AI papers — 2026-05-29

Problem: Modern open-world agents (e.g., OpenClaw) introduce broad new safety risks while advanced frontier models lower attack barriers, making current agent alignment frameworks inadequate for real-world deployment.

Method: AgentDoG 1.5 is a taxonomy-guided alignment framework that updates the agent safety taxonomy to cover emergent risks from Codex and OpenClaw execution scenarios. A taxonomy-guided data engine with influence-function purification trains four lightweight AgentDoG 1.5 variants (0.8B, 2B, 4B, and 8B parameters) using only ~1k samples via SFT and RL. A highly efficient Docker-level agentic safety training environment is constructed, and AgentDoG 1.5 is deployed as a training-free online guardrail for real-time safety moderation.

Key results:

• Four model variants trained: 0.8B, 2B, 4B, and 8B parameters
• Trained on only ~1k samples while achieving performance comparable to GPT-5.4
• Docker-level deployment overhead reduced by two orders of magnitude compared to prior environments
• State-of-the-art performance reported across diverse interactive agentic scenarios

Why it matters / caveats: As LLM-based agents become increasingly autonomous, systematic and lightweight safety guardrails are critical for responsible deployment; AgentDoG 1.5's open release of models and datasets lowers the barrier for practitioners. The framework's effectiveness depends on the quality and coverage of the updated taxonomy, and "comparable to GPT-5.4" is an informal benchmark without a standardized safety leaderboard.

Problem: Embodied intelligence is fragmented across specialized models for individual tasks (manipulation, navigation, etc.), resulting in limited generalization across tasks, environments, and robot embodiments.

Method: Qwen-VLA extends Qwen's vision-language stack to continuous action and trajectory generation via a DiT-based action decoder, unifying manipulation, navigation, and trajectory prediction into a single framework. It is trained with large-scale joint pretraining over robotics manipulation trajectories, human egocentric demonstrations, synthetic simulation data, VLN data, and auxiliary vision-language data. Embodiment-aware prompt conditioning uses robot-specific textual descriptions to specify current embodiment and control conventions, enabling cross-embodiment generalization.

Key results:

• LIBERO manipulation benchmark: 97.9%
• Simpler-WidowX: 73.7%
• RoboTwin-Easy / RoboTwin-Hard: 86.1% / 87.2%
• R2R navigation (OSR): 69.0%
• RxR navigation (SR): 59.6%
• Real-world ALOHA OOD average success: 76.9%
• DOMINO zero-shot success: 26.6%

Why it matters / caveats: Qwen-VLA demonstrates that a single unified foundation model can span manipulation and navigation across diverse embodiments, a key step toward general-purpose robots. The 26.6% zero-shot score on DOMINO indicates that cross-task generalization to novel settings remains substantially below in-distribution performance.

Problem: Real-world information needs require access to structurally diverse knowledge sources (unstructured text, relational tables, knowledge graphs, property graphs), but existing retrievers operate over only one source type at a time.

Method: OmniRetrieval is a unified retrieval framework that accepts any natural-language query, identifies the appropriate knowledge source type, and dispatches source-native queries (preserving schemas, ontologies, and compositional operators) to native execution engines rather than collapsing all sources into a shared embedding space. The approach maintains each source's structural affordances (e.g., relational schemas for tables, graph traversal for KGs) rather than homogenizing them.

Key results:

• Evaluated across 13 datasets and 309 distinct knowledge bases spanning text, relational, and graph-structured sources
• Exceeds single-source baselines across all evaluated source types
• Specific numeric margins over baselines not stated in the paper abstract

Why it matters / caveats: A retrieval system that natively handles heterogeneous sources without source-type-specific pipelines simplifies enterprise RAG deployments. Specific quantitative margins over baselines are not reported in the abstract, making it difficult to gauge how large the improvements are in practice.

Problem: Storing and dynamically loading numerous effect-specific LoRAs increases deployment overhead, and cascading them with acceleration modules causes parameter interference leading to concept bleeding and style degradation.

Method: CollectionLoRA is a multi-teacher on-policy distillation framework that consolidates up to 50 different effect LoRAs plus few-step generation into a single LoRA using three components: (i) Probabilistic Dual-Stream Routing randomly switches between data sources during training to enhance generalization; (ii) Asymmetric Orthogonal Prompting isolates individual concepts within the prompt space; (iii) a Coarse-to-Fine Distillation Objective bridges the teacher-student distribution gap across training stages.

Key results:

• Up to 50 distinct effect LoRAs distilled into a single LoRA
• Concept fidelity comparable to or better than independently trained teacher models
• Enables few-step fast generation in the same single LoRA
• Specific quantitative metric scores (e.g., FID, CLIP) not stated in the abstract

Why it matters / caveats: Reducing 50 specialized LoRAs to one dramatically cuts storage and inference overhead for production diffusion-model deployments. Reported results are primarily qualitative comparisons; numerical evaluation scores on standard image quality benchmarks are not cited in publicly available materials.

Problem: High-quality video diffusion models are bidirectional and slow, making it challenging to convert them into real-time interactive camera-controllable world models.

Method: minWM is a full-stack open-source pipeline that converts bidirectional T2V/TI2V video foundation models into camera-controllable few-step autoregressive world models in three stages: (1) fine-tune a bidirectional video diffusion model with camera control; (2) apply Causal Forcing / Causal Forcing++ (AR diffusion training with causal ODE or consistency distillation); (3) compress via asymmetric DMD (Distribution Matching Distillation) for low-latency streaming rollout. The framework is instantiated on Wan2.1-T2V-1.3B (cross-attention condition injection) and HY1.5-TI2V-8B (MMDiT-style architecture).

Key results:

• Framework open-sourced with code, checkpoints, and documentation (356 GitHub stars at submission)
• Supports both 1.3B and 8B base model scales
• Practical ablations on camera trajectory quality, controllability training steps, and minimal batch-size requirements provided
• Specific FPS or FVD quality numbers not stated in the abstract

Why it matters / caveats: Open-sourcing a full pipeline from a bidirectional video diffusion model to a real-time interactive world model lowers the barrier for simulation and robotics research. Concrete real-time latency and video quality benchmarks versus prior interactive world models are not cited in the abstract.

Problem: It is unclear whether state-of-the-art video diffusion models genuinely understand physical causality or merely overfit to statistical temporal patterns in training data.

Method: YoCausal is a two-level benchmark inspired by the Violation of Expectation (VoE) paradigm from cognitive science, using temporally reversed real-world videos as cost-free natural counterfactual samples. Level 1 introduces the Reverse Surprise Index (RSI), which quantifies arrow-of-time perception via denoising loss on forward vs. reversed videos. Level 2 introduces the Causality Cognition Index (CCI), which uses a VLM to disentangle genuine causal reasoning from temporal bias. Thirteen state-of-the-art VDMs were evaluated.

Key results:

• Evaluated 13 state-of-the-art video diffusion models
• Key finding: perceiving the arrow of time (high RSI) does not imply causal understanding (high CCI)
• Significant gap between all evaluated models and human-level causal cognition
• Specific per-model RSI/CCI scores not stated in the abstract

Why it matters / caveats: This work provides a principled, cognitively grounded diagnostic separating temporal pattern matching from genuine causal reasoning, which is essential for evaluating video models as world models. The benchmark currently relies on temporally reversed real-world videos, and the extent to which this proxy fully captures causal understanding remains an open question.

Problem: Current image generation agents rely on iterative prompt rewriting cycles without direct structural control over the canvas, limiting spatial precision and interpretability.

Method: GenClaw introduces a three-stage code-driven agentic image generation paradigm: (1) Conceptualization — the agent builds domain knowledge via search and reasoning; (2) Sketching — executable visual sketches are rendered via code (SVG, HTML, Three.js), providing a controllable intermediate canvas; (3) Coloring — a diffusion-based image generation model adds textures, materials, and photorealism to the structural sketch.

Key results:

• Code-driven intermediate representation enables direct spatial control not possible with prompt-only pipelines
• Demonstrated across SVG, HTML, and Three.js rendering modalities
• Specific quantitative benchmark scores (e.g., GenEval, T2I-CompBench) not stated in the abstract

Why it matters / caveats: Using executable code as a controllable intermediate canvas is a novel abstraction that could substantially improve spatial layout fidelity in complex compositional scenes. Without reported standard benchmark numbers it is difficult to quantify the improvement over prompt-rewriting baselines.

Problem: VLMs achieve strong scores on spatial reasoning benchmarks, but it is unclear whether this reflects genuine structured 3D understanding or exploitation of statistical shortcuts in natural image correlations.

Method: This paper introduces a representation-level analysis framework using minimal contrastive image pairs to measure how spatial axes (vertical position vs. depth/distance) are organized and disentangled within VLM embeddings. It introduces SpatialTunnel, a synthetic benchmark removing natural image correlations to isolate model-intrinsic spatial biases. Experiments span multiple model families to characterize the vertical-distance entanglement phenomenon.

Key results:

• Consistent vertical-distance entanglement identified: VLMs conflate vertical image position with depth, mirroring photographic perspective bias
• Bias worsens with data scaling despite improved benchmark scores, confirming it is model-intrinsic
• SpatialTunnel exposes the entanglement in models that appear strong on natural-image spatial benchmarks
• Specific numeric entanglement scores per model not stated in the abstract

Why it matters / caveats: Identifying that benchmark gains can mask a persistent perspective-shortcut bias has direct implications for robotics, navigation, and embodied AI applications. The study diagnoses the problem but does not yet propose a mitigation.

Problem: Video-LLM deployment is bottlenecked by massive visual token counts, and the vision encoder itself contributes a disproportionate share of time-to-first-token (TTFT) that prior late-stage compression methods leave unaddressed.

Method: EarlyTom is a training-free token compression framework that performs early-stage visual token compression inside the vision encoder rather than after it, targeting the encoding bottleneck directly. It additionally introduces a decoupled spatial token selection strategy that selects informative spatial tokens independently of the temporal compression path, improving overall compression effectiveness without fine-tuning.

Key results:

• TTFT reduced by up to 2.65× on LLaVA-OneVision-7B on a single NVIDIA A100 GPU
• FLOPs reduced by up to 61%
• Accuracy maintained comparable to the full-token baseline
• Accepted to CVPR 2026

Why it matters / caveats: Cutting TTFT by 2.65× and FLOPs by 61% without training or accuracy loss makes EarlyTom immediately deployable in production Video-LLM systems. Results are reported on a single model (LLaVA-OneVision-7B); generalization to other vision encoder architectures is not yet characterized.

Problem: LoRA is widely used for continual knowledge updates in LLMs, but the quantitative capacity limits and underlying dynamics of exact parametric memory storage in LoRA are unexplored.

Method: This work uses LoRA as a controlled memory capacity probe in latent space to derive the Parametric Memory Law — a robust power law linking loss reduction (ΔL) to LoRA rank, sequence length, and data quantity. Token-level analysis reveals a phase transition at prediction probability p > 0.5 as a sufficient condition for verbatim recall under greedy decoding. MemFT is introduced as a threshold-guided optimization strategy that dynamically redistributes training budget toward sub-threshold tokens to improve memory efficiency.

Key results:

• Power law ΔL ∝ (effective parameters) × (sequence length) empirically validated
• Phase transition at p > 0.5 per-token probability establishes a clear threshold for verbatim recall
• MemFT improves memory fidelity and efficiency over standard LoRA training
• Specific improvement magnitudes for MemFT not yet stated (work in progress)

Why it matters / caveats: A quantitative memory law for LoRA provides principled guidance for rank selection and training budget allocation, practically valuable for continual learning. The paper is marked as ongoing work, so results may change before final publication.

Problem: Joint audio-video generation requires temporally synchronized and semantically coherent content, but existing dual-tower designs weaken fine-grained co-evolution while unified tri-modal designs entangle semantic conditioning with low-level synchronization.

Method: NAVA (Native Audio-Visual Alignment) uses a context-conditioned Align-then-Fuse MMDiT architecture: it first establishes audio-video correspondence in a dedicated modality-aware interaction space, then transitions to modality-shared joint denoising conditioned on external context. Timbre-in-Context Conditioning associates reference timbre cues with corresponding speech spans for controllable speech timbre. The model totals 6.3B parameters.

Key results:

• 6.3B parameter model evaluated on Verse-Bench and Seed-TTS benchmarks
• Superior video quality, precise audio-visual synchronization, competitive audio quality, and stronger reference-timbre controllability vs. prior methods
• Specific numeric scores on Verse-Bench and Seed-TTS not stated in the abstract

Why it matters / caveats: Decoupling the alignment step from the conditioning step in a unified MMDiT is a principled architectural improvement for joint AV generation. Exact metric values on standard benchmarks are absent from the abstract, making head-to-head comparison difficult.

Problem: Activation-based LLM control methods rely on fixed steering directions or task-specific modules, limiting flexibility across diverse behavioral targets.

Method: UniSteer learns a single conditional distribution over residual-stream activations from natural-language conditions via text-guided activation flow matching, training a universal conditional velocity field in activation space. At inference, it performs flow inversion by partially transporting the source activation toward a latent state and regenerating it under the target textual condition before re-injecting it into the frozen LLM. The same conditional model also supports activation-space classification by selecting the textual label with the lowest reconstruction energy.

Key results:

• Single model covers five steering tasks: behavioral control, truthfulness steering, fine-grained concept steering, multi-constraint instruction following, and activation-space classification
• Outperforms fixed-direction and task-specific activation intervention baselines across three target LLMs
• Specific accuracy/success-rate numbers not stated in the abstract

Why it matters / caveats: A single text-conditioned flow model that steers arbitrary LLM behaviors without retraining is a flexible and deployable alternative to per-task steering vectors. Quantitative comparisons to baselines such as CAA and ITI are not enumerated in the abstract.

Problem: Data contamination in RL post-training undermines generalization and evaluation reliability of reasoning LLMs, and output-level detection signals (likelihood, entropy) are unreliable for RL-trained models.

Method: LaRA is a layer-wise representation analysis framework detecting contamination via three geometric metrics measured under controlled perturbations: perturbation sensitivity, directional collapse, and local representation rigidity. Contamination produces progressive geometric deviations across layers in all three metrics, providing a more reliable signal than output-level methods.

Key results:

• LaRA outperforms existing output-level contamination detection baselines on RL-trained reasoning models
• Contamination signature: amplified perturbation sensitivity + stronger directional collapse + enhanced local rigidity across layers
• Specific AUC or detection accuracy figures not stated (work in progress)

Why it matters / caveats: As RL post-training becomes standard for reasoning models, reliable contamination detection at the representation level is essential for valid benchmarking. The paper is marked as work in progress, so results remain preliminary.

Problem: VLMs exhibit carrier sensitivity — replacing a textual question with its rendered-image equivalent causes dramatic performance degradation — because training data assigns asymmetric roles to text (linguistic query) and images (visual reference).

Method: LoMo is a data curation paradigm that provides supervision for cross-modal representational invariance during SFT. It reformulates single-modality prompts into interleaved multimodal sequences by dynamically selecting text spans and recasting them as rendered images, preserving semantics across text-visual-text carriers. The method is lightweight and architecture-agnostic, requiring no changes to model architecture.

Key results:

• +2.67 average points over standard SFT on LLaVA-OneVision-1.5-8B across 13 multimodal benchmarks
• +2.82 average points over standard SFT on Qwen3.5-9B across 13 multimodal benchmarks
• Consistent improvements across all 13 benchmarks tested

Why it matters / caveats: A simple data curation fix yielding consistent 2–3 point gains across 13 benchmarks without architectural changes is immediately applicable to any SFT pipeline. The approach addresses cross-modal invariance at training time but does not resolve deeper asymmetries in pre-training data distributions.

Problem: Skill-based RL methods force an all-or-nothing choice between full skill externalization (high context overhead) and full skill internalization (overfitting and knowledge conflicts), limiting out-of-distribution generalization.

Method: Skill0.5 is an agentic RL framework that differentiates skill treatments based on type and task difficulty. A dynamic difficulty-aware router streams tasks into mastery tiers: for hard tasks, general skills are internalized via privileged distillation to build a cognitive foundation; for easy tasks, diagnostic probing penalizes shortcuts and enforces task-specific skill utilization from external context.

Key results:

• Outperforms both memory-based and skill-based RL baselines on ALFWorld and WebShop
• Gains on both in-distribution and out-of-distribution evaluation scenarios
• Specific success-rate numbers not stated in the abstract

Why it matters / caveats: Skill0.5 provides a principled middle ground between the two extremes of skill management in RL agents, with demonstrated OOD gains. Concrete numeric margins versus baselines are not cited in the abstract.

Problem: Long-horizon interactions require LLMs to manage when to update, preserve, or ignore beliefs based on evidence, but current models lack principled mechanisms for Contextual Belief Management (CBM).

Method: The paper formalizes CBM and introduces BeliefTrack, a closed-world benchmark with two task types — Rule Discovery and Circuit Diagnosis — using a finite belief space and symbolic verifiers for exact turn-level evaluation. It identifies three failure modes: Failed Stay, Failed Update, and Failed Isolation. Two interventions are evaluated: RL training with belief-state rewards, and representation-level steering on internal model states.

Key results:

• RL with belief-state rewards reduces failure rates by 70.9% on average
• Representation-level steering reduces failure rates by 46.1% across two tasks
• Vanilla models exhibit severe CBM failures; explicit belief-tracking prompts provide only marginal improvement

Why it matters / caveats: Reducing CBM failures by ~71% with RL rewards is a strong result for improving reliability in multi-turn reasoning and dialogue agents. The benchmark uses closed-world symbolic tasks; how well findings transfer to open-ended real-world belief updating is not yet demonstrated.

Problem: Dense retrievers assign relevance scores through opaque high-dimensional embeddings, making it difficult to explain why a retriever ranks a document highly for a given query.

Method: Xetrieval is an embedding-level mechanistic interpretability framework with two components: (1) a lightweight reasoning internalizer that approximates Chain-of-Thought reasoning directly in embedding space with a single forward pass, enriching embeddings without expensive autoregressive generation; (2) a sparse feature decomposition module that decomposes the reasoning-enhanced embeddings into sparse, human-interpretable features each associated with a coherent natural language description. Feature-level explanations are produced by aggregating sparse feature overlaps across multiple document-side views.

Key results:

• Uncovers coherent interpretable features across diverse retrievers and benchmarks
• Yields stronger pair-level intervention effects than baseline explanation methods
• Supports task-level feature steering for retrieval control
• Specific retrieval accuracy or explanation quality scores not stated in the abstract

Why it matters / caveats: Mechanistic interpretability for dense retrieval addresses a real gap in enterprise search and RAG systems where understanding retrieval decisions is critical for auditing and debugging. Numerical comparisons to prior explanation methods are not cited in the abstract.