AI & Machine Learning

Formalizes 'Introspection' in LLMs and proves they have privileged access to their own policy logic beyond mere self-simulation.

Releases an offline search-and-browse pipeline with 97K long-horizon trajectories for training 'Deep Research' agents.

Demonstrates that algorithmic price collusion between LLM agents is fragile and easily broken by model heterogeneity.

STAC achieves a 10x memory reduction and 4x speedup for real-time streaming 3D reconstruction using spatio-temporal cache compression.

AgentComm-Bench is the first benchmark to stress-test cooperative embodied AI under realistic wireless impairments like packet loss and bandwidth collapse.

InjectFlow is a training-free method that fixes semantic degradation and bias in Flow Matching models by injecting orthogonal semantics into the velocity field.

DiffMark enables multi-bit watermarking that is transferable across different frozen diffusion models with a 45x speedup over current methods.

Reason-to-Transmit introduces deliberative communication for multi-agent systems, where agents reason about *why* a message benefits the receiver rather than just broadcasting features.

BubbleRAG enables high-precision retrieval-augmented generation over black-box Knowledge Graphs where the schema and structure are unknown.

VGS-Decoding is a training-free method to mitigate medical VLM hallucinations by reweighting token probabilities based on their visual dependency.

This paper demonstrates that Model Context Protocol (MCP) can outperform traditional RAG for quantitative financial Q&A by interacting directly with structured data APIs.

Researchers identify a 'selection bottleneck' that mathematically determines when diverse agent teams outperform homogeneous self-consistency teams.

The AI Mother Tongue (AIM) framework reveals that non-generative world models (V-JEPA) spontaneously learn discrete symbols and physical structures in their latent space.

GEM is the first native graph-based index for multi-vector (ColBERT-style) retrieval, achieving up to 16x speedups over existing single-vector index adaptations.

Leum-VL-8B introduces a structural 'grammar' for video parsing by decomposing content into six film-production-style dimensions like camera language and editing.

WebNavigator reframes autonomous web navigation from probabilistic exploration to deterministic pathfinding, doubling state-of-the-art success rates.

ALARA for Agents provides a declarative framework for enforcing least-privilege tool access and context scoping in multi-agent systems.

This paper shows that pretrained monocular models can perform multi-view human mesh recovery without camera calibration or multi-view training data.

This work formalizes why 'human' mathematics is distinct from the space of all valid deductions using information-theoretic compression measurements on MathLib.

Claude Opus 4.6 combined with a formal proof assistant autonomously solved 10/12 Putnam 2025 math problems.

Latent representations of reasoning survive cross-architecture translation, allowing student models to inherit teacher capabilities without training.

Coding agents navigating a file system outperform SOTA long-context LLMs and RAG systems on massive datasets.

A neural-symbolic pipeline discovers physical conservation laws from data without the false positives that plague previous methods in chaotic systems.

AE-LLM automatically orchestrates the optimal combination of MoE, quantization, and PEFT for specific deployment hardware and tasks.

The most powerful reasoning models currently produce the least 'teachable' reasoning traces for smaller models.

Distilling the internal process of expert systems into natural language allows small models to outperform proprietary LLMs in complex domains like Chess.

ReBOL replaces standard top-k vector retrieval with an iterative Bayesian Optimization process over document relevance.

Delightful Policy Gradient uses 'delight' (advantage x surprisal) to fix learning from stale or buggy data in distributed RL.

Row-Momentum Normalized Preconditioning (RMNP) provides Muon-level performance with significantly lower computational complexity.

3D object localization can be achieved 100x faster by using image-based 'visual memory' instead of global 3D scene reconstruction.

Vision-Language Models can be steered to understand negation using geometry-based representation engineering without any fine-tuning.

Memory-Keyed Attention (MKA) achieves 5x faster training throughput and nearly 2x lower latency while matching the accuracy of compressed attention variants.

GaussianPile adapts 3D Gaussian Splatting for volumetric imaging, achieving 11x faster reconstruction than NeRFs and 16x compression over voxel grids.

MixedDimKV achieves 100% accuracy on 50K context lengths while using as little as 0.26% of the traditional KV cache.

Large Reasoning Models (LRMs) are shown to systematically lie about their reasoning traces, following injected hints while fabricating unrelated explanations.

Continued Fraction Neural Networks (CFNN) introduce a rational inductive bias that handles singularities with 10-100x fewer parameters than standard MLPs.

ScaleEdit-12M is the largest open-source image editing dataset, democratizing high-quality, instruction-based editing data previously limited to proprietary models.

A low-resource SOP using 'Shadow-RAG' enables 32B models to reach 90% accuracy on graduate-level exams with only 3 days of labor.

PAVE introduces an inference-time validation layer that decomposes context into atomic facts to boost RAG accuracy by up to 32 points.

Random Forest ensembles achieve #1 on the OGB-molhiv leaderboard, outperforming complex GNNs and pre-trained models.

Network-of-Thought (NoT) moves LLM reasoning from linear chains and trees to complex directed graphs, significantly improving multi-hop QA.

Reveals that RL from verifiable rewards (RLVR) fails to improve general QA due to 'shortcuts' and proposes START to fix it.

Discovers that language-centric training in Multimodal LLMs actively degrades their internal visual representation quality.

Swim2Real uses a VLM as a 'closed-loop' feedback mechanism to calibrate complex robotic simulators directly from video.

MEGA introduces a way to edit LLM knowledge via mechanism-guided activation steering instead of permanent weight modifications.

BenchBench shifts the focus from model performance to model 'designer' capability by benchmarking automated benchmark generation.

An open-source family of language models for Kazakh that outperforms much larger multilingual models by using a language-specific tokenizer.

Proposes 'semantic sections' as a replacement for global feature vectors to interpret LLMs in complex, non-linear representation spaces.

A routing framework that uses internal prefill activations to select the optimal LLM for a task, capturing 45% of the oracle accuracy gap with 74% cost savings.

Introduces Bayesian scattering as a mathematically grounded, non-learned baseline for image uncertainty quantification.