MARS is a multimodal system for the CASTLE 2026 challenge that reasons over 4 days of activity, 15 synchronized perspectives, transcripts, and auxiliary modalities (photos, videos, gaze, thermal imagery, heartrate). The approach uses DeepSeek for video summaries and a GPT-5.4 agent to select evidence sources. The system achieved second place on the final leaderboard.
MoleCode is an LLM-native molecular language representing molecules as explicit graphs with typed entities and direct relations, replacing implicit SMILES strings. Training-free, it improves frontier LLMs on molecular reasoning, editing and generation tasks, especially for unfamiliar molecules, topology-sensitive operations and larger structures.
TinySAM 2 compresses SAM 2 for efficient video segmentation. Memory quality management mechanism + joint spatial-temporal token compression. Achieves 90% of SAM 2.1 performance with 7% memory tokens and 3% training data. Reduces parameters, computational load, and deployment costs.
Method to expand LLMs to new languages without costly alignment phase. Converts dense model to Mixture-of-Experts architecture with language-specific experts, then transfers alignment capabilities via post-training delta fusion. Improves performance on new languages while preserving original abilities.
EmoMind decodes affective captions directly from brain fMRI signals. The system first retrieves a neutral scene description from brain-decoded visual features, then rewrites it using a continuous 34-dimensional emotion vector extracted from the same fMRI recording. Evaluated on two independent emotion fMRI datasets, EmoMind outperforms GPT-4 with discrete emotion labels across all validation axes.
Asynchronous RAG framework predicting when and what to retrieve using three components (retrieval predictor, context monitor, query generator). Achieves 43.5% end-to-end latency reduction and 62.4% time-to-first-token improvement while maintaining answer quality.
Empirical study comparing 11 LLMs (GPT-4o, Gemma-3, Phi-4, etc.) across 10 everyday occupational tasks. GPT-4o delivers superior performance but at higher cost; smaller models achieve strong results with better efficiency. Proposes task-aware sufficiency assessments over performance-maximizing benchmarks.
Systematic study of code as AI agent infrastructure. Three layers: harness interface (code connects reasoning, action, environment modeling), mechanisms (planning, memory, tool use, feedback), and multi-agent scaling. Applications: coding assistants, GUI/OS automation, embodied agents, scientific discovery, DevOps.
arXiv paper proposing a formal statistical framework to combine LLM and human evaluations. Uses a doubly robust estimator (missing data approach) to determine optimal sample sizes of human ratings needed for benchmark validation, based on LLM judgment predictability.
Critical audit of SAEBench, the de-facto standard evaluation suite for sparse autoencoders (SAEs). TPP and SCR metrics fail multiple reliability tests and should not be used. Other metrics show higher reseed noise and lower discriminability than assumed. Only sae-probes demonstrates acceptable reliability, but struggles to distinguish architecture variants.
New GPRL (General Preference Reinforcement Learning) method replaces scalar reward models with General Preference Model (GPM) using k skew-symmetric subspaces. Tested on Llama-3-8B-Instruct: 56.51% win rate AlpacaEval 2.0, outperforms SimPO and SPPO on Arena-Hard, MT-Bench, WildBench by preventing single-axis reward hacking.
CheckSupport is an open-source system using locally-deployed LLMs to automate reporting checklist recommendation and completion for scientific manuscripts. Evaluated on peer-reviewed manuscripts, it achieves 90% accuracy for checklist recommendations and 88% for item-level completion, processing each manuscript in 12.5 seconds on CPU-only hardware.
Study of adversarial action removal attacks in self-play reinforcement learning. An attacker selectively masks legal actions from the victim's action set. Experiments on poker (6 to 5,531 states) and two non-poker domains: learned masking causes substantially more damage than random masking, persists across Q-learning/PPO/NFSP/DQN, transfers between agents, and is amplified by self-play.
F³A is a training-free router for visual token pruning in vision-language models. It selects relevant visual tokens via question-conditioned cues without extra LLM forward passes, reducing inference costs while maintaining performance across model scales.
AMARIS introduces persistent evaluation memory to improve rubrics in LLM RL fine-tuning. The system accumulates evaluation diagnostics over time, uses static and dynamic retrieval to contextualize rubric modifications, and adds ~5% time overhead. Experiments show consistent gains across closed and open-ended domains.
Neuro-symbolic study comparing LLMs constrained by deterministic logic scaffolds (Rulemapping) versus unconstrained prompting for hate speech moderation under German Criminal Code (§130). Rulemapping achieves precision 0.80-0.86 and recall 0.82-0.89 versus 0.34-0.49 with unconstrained prompting, eliminating conflation of moral offense with legal illegality.
TierCheck is a three-tier checkpointing system for LLM training. It maintains lightweight differential checkpoints in local/peer memory for fast recovery, asynchronously migrates base checkpoints to remote storage, and ensures global consistency without stalling training. On models up to 40B parameters, it reduces checkpointing time to under 10s.
Longitudinal study of 235M posts from 4M Bluesky users showing algorithmic feed exposure (News, Science, Blacksky) measurably shapes user language: stylistic accommodation, semantic alignment, register formalization. Effects vary by feed; reposting is the strongest predictor of linguistic convergence.
SWIM aligns vision-language representations for fine-grained video object understanding from text prompts alone. Uses mask supervision during training to guide cross-modal attention. Constructs NL-Refer dataset with precise natural language referring expressions. Outperforms visual-prompt-based methods on fine-grained benchmarks.
SonarSweep fuses sonar and vision for underwater 3D reconstruction via plane sweeping. The end-to-end deep learning framework overcomes single-modality limitations by adapting the plane sweep algorithm for cross-modal fusion. Results in simulation and real-world environments, public release of code and first synchronized stereo-camera and sonar dataset.
ECHO is a diffusion-based model for chest X-ray report generation. It uses Direct Conditional Distillation and Response-Asymmetric Diffusion training to enable one-step-per-block inference, achieving 8× inference speedup while improving RaTE (+64.33%) and SemScore (+60.58%) over autoregressive baselines with negligible clinical accuracy loss.
First machine unlearning framework for masked diffusion language models (LLaDA, Dream). MDU minimizes KL divergence from prompt-conditional to prompt-masked unconditional distribution at each masked position, with temperature scaling for privacy-utility trade-off. Code released.
Analysis of 6.8M solutions across 6.7K tasks in 9 domains (materials, structural biology, AI, computational biomedicine, data science, theoretical CS, F1, wheel building). Three universal patterns: heavy-tailed waiting times, sublinear record accumulation, temporal correlation of breakthroughs. Minimal model unifies radical innovation and incremental refinement.
Sherpa.ai introduces a multi-party protocol for privacy-preserving entity alignment in vertical Federated Learning. The method hides intersection membership while enabling exact and typo-tolerant matching, without revealing which samples are shared across parties.
StyleText is a dataset of 28,518 image-mask-prompt triplets for scene text inpainting with style preservation. Automated pipeline combines LLM templating, Flux with KV-cache injection, OCR, polygon mask extraction, and FluxFill augmentation. FluxFill+LoRA baseline substantially improves OCR accuracy while maintaining scene style consistency.
NavOne reformulates vision-language navigation (VLN) as one-step global path planning on top-down maps. The framework directly predicts dense path probabilities via single end-to-end forward pass, achieving 8x speedup over map-based methods and 80x over egocentric approaches. Introduces R2R-TopDown dataset.
CounterRefine is a lightweight repair layer for RAG that treats the first answer as a hypothesis to test. The system issues answer-conditioned expansion queries to retrieve candidate-specific evidence, then applies a deterministically-validated KEEP/REVISE refinement step. On SimpleQA, it improves a matched one-pass RAG baseline by up to 5.8 correct-rate points.
Novel VAE method using spherical shell geometry and cluster-aware constraints to mathematically eliminate posterior collapse. Transforms data to spherical shell, applies K-means, defines feasible region between within-cluster variance and collapse loss. Guarantees 100% collapse prevention without computational overhead or decoder output restrictions.
Rover is a code merge conflict resolution system combining program analysis with LLMs. It introduces Multi-layer Code Property Graph (MtCPG) to capture inter-file dependencies and uses graph connectivity algorithms to create meaningful contexts. Evaluation: Rover outperforms standalone LLMs, MergeGen, and WizardMerge at character, lexical, and semantic levels.
SkyPart, a lightweight head for vision transformers, improves cross-view drone-to-satellite geo-localization by explicitly separating layout and texture via learnable prototypes. At 26.95M parameters, it achieves state-of-the-art on SUES-200, University-1652, and DenseUAV with enhanced robustness under weather corruptions.
Theoretical paper on recovering latent actions and environment dynamics from offline trajectories without action observations. Authors leverage demonstrator diversity (each following distinct policies) to identify latent transition kernels via nonnegative matrix factorization. Identifiability proven under rank and policy diversity conditions.
New perspective on agentic AI systems: code as central infrastructure. This research paper organizes a unified framework around three layers — harness interface (code connecting reasoning and action), mechanisms (planning, memory, feedback), and multi-agent scaling. Applications: coding assistants, GUI/OS automation, embodied agents, scientific discovery.
UCSF-PDGM-VQA is a clinical VQA benchmark for brain tumor MRI interpretation. Dataset of 2,387 QA pairs from 473 glioma studies. Evaluation of 6 VLMs: all fail on multi-sequence 3D MRI, suffer modality collapse and over-reliance on language priors.
HIBCG introduces a theoretically grounded approach to learn sparse coordination graphs in multi-agent RL. Using graph information bottleneck, it determines edge existence and message capacity with formal guarantees on learned topology and differential capacity allocation across agent groups.
An arXiv study challenges the assumption that Mixture of Experts models achieve domain specialization through sparse routing. The COMMITTEEAUDIT framework reveals a domain-invariant "Standing Committee"—a compact coalition of experts capturing most routing mass across domains, layers, and budgets. Peripheral experts handle domain-specific knowledge alone.
Paper introducing CAT (Context-Agent-Tool), a data layer for managing multi-agent teams. Applies ALARA principle (as low as reasonably achievable exposure) to context. Evaluates 22 models (0.6B–35B parameters) on 115 practical tasks via npcsh, a CLI shell. ~2500 executions test file operations, web search, multi-step scripting, tool chaining, and inter-agent delegation.
LLMs spontaneously reconstruct graph topology via sawtooth attention patterns, but this structural understanding is diluted by attention sink. SLASH, a training-free solution, re-amplifies this understanding through plug-and-play attention redistribution, showing significant gains on graph tasks and molecular prediction across diverse LLMs.
DARC is a retraining-free inference-time method that reformulates response selection as distributionally robust optimization under annotator disagreement. It reranks candidates by maximizing KL-robust satisfaction objectives, with deployment controls to cap entropic risk premium without retraining.
Babel is a black-box jailbreak method exploiting a vulnerability in LLM safety alignment: safety relies on sparse attention heads, leaving representational space weakly monitored. Through optimized obfuscation and iterative refinement, Babel achieves 82.67% success on GPT-4o and 78.33% on Claude-3-5-haiku within ~40 queries.
Theoretical paper on generative model evaluation. Authors show standard criteria (marginal matching) don't guarantee covariance structure preservation. They introduce D_Sigma = ||Sigma_P - Sigma_Q||_F to measure dependence fidelity, with formal proofs and validation on Fashion-MNIST VAE, RNA-seq (TCGA-BRCA, n=1111), and Alzheimer's data (n=113).