AI & Machine Learning

VAE tokenizers in Latent Diffusion Models create 'overly compact' manifolds that cause variance collapse, leading to unstable generative sampling.

Introduces 'Mixture of Chapters' to scale Transformer memory to 262K tokens without the quadratic cost of standard attention.

CounterScene endows generative world models with explicit counterfactual reasoning for safety-critical driving evaluation.

A training-free visual token pruning framework for Large Vision-Language Models that preserves geometric structure through subspace reconstruction.

Free Sinewich enables parameter-efficient multi-task learning using frequency-based weight modulation with near-zero overhead.

Reveals that state-of-the-art MLLMs fail to maintain stable spatial representations under simple counterfactual viewpoint changes.

LiFR-Seg achieves high-frame-rate semantic segmentation using low-frame-rate cameras by propagating features through asynchronous event streams.

Proposes multi-cluster memory for test-time adaptation, proving that a single unstructured memory pool is fundamentally insufficient for non-i.i.d. data streams.

ORACLE uses symbolic reasoning engines to verify intermediate reasoning steps in synthetic data generation, moving beyond simple answer-correctness filtering.

AlphaAdj uses a VLM to dynamically adjust Control Barrier Function parameters in real-time for safe and efficient robotic navigation.

BadGraph demonstrates that LLMs can generate universal adversarial attacks that exploit vulnerabilities in both GNN and PLM architectures on graph data.

SPECTRE-G2 is a unified anomaly detector that uses eight complementary signals to detect 'unknown unknown' structural anomalies.

Restores monotonic scaling in LLM tree search by replacing standard MCTS selection with Gumbel sampling and Sequential Halving.

A training-free system for 3D scene reconstruction and editing from sparse RGB images using 3D-aware diffusion models to fill geometric gaps.

Introduces the Neural Zeroth-order Kernel (NZK) to provide a theoretical foundation for training models without backpropagation.

Shows that a simple pruned adaptation module (PAM) outperforms complex SOTA foundation-model-based continual learning methods.

Demonstrates that entropy-based uncertainty is insufficient for safe selective prediction and proposes combining it with correctness probes.

Reframes plasticity loss in Reinforcement Learning as an optimization problem where networks get trapped in local optima of previous tasks.

Introduces Reward Sharpness-Aware Fine-Tuning (RSA-FT) to mitigate reward hacking in diffusion models without retraining reward models.

GIDE enables precise, training-free image editing for discrete Diffusion LLMs by introducing a novel Discrete Noise Inversion mechanism.

Prompt Replay speeds up GRPO training by selectively reusing 'medium difficulty' prompts to maximize learning signal in RL rollouts.

Repurposes a 2B-parameter latent video transformer as a differentiable physics simulator for urban wind flow optimization.

Provides the first empirical evidence of a 'Quality-Homogenization Tradeoff' where AI-assisted writing strips structural diversity from human thinking.

Challenges the widespread assumption that auxiliary dynamics supervision creates useful latent structures for robotics.

Proves that structured retrieval is exponentially more efficient than sequential context scanning for agentic reasoning.

Proposes replacing flat conversation histories with a tree-based architecture to solve 'logical context poisoning.'

Breaks the massive compute barrier for medium-range weather forecasting, training on a single consumer-grade GPU.

Enables multimodal models to self-evolve their reasoning without human labels or external reward models.

Replaces self-attention with Reaction-Diffusion PDEs as the predictive engine for world models.

Identifies architectural 'stream separation' as the key to making linear safety interventions effective.

An autonomous agent loop that optimizes GPU kernels to outperform human-expert and compiler-generated baselines.

Reconceptualizes human-agent interaction as dynamically generated software rather than just chat.

Exposes that LLMs solve complex puzzles via 'reduction' to known patterns rather than true epistemic reasoning.

Introduces AgentHER, a framework that salvages 'failed' agent trajectories by relabeling them as successful demonstrations for alternative goals.

ADARUBRIC generates task-specific evaluation rubrics on the fly, significantly outperforming static rubrics in human correlation and agent training outcomes.

TIDE is a post-training early-exit system that allows individual tokens to skip unnecessary layers, improving throughput by up to 8% with minimal calibration.

PivotRL identifies 'pivot' turns in agent trajectories where actions matter most, enabling compute-efficient reinforcement learning that matches end-to-end RL at 4x lower cost.

Discovers 'silent commitment failure,' where some model architectures produce confident, incorrect outputs with zero detectable warning signals before execution.

Provides a causal explanation for 'embedding collapse' in Transformers, linking it to the concept of semantic shift rather than just text length.

KG-Hopper enables 7B-parameter models to outperform 70B systems on complex Knowledge Graph reasoning by embedding the entire multi-hop process into a single 'thinking' stage.

Introduces Cross-Context Verification (CCV) to detect benchmark contamination, finding that contamination is binary: models either recall solutions perfectly or lack reasoning entirely.

DSPA performs preference alignment at inference time by steering Sparse Autoencoder (SAE) features, bypassing the need for expensive weight-update training.

DRTriton uses large-scale synthetic data and curriculum RL to automatically generate highly optimized Triton kernels, significantly outperforming top-tier LLMs.

Introduces git-inspired primitives to enable truly asynchronous and non-interfering multi-agent software engineering collaboration.

Demonstrates that learning systems can stably converge to incorrect solutions when feedback reliability is unobservable.

Achieves state-of-the-art open-vocabulary segmentation using a training-free, purely geometric projection and propagation method.

Reveals that 'erasing' concepts from video diffusion models only suppresses output rather than removing the underlying representations.

Solves the 'recursive drift' problem in self-improving LLMs by using symbolic verification to gate training data quality.

Introduces a counterfactual framework for precise individual credit assignment in collaborative multi-agent LLM systems.

Provides the first unified theoretical formalism for hierarchical memory systems used by long-context language agents.