ICLR 2027 review scores

TOMC builds memory for long interaction histories on CPU by executing state, relation, and count operations and packing the results with selected source text; it outperforms six baselines on BEAM with three strong LLM...

BLADE learns N:M sparsity masks block-by-block with error compensation and weight re-optimization, matching end-to-end learned mask quality (MaskLLM) on a single GPU at ~56x less compute.

WindWeaver uses coarse winds to guide cross-time feature exchange in flow matching, improving multivariate weather super-resolution.

This paper introduces a pretrain-then-calibrate framework for safe contextual optimization.

We further combine diffusion with DETR through self-trajectory denoising from a noisy encoder prior, which reduces exposure bias and, with DETR training practices, surpasses DINO and DiffuDINO on COCO and LVIS under t...

We identify two efficiency failures that survive recalibration and introduce pooled cost projection, which preserves coverage, bounds added labels, and improves set size under matched budgets in the evaluated tasks.

A novel framework called GASplat, which reconstructs static scene with sparse input containing distractors.

We treat a model's parameters as computable physical particles and other physical quantities, model their changes across iterations with mathematical and physical laws, and thereby derive the parameter update rule wit...

We show that GNNs are not invariant to node IDs althought they should, and propose a method to achieve invariance to node IDs.

pipeline for comprehensive image captioning

Interactive World Model through training free methods

We introduce SimpleARM, a training-free, instruction-conditioned memory that maintains and selectively retrieves task-relevant state, including entity relations, progress, and trajectories.

Across four agent memory frameworks, we trace selective deletion from deletion operations through stored memory and context to task actions, examining the relationships between these stages.

We propose Pair-Aware Contextual Tokens (PACT) framework, a relation representation that preserves global scene context while being aware of the target pair in Video Visual Relation Detection

We show that strategy collapse is a fundamental failure mode of RLVR, explain its mechanism through optimization dynamics and information theory, and propose Mesh Learning to prevent it.

A lightweight class-incremental learning method that statistically models class-conditional patch distributions in the representation space.

KFTP makes video depth predictions more stable over time by carrying information from previous frames into the current frame with uncertainty-aware Kalman fusion.

We introduce the first linguistically and behaviorally motivated fine-tuning framework for identifying beneficial code-switched reasoning behaviors in large language models and teaching these models to code-switch mor...

Global coordination improves agents' capability for automated theorem proving

ScaleComposer combines relational representations with scale-resolved residual learning to predict scalar air-speed fields for configurations with more sources than seen during training.

We introduce RATTS, a reliability-aware test-time scaling method that allocates solution-generation compute only to useful, complementary reasoning concepts.

FactorBench evaluates whether advances in automated factor mining yield signals that generalize, contribute distinct information, and improve portfolio performance after trading costs.

A frozen SAM 3, guided by image-learned forensic evidence and a causal memory bank, localizes video forgeries in space and time, scored by a new gated protocol.

We introduce a path law regularizer for Joint-Embedding World Models (JEPA) that eliminates static latent collapse and significantly improves long-horizon visual planning.

CFF+MAE introduces block-wise ViT pretraining that jointly learns contrastive and masked reconstruction objectives without end-to-end backpropagation, enabling strong representations with much lower training memory.

We developed two boundary detection methods for high-dimensional data, and we applied them to spatial transcriptomics data and LLM activations.

Zero-shot cross-city mobility generation by transferring activity patterns as discrete, city-independent activity prototypes, using only the target city's static map and residential distribution.

We propose a novel framework for evaluating agent from relative feedback that represents the collective preferences of the entire population.

The value signal determines where along the flow guidance works, while the number of guidance determines how well the guidance budget is used.

Time series attribution has no convention-free answer; what does is how much of a declared value function an explanation of a given size carries. We measure that distortion-rate curve, refute locality for low rank, an...

AWDM evaluates whether computer-use agents can keep making progress as user intent evolves without acting beyond what the user has decided.

We propose SAG, a structured retrieval architecture that activates latent hyperedges over event–entity relations with query-time SQL joins.

We introduce CellKAFC, a knowledge-aware flow-correction framework that integrates gene-level biological priors with flow correction.