language model (LLM) agents to interact with environments and solve multi-turn tasks. However, effective agentic RL remains challenging: sparse outcome-only rewards provide limited guidance for assigning credit to individual steps within long interaction trajectories. Existing approaches often introduce dense intermediate supervision, such as process reward models or auxiliary self-supervised signals, which increases supervision and tuning complexity and may limit generalization across tasks and domains. We present AEM, a supervision-free credit assignment method that adaptively modulates entropy dynamics during RL training to improve the exploration-exploitation trade-off. Since in agentic RL the environment is typically affected by a complete response, rather than an individual token, our analysis lifts entropy dynamics from the token level to the response level, aligning uncertainty estimation with the effective action granularity of LLM agents and reducing sensitivity to token-level sampling noise. We further show that entropy drift under natural-gradient updates is governed by the interaction between the sampled-response advantage and its relative surprisal. Motivated by this result, AEM derives a practical response-level uncertainty proxy and uses it to rescale advantages, leveraging the evolving balance between positive and negative samples to naturally transition from exploration to exploitation. Extensive experiments on ALFWorld, WebShop, and SWE-bench-Verified with models ranging from 1.5B to 32B demonstrate that AEM consistently improves strong RL baselines, including a +1.4\% gain when integrated into a state-of-the-art software-engineering RL training framework.

The KV cache is a major memory bottleneck in LLM inference, especially under large batch sizes and long contexts. While quantization is a promising solution, existing static group-wise quantization methods are constrained by a trade-off between metadata overhead and accuracy: increasing the group size reduces per-group metadata (e.g., zero-points, scaling factors, and outlier indices), but exacerbates quantization errors under the sinusoidal-like temporal drift of post-RoPE outlier channels; conversely, using smaller groups improves accuracy but incurs substantial metadata overhead. In this paper, we propose ADKV, a low-overhead adaptive delta quantization framework for KV cache in LLM inference, which leverages the temporal structure of the KV cache. ADKV tracks per-channel zero-points and scaling factors via online EMA, adapting to the sinusoidal-like temporal drift of post-RoPE outlier channels, thereby eliminating the rigidity of static group-wise quantization. This adaptive mechanism further enables delta coding to embed metadata directly into the quantized sequence, decoupling metadata overhead from context length. To minimize the MSE of attention outputs, we design an SGD-based calibration algorithm that models ADKV as an RNN. Experiments on generation and long-context benchmarks show that, while matching the accuracy of static group-wise quantization, ADKV reduces the metadata memory footprint by $\sim 12.8\times$. A custom fused ADKV-GEMV CUDA kernel optimized for memory-bound autoregressive generation achieves $\sim 1.96\times$ throughput over the FP16 baseline.

Language model post-training is often bottlenecked by the need for human-collected preference data, which is expensive and difficult to scale. RLAIF-style approaches that leverage pseudo labels offer an abundant alternative but introduce systematic biases that degrade downstream alignment. Recent general-purpose semi-supervised methods correct for teacher bias using a small set of human-labeled examples, but suffer from high variance especially when human annotations are scarce. To this end, we propose ABC-Align, leveraging abundant pseudo label signal to minimize variance and applying a lightweight, adaptive correction grounded in the human-labeled subset. The correction strength is tuned automatically during training using plug-in estimates of the relevant bias--variance quantities. On LLM alignment with RLHF and DPO where human feedback is scarce, we empirically demonstrate that ABC-Align achieves superior performance over prior semi-supervised baselines.

We apply offline reinforcement learning to large-scale warehouse consolidation, an application that requires ranking tens of thousands of candidates at each decision epoch with an architecture that scales gracefully with variable-cardinality action spaces. Since the available reward signal is only an approximation of the true target-domain objective, the policy must also transfer robustly. We propose a two-stage framework, Asymmetric Implicit Q-Learning (AIQL) and its adaptive cross-domain extension A$^2$IQL. First, AIQL learns a scoring policy from fixed offline consolidation logs by decoupling a per-candidate scoring actor from an aggregate-statistic critic, keeping value estimation tractable and within the support of logged data. Second, A$^2$IQL adapts the source-domain-trained policy to a new target objective using scarce target-domain data: a Bayesian reward predictor supplies an uncertainty-penalized conservative bonus that modifies only the advantage-weighted actor extraction, leaving the source critic unchanged. On a single-domain benchmark, AIQL improves throughput by 20% over a heuristic baseline using offline data alone. When adapted to a new target objective via A$^2$IQL, the method outperforms source-only training, data pooling, and off-dynamics baselines by up to 23% under a $30{:}1$ source-to-target data ratio while achieving the lowest variance across all methods.

Agents built on large language models (LLMs) rely on a range of reliability techniques, including retry, majority voting, and self-consistency, that have been developed in parallel rather than within a common analytical framework. We observe that an LLM sampled at temperature $T$ is a discrete stochastic channel $p(y \mid x)$ in the sense of Shannon's coding theory, and use this identity as the entry point for such a framework grounded in communication theory. Each of these techniques is a special case of one of six classical reliability operators: diversity combining, hybrid retransmission, iterative generator-critic decoding, rateless sampling, structured redundant verification, and difficulty-adaptive routing. Within the framework we give two closed-form results: a noise-variance threshold above which uniform averaging beats quality-weighted averaging, and a contractivity criterion for generator-critic refinement, consistent with a contractive-to-divergent transition we observe between 3B- and 14B-parameter models. We further introduce a cost-aware semantic-nearest-neighbor router whose single Lagrangian knob traverses the quality-cost frontier without retraining. Across six channel configurations spanning local and cloud models on 69 hard tasks, no fixed model-technique-budget choice dominates, motivating per-task allocation. On a 300-item hard split of MMLU, GSM8K, and HumanEval, our router occupies the full empirical Pareto frontier: at matched quality, its normalized cost is ${\approx}56$\% lower than the strongest fixed technique; at matched normalized cost, it improves quality by ${\approx}7$\% ($26$\% over single-shot decoding). These results argue for consolidating these reliability techniques into a single tunable layer informed by channel coding.

Many modern learning approaches are still struggling with spatial reasoning tasks, i.e. they lack the ability to utilize geometric information of perceived entities and their spatial relation to each other to solve problems. We introduce a novel Adaptive Neural Cellular Automata (aNCA) architecture which uses deformable convolutions to dynamically adapt the perceptive field and iteratively reason over 2D spatial relations on grid-like data structures (e.g. images). Empirical results on public benchmarks show state of the art comprehensible results with high generalization abilities for solving image based puzzles like Sudoku or finding the shortest path in a maze.