ArXiv AI Papers
Latest artificial intelligence and machine learning research papers from ArXiv.
Showing 50 of 183 items
Program-as-Weights: A Programming Paradigm for Fuzzy Functions
You describe a fuzzy task in plain language and a 4B "compiler" turns it into a tiny neural add‑on for a small model. Those add‑ons match a 32B model on the tasks while using about fifty times less memory and running comfortably on a laptop. Use this to turn one‑off prompts into cheap, reusable local tools for routing, triage, and other judgment calls.
Multi-Resolution Flow Matching: Training-Free Diffusion Acceleration via Staged Sampling
MrFlow speeds up image generators by first making a rough low‑res image, then upscaling in pixel space, adding light noise, and doing one last high‑res pass. It reports 10–25× faster full pipelines on FLUX and Qwen‑Image with almost no quality loss. If you ship image models, this is a blueprint for big speedups without retraining.
Optimizing Visual Generative Models via Distribution-wise Rewards
Most image RL setups reward each sample alone and end up with mode collapse or weird artifacts. This work instead rewards the whole batch distribution, pushing models toward diverse, high‑quality images. If you tune image generators, this is a recipe for reducing reward hacking while keeping variety.
AgenticDataBench: A Comprehensive Benchmark for Data Agents
AgenticDataBench collects 344 real and synthetic data‑science tasks across 15 domains with 433 labeled "skills" like joins, cleaning, and feature work. It lets you see not just whether a data agent solves a task, but which skills it uses or misses. Use this if you are building spreadsheet or SQL copilots and want fine‑grained eval instead of single scores.
PACE: A Proxy for Agentic Capability Evaluation
Running full agent benchmarks like SWE‑Bench or GAIA can cost thousands of dollars per run. PACE learns a tiny set of cheap test items whose scores predict those heavy agent scores with strong rank correlation. Use this to triage models and routes before paying for full agent evaluations.
AgenticSTS: A Bounded-Memory Testbed for Long-Horizon LLM Agents
The authors build a Slay the Spire 2 benchmark where an agent must make hundreds of connected decisions with tight memory limits. They swap the usual endlessly growing context for a "bounded contract" that pulls only typed, relevant records into each step. Use this testbed to stress‑test your agent memory designs instead of guessing from toy games.
IsoSci: A Benchmark of Isomorphic Cross-Domain Science Problems for Evaluating Reasoning versus Knowledge Retrieval in LLMs
IsoSci pairs science questions that share the same logical structure but live in different domains. It shows that over 90% of "reasoning mode" gains really come from extra knowledge, not better logic. Use this benchmark before claiming your model "reasons" better just because chain‑of‑thought scores went up.
CausalMix: Data Mixture as Causal Inference for Language Model Training
CausalMix treats data mixing for LLM training as a causal problem instead of a guessing game. It estimates how different data buckets change downstream scores, then picks mixtures based on those estimated effects. If you are curating giant training pools, this gives you a more principled way to choose what to oversample.
Prism Transformer: Progressive Head Schedules for Hierarchical Attention Processing
Keeps total parameters fixed but redistributes attention heads: few wide heads early, many narrow heads late. This simple change consistently beats standard layouts on language benchmarks. If you're running expensive training runs, this is a cheap architectural tweak to test. ([arxiv.org](https://arxiv.org/list/cs.LG/new))
ToE: A Hierarchical and Explainable Claim Verification Framework with Dynamic Multi-source Evidence Retrieval and Aggregation
Builds a "Tree of Evidence" system: one agent expands a claim into subclaims and retrieves evidence, another scores evidence, then a tree algorithm aggregates everything. It beats baselines especially on poisoned or adversarial inputs. If you rely on web search or RAG, the architecture is worth copying. ([arxiv.org](https://arxiv.org/list/cs.AI/new))
Position: The Term "Machine Unlearning" Is Overused in LLMs
Argues that "machine unlearning" should mean only one strict thing: making a model behave as if a specific dataset was never seen. Most current LLM "unlearning" work does not meet that bar. If you work on data deletion or policy compliance, this sharpens the vocabulary and bar for evidence. ([arxiv.org](https://arxiv.org/list/cs.CL/new))
Not All Relations Rotate Alike: Transformation-Aware Decoupling for Viewpoint-Robust 3D Scene Graph Generation
Separates 3D scene relations into ones that should rotate with the camera and ones that should stay fixed. This decoupling makes scene graphs much more stable under viewpoint changes. If you're building embodied agents or 3D perception, this is a practical robustness trick. ([arxiv.org](https://arxiv.org/list/cs.CV/new))
The Curse of Multiple Mediators: Hidden Interaction Effects in Activation Patching
Shows that standard activation patching mixes the effect of a unit with how it interacts with many others, not just its direct influence. These interaction terms can hide or fake "important" neurons. If you run mechanistic interpretability experiments, this paper says: treat patching results with more skepticism. ([arxiv.org](https://arxiv.org/list/cs.LG/new))
PairSAE: Mechanistic Interpretability from Pair Representations in Protein Co-Folding
Adapts sparse autoencoders to the "pair" tensors in protein co-folding models by compressing them into token-level features first. Recovers features aligned with biological structure and binding signals. If you care about interpretability beyond plain transformers, this is a useful template. ([arxiv.org](https://arxiv.org/list/cs.LG/new))
Multi-Objective Molecular Generation with Frequency-Controlled Evolutionary Dynamics
Represents molecules in a Fourier basis and evolves them with a multi-objective evolutionary algorithm. Separates coarse scaffold changes from fine local tweaks in a clean way. If you care about drug discovery or interpretable molecule search, this offers a training-free alternative to diffusion models. ([arxiv.org](https://arxiv.org/list/cs.NE/new))
Developmental Approach Reveals the Statistical Learning of Neural Language Models
Trains transformers on a synthetic grammar and snapshots them over time to see how they internalize patterns. Finds that they first pick up global statistics, then refine local rules. If you care about curriculum design or "how models learn," this gives concrete evidence, not just anecdotes. ([arxiv.org](https://arxiv.org/list/cs.CL/new))
MER-R1: Multimodal Emotion Reasoning via Slow-Fast Thinking Synergy
Shows that explicit chain-of-thought often hurts emotion recognition compared with quick answers from the same model. Introduces a training setup that combines "fast" and "slow" heads so they work together instead of fighting. If you build emotional or social agents, this is a blueprint for more stable behavior. ([arxiv.org](https://arxiv.org/list/cs.AI/new))
Internalizing the Future: A Unified Agentic Training Paradigm for World Model Planning
Trains LLM agents to "imagine" future states and score plans, not just react step by step. They use a three-stage pipeline to inject forecasting, format it, then harden it with reinforcement learning. If you build agents that plan, this is a concrete recipe for giving them an internal world model. ([arxiv.org](https://arxiv.org/list/cs.AI/new))
Towards Reliable and Robust LLM Planning: Symbolic Feedback-Driven Iterative Self-Refinement Framework
Wraps an LLM planner with a symbolic checker that spots broken plans and feeds back fixes in plain language. Adds a goal recognizer to catch impossible targets. If your agents keep hallucinating steps, this paper gives a concrete control loop to bolt on. ([arxiv.org](https://arxiv.org/list/cs.AI/new))
AI-Model Network: Concept, Current State and Future
Proposes an "AI-ModelNet" that connects many smaller models into a network that can share skills, route requests, and collaborate like the internet of AIs. Useful if you're thinking beyond one giant model and toward fleets of specialized models that talk to each other. ([arxiv.org](https://arxiv.org/list/cs.AI/new))
Formalizing Latent Thoughts: Four Axioms of Thought Representation in LLMs
Defines four testable criteria for what a "good" internal thought representation should satisfy, separate from task scores. Finds that current models systematically fail these tests. If you probe activations or build latent-thought pipelines, this gives a sharper evaluation target. ([arxiv.org](https://arxiv.org/list/cs.CL/new))
DysLexLens: A Low-Resource LLM Framework for Analysing Dyslexic Learners Insights from Online Forums
Builds a full pipeline that cleans messy Reddit posts, builds a knowledge graph, and uses an LLM to study how dyslexic learners use AI tools. Treat it as a pattern for responsible analysis of niche online communities when data and labels are scarce. ([arxiv.org](https://arxiv.org/list/cs.AI/new))
Unified Zero-Shot Time Series Forecasting: A Darts Foundation
Plugs several time-series foundation models into the Darts library under one standard interface. Lets teams swap in Chronos-2, TimesFM 2.5, TiRex, and PatchTST-FM with a name change. If you forecast anything, you can now A/B strong zero-shot models without glue code. ([arxiv.org](https://arxiv.org/list/cs.LG/new))
HumanScale: Egocentric Human Video Can Outperform Real-Robot Data for Embodied Pretraining
Shows that carefully processed first-person human videos can beat expensive teleoperated robot data for pretraining embodied models. If robot data collection is killing your budget, start experimenting with egocentric video pipelines.
LedgerAgent: Structured State for Policy-Adherent Tool-Calling Agents
Adds a separate “ledger” of task state so customer-service agents stop forgetting rules and context between tool calls. If you deploy tool-calling agents in regulated domains, this is a ready-made design pattern.
Rethinking Shrinkage Bias in LLM FP4 Pretraining: Geometric Origin, Systemic Impact, and UFP4 Recipe
Argues today’s popular 4‑bit number format systematically underestimates values and destabilizes large-model training. Proposes a uniform 4‑bit recipe that stays closer to BF16 while saving memory and compute.
DragMesh-2: Physically Plausible Dexterous Hand-Object Interaction with Articulated Objects
Introduces a contact-driven control scheme so robot hands move articulated objects by managing physical contact, not replayed trajectories. Useful if you’re training hands for doors, drawers, and other jointed objects without rich tactile sensors.
Playful Agentic Robot Learning
Robots learn skills by self-directed play, then reuse those skills to solve real tasks with little extra training. If you run embodied agents, steal this curriculum idea to cut hand-designed task scripts.
Context-Aware RL for Agentic and Multimodal LLMs
Trains models to pick the right evidence, not just the right final answer, using contrastive context pairs. If your agents drown in long traces or busy images, this offers a concrete RL recipe.
DF3DV-1K: A Large-Scale Dataset and Benchmark for Distractor-Free Novel View Synthesis
Releases a 1,048-scene dataset tuned for “clean” novel view synthesis without background clutter. If you train 3D or radiance-field models, this is a ready-made stress test and fine-tuning set.
Beyond Static Leaderboards: Predictive Validity for the Evaluation of LLM Agents
Shows that single-number leaderboards for agent benchmarks often fail to predict how agents behave in new settings. If you run evals, you should copy their predictive-validity approach, not just chase top scores.
S-Agent: Spatial Tool-Use Elicits Reasoning for Spatial Intelligence
Builds a spatial “tool-using” agent that lets vision-language models maintain a live 3D picture of the world over time. Use this if your agents constantly forget what’s where in multi-view or video settings.
ENPIRE: Agentic Robot Policy Self-Improvement in the Real World
Wraps real robots in a closed-loop system where coding agents iteratively reset scenes, run policies, check results, and improve code. If you’re serious about autonomous robot labs, this is basically a blueprint.
SpatialClaw: Rethinking Action Interface for Agentic Spatial Reasoning
The authors build SpatialClaw, a code-driven agent that uses a stateful Python kernel plus vision tools to solve 3D and 4D spatial puzzles. It beats prior spatial agents across 20 benchmarks and six vision-language backbones, showing that the action interface design can unlock much stronger spatial reasoning.
HYDRA-X: Native Unified Multimodal Models with Holistic Visual Tokenizers
HYDRA-X unifies image and video tokenization inside a single vision transformer, then uses that tokenizer to drive a 7B multimodal model. It hits strong scores on both understanding and generation while proposing a better way to edit content by operating in the tokenizer’s latent space instead of inside the language model.
Objects Before Words: Object-First Inductive Biases for Grounding Language in Child-View Video
BabyMind learns word meanings from egocentric child video by first building object tracks, then aligning utterances to bags of objects. This object-first bias beats prior contrastive baselines on SAYCam and out-of-distribution tests, suggesting more human-like pathways for grounded language learning.
HairPort: In-context 3D-aware Hair Import and Transfer for Images
HairPort separates hair removal from transfer using a LoRA-adapted image model that first creates realistic bald faces. It then reconstructs the source hairstyle in 3D and re-renders it from the target viewpoint, enabling robust hair transfer under large pose and scale changes.
Visual Place Recognition in Forests with Depth-Aware Distillation
The authors improve visual place recognition in forests by distilling depth cues into a DINOv2 descriptor. Their lightweight model keeps the original embedding space but becomes more robust to lighting and seasonal changes on a recent forest benchmark.
Language-Guided Abstraction for Visual Reasoning
This paper tackles ARC-style abstract reasoning by adding a language branch that refines human-written task descriptions with a large model. The system uses those cleaned descriptions to guide a compact 18M-parameter visual model, outperforming prior ARC methods while keeping the final model lightweight.
GRIP: Feedback-Guided Prompt Retrieval for Large Multimodal Models
GRIP trains a retriever to pick in-context examples that actually improve a multimodal model’s answers, instead of just being visually similar. The retriever learns from model feedback and then transfers across different vision-language models, boosting accuracy on classification, captioning and VQA.
Revisiting Vehicle Color Recognition in Long-Tailed Surveillance Scenarios
This work studies vehicle color recognition when rare colors are heavily underrepresented in surveillance data. It mixes synthetic images from diffusion and commercial image editors with tailored training tricks to lift macro accuracy by over eight points on a real-world dataset.
DIMOS: Disentangling Instance-level Moving Object Segmentation
DIMOS fuses event camera streams with RGB frames and explicitly separates motion from appearance features in both modalities. That dual disentangling plus cross-modal alignment sharply improves moving-object segmentation, especially for small or fast targets in low light.
EvoArena: Tracking Memory Evolution for Robust LLM Agents in Dynamic Environments
EvoArena builds a three-domain benchmark where agents must keep working as terminals, codebases, and user preferences change over time. The companion EvoMem memory system logs non-additive updates as patches, giving measurable gains on both step-level and chain-level success in evolving tasks.
Reasoning for Mobile User Experience with Multimodal LLMs: Task, Benchmark, and Approach
The authors introduce a benchmark where multimodal models must judge mobile app UX directly from full UI screenshots. They also propose a baseline model that reasons over layout, text and visual cues, highlighting how current systems miss many usability issues humans spot instantly.
Learning to Reason by Analogy via Retrieval-Augmented Reinforcement Fine-Tuning
This paper trains a retriever to select past reasoning traces that actually help solve a new problem, then uses those traces during reinforcement-based customization. On hard math benchmarks like AIME, their analogy-aware method beats standard reinforcement setups by several points, showing that reasoning-aware retrieval is a real lever.
Video2LoRA: Parametric Video Internalization for Vision-Language Models
Video2LoRA watches a video once, then predicts LoRA weights that let a frozen vision-language model handle that video efficiently. You keep quality while cutting the visual token budget and time-to-first-answer. This is very relevant if video-heavy agents are choking your GPU bill.
MAOAM: Unified Object and Material Selection with Vision-Language Models
MAOAM uses a single vision-language model to select both objects and materials in images from text or clicks. That enables more precise, flexible photo and video editing than today’s simple masks. If you build creative tools, this points to where AI-powered selection is heading.
AffordanceVLA: A Vision-Language-Action Model Empowering Action Generation through Affordance-Aware Understanding
AffordanceVLA bakes an understanding of "what you can do with this object" directly into a robot’s vision-language-action model. It predicts affordances first, then uses them to guide actions. This is a concrete recipe if you want robots that manipulate objects reliably instead of just describing them.
Rethinking Continual Experience Internalization for Self-Evolving LLM Agents
This work studies how agents should turn past trajectories into lasting skills without collapsing or forgetting. It compares different ways to chunk experience, inject it back into the model, and train on it off-policy. If you’re building agents that learn over weeks or months, the design choices here matter more than raw model size.
The Road Ahead in Autonomous Driving: The KITScenes Multimodal Dataset
KITScenes offers a rich European driving dataset with high-res cameras, long-range lidar, 4D radar, and dense HD maps. It fixes gaps in current driving sets around sensor quality and map coverage. If you’re training driving or world models, this is a new high-end reference point.