AI papers — 2026-05-14

Problem: Existing multimodal tabular learning benchmarks fail to test scenarios where unstructured modalities (text, images) require task-specific representation tuning, as they rely on frozen pre-trained embeddings that lose fine-grained, task-relevant details.

Method: The paper introduces a curation pipeline with four experimental conditions—Unimodal Structured, Unimodal Unstructured, Joint Frozen, and Joint TAR (Target-Aware Representations)—where TAR fine-tunes the last 3 transformer layers of encoders (e5-v2-small for text, DINO-v3-small for images) using LoRA before feeding representations to tabular learners. Datasets qualify for the benchmark only if Joint Frozen outperforms both unimodal baselines and Joint TAR improves over Joint Frozen, verified across at least 3 of 5 tabular learners (LightGBM, CatBoost, TabM, TabPFNv2, TabPFN-2.5). Embeddings are compressed to 30 PCA components for efficiency.

Key results:

• Benchmark comprises 40 datasets (20 image-tabular, 20 text-tabular), balanced between classification and regression, with 400–114,000 rows per dataset
• Only 41% of 56 text-tabular candidates and 31% of image-tabular source datasets passed both mandatory criteria
• TAR consistently improves over frozen embeddings across all tabular learners; TAR with small models outperforms frozen large embeddings
• Attention maps confirm TAR shifts focus to task-relevant regions (e.g., optic disc for glaucoma, lung patterns for pneumonia)

Why it matters / caveats: This work establishes the largest image-tabular benchmark to date and introduces rigorous criteria for distinguishing genuinely multimodal problems from coincidental co-occurrence. Key limitations include selection bias (the criteria conflate task properties with algorithmic solutions), computational cost of LoRA fine-tuning, and no per-dataset hyperparameter tuning, meaning results represent conservative performance lower bounds.

Problem: Managing millions of trained LLM policies derived from shared base models requires efficiently moving LoRA adapter revisions through training, serving, scheduling, and data movement without materializing full model checkpoints for each policy.

Method: MinT operates across three scaling dimensions: Scale Up extends LoRA reinforcement learning (GRPO) to models up to 1T parameters with support for dense and MoE architectures and specialized attention mechanisms (MLA and DSA); Scale Down exports only LoRA adapters (under 1% of base-model size at rank-1) rather than full checkpoints via concurrent multi-policy GRPO training; Scale Out separates policy addressability from computational resources using tensor-parallel deployments to manage millions of addressable adapters while keeping only thousands actively loaded.

Key results:

• Adapter handoff speedup: 18.3× on 4B dense models; 2.85× on 30B MoE
• Training wall time reduction: 1.77× (dense) and 1.45× (MoE) via concurrent multi-policy GRPO
• Packed MoE LoRA tensors improve live engine loading by 8.5–8.7×
• Supports 10⁶-scale addressable adapter catalogs, validated with 100K single-engine measurements

Why it matters / caveats: MinT enables practical enterprise-scale LLM policy management by avoiding full checkpoint materialization, advancing efficient customization at million-model scale. Explicit limitations are not stated in the abstract, though implementation-specific constraints likely exist at the scale described.

Problem: Consistency-distilled video generation models degrade when more sampling steps are allocated at test time, because consistency distillation replaces the probability-flow ODE trajectory with a consistency-sampling trajectory that undermines step-budget scaling.

Method: AnyFlow shifts the distillation target from endpoint consistency mapping (z_t → z_0) to flow-map transition learning across arbitrary time intervals (z_t → z_r), enabling optimization of the full ODE sampling trajectory. Flow Map Backward Simulation decomposes full Euler rollouts into shortcut flow-map transitions for efficient on-policy distillation, reducing both discretization error in few-step sampling and exposure bias in causal generation. The framework supports both bidirectional and causal video architectures and is validated at model scales from 1.3B to 14B parameters.

Key results:

• AnyFlow matches or surpasses consistency-based counterparts in the few-step (low-budget) regime
• Unlike consistency distillation, AnyFlow scales monotonically with increasing sampling step budgets
• Demonstrated across 1.3B to 14B parameter models (specific FVD/benchmark numbers not provided in the abstract)

Why it matters / caveats: AnyFlow is claimed to be the first any-step video diffusion distillation framework based on flow maps, resolving a previously unresolved capability gap where users could not benefit from allocating more compute at inference. Specific quantitative benchmark comparisons were not accessible from the abstract alone.

Problem: While frontier vision-language models now support 128K+ token contexts, practical training recipes for extending context in VLMs remain insufficiently explored, particularly regarding data mixture design and length distribution.

Method: The paper extends Qwen2.5-VL-7B from 32K to 128K context via continued pre-training, scaling RoPE base frequency from 1×10⁶ to 4×10⁶ via Dynamic-NTK heuristics. The training data synthesis pipeline creates Long-Document VQA across three task types (single-page extraction, multi-page extraction, numerical reasoning) by sampling 8–15 page segments from 1.5M+ PDFs, generating QA pairs via Seed 2.0, then reinserting them into full-document context. The final recipe uses a 5B-token budget with an 8:2 extraction-to-reasoning ratio; pool-native length distribution (32K–128K naturally sampled) outperforms data concentrated at 128K; no short-context mixing is used. The resulting model is named MMProLong.

Key results:

• MMLongBench: MMProLong scores 57.70 vs. baseline 50.59 (+7.11 points)
• Generalizes without additional training to 256K (55.09) and 512K (52.52) contexts
• MM-NIAH (needle-in-haystack): average score improves from 20.0 to 49.4
• VTCBench (vision-text compression): 48.23 → 52.73
• Video benchmarks (Video-MME, MLVU, LongVideoBench): consistent improvements
• Short-context degradation minimal: 66.47 → 65.48 on general VQA/reasoning/OCR

Why it matters / caveats: The study establishes that long-document VQA supervision substantially outperforms OCR transcription as training signal, and diverse length distributions generalize better than target-length specialization. The recipe is primarily validated on Qwen2.5-VL-7B and Qwen3-VL-8B; transferability to other architectural families remains partially unexplored.

Problem: Given a black-box AI agent whose underlying LLM, prompts, and control logic are unknown, can a predictor accurately forecast its next decision from only a few prior interactions in language-mediated bargaining and negotiation games?

Method: The approach frames agent prediction as target-adaptive text-tabular learning where each decision point becomes a table row combining three feature modalities: structured game-state features (public configuration, current offer, round number, prior offers), semantic dialogue embeddings from a sentence encoder with PCA reduction, and an LLM-as-Observer whose hidden states (not generated outputs) serve as decision-oriented features. These features feed into TabPFN as the final tabular foundation model predictor. The system adapts to new target agents using K observed games (K ∈ {0, 2, 4, 8, 16}) plus source-population labeled rows from 13 frontier LLMs in the GLEE tournament, evaluated on 91 held-out hackathon agents.

Key results:

• Response prediction AUC at K=16 (bargaining): game+text baseline 0.791; best Observer (Gemma) 0.831 (+4.0 points); LLM-as-Predictor 0.770 (worse than Observer by 2.1 points)
• Response prediction AUC at K=16 (negotiation): game+text baseline 0.803; best Observer (Qwen3) 0.852 (+4.9 points); LLM-as-Predictor 0.785 (−6.7 points vs. best)
• Proposal prediction R² at K=16 (bargaining): baseline 0.622; best Observer 0.676; typical dollar error reduced from $552 to $473 (14%)
• Observer hidden states outperform direct logits: AUC 0.793 vs. 0.767 on bargaining with game features

Why it matters / caveats: The key finding—that frozen LLM hidden states are stronger predictive features than the model's own generated outputs—has broader implications for leveraging LLM representations in downstream tasks without fine-tuning. The study is limited to controlled bargaining abstractions rather than real markets, and requires access to a source population for training.

Problem: High-compression Variational Autoencoders (VAEs) for image generation suffer from reconstruction bottlenecks and convergence challenges that limit reconstruction fidelity and latent space suitability for diffusion modeling.

Method: Qwen-Image-VAE-2.0 introduces Global Skip Connections (GSC) and expanded latent channels to enhance reconstruction quality, uses an asymmetric and attention-free encoder-decoder backbone to reduce computational overhead, and implements an enhanced semantic alignment strategy to improve latent space structure for diffusion modeling. Training scales to billions of images and integrates a synthetic rendering engine targeting text-rich document scenarios. The model is also evaluated on a newly introduced benchmark, OmniDoc-TokenBench, which uses OCR-based metrics for document-heavy scenarios.

Key results:

• Claims state-of-the-art reconstruction performance on public benchmarks at high compression ratio
• Downstream diffusion experiments show superior diffusability, significantly accelerating convergence compared to existing high-compression baselines
• Exceptional capabilities in both general domains and text-rich scenarios (specific FID or PSNR numbers not provided in the available abstract)

Why it matters / caveats: The work introduces OmniDoc-TokenBench as a new evaluation standard for text-rich VAE scenarios, addressing a gap in existing benchmarks. No explicit limitations are stated; specific numerical comparisons were not accessible from the abstract.

Problem: Existing image editing benchmarks fail to reflect human judgment for frontier models due to limited task difficulty and coarse-grained evaluation, while reward model benchmarks suffer from distribution mismatch with realistic RL training scenarios.

Method: Edit-Compass contains 2,388 annotated instances spanning six task categories of progressive difficulty: General Tasks, Dynamic Manipulation, World Knowledge Reasoning, Algorithmic Visual Reasoning, Multi-Image Tasks, and Complex Tasks; evaluation uses three dimensions: Instruction Awareness, Visual Consistency, and Visual Quality. EditReward-Compass contains 2,251 preference pairs constructed via a two-stage process: sampling using FlowGRPO-inspired strategies, followed by human annotation requiring unanimous agreement across 5 of 8 expert annotators per retained pair.

Key results:

• 29 image editing models evaluated; best proprietary model (Nano Banana Pro): 3.99 overall score; best open-source model (Qwen-Image-Edit): 2.69 overall score
• Large performance gap on World Knowledge Reasoning: proprietary 3.89 vs. open-source 1.74
• 21 reward models evaluated; Qwen3.5-27B with thinking-enabled inference: 0.6998 average score
• Thinking-enabled inference improved Qwen3.5-9B by 9.83 points on EditReward-Compass
• Native multimodal models outperformed preference-trained alternatives on reward modeling

Why it matters / caveats: The benchmark reveals that frontier models perform adequately on perception tasks but struggle with deeper reasoning, world knowledge, and complex multi-image editing. A noted limitation is that evaluation relies on API-based MLLM judges, introducing reproducibility concerns tied to model version updates; the authors propose developing dedicated image-editing judge models as future work.

Problem: Dense 3D point tracking from monocular video requires reference-anchored representations that follow physical points across time, whereas pre-trained video diffusion transformers (DiTs) produce frame-anchored outputs—a fundamental mismatch that existing approaches address only with iterative or scratch-trained methods.

Method: TrackCraft3R uses a dual-latent representation: geometry latents (concatenated RGB and pointmap encodings per frame) and track latents (first-frame geometry replicated across all timestamps as dense query anchors). Both are concatenated and processed by the Wan 2.1-T2V video DiT fine-tuned with LoRA (rank 1024). Temporal RoPE Alignment assigns identical temporal indices to corresponding geometry and track latents so track latents attend to the correct timestamp's geometry. The model predicts residual displacements (Δⱼ = P₀(tⱼ) − P₀(t₀)) and visibility maps using MSE loss + 0.1-weighted BCE. Training uses a two-stage recipe (Stage 1: 3 days, frozen VAEs, lr 1e-4; Stage 2: 2 days, unfrozen VAE components) on Kubric, PointOdyssey, DynamicReplica, and TartanAir datasets at 480×832 on 12-frame clips.

Key results:

• Average AJ: TrackCraft3R+ViPE 0.5639 vs. DELTAv2+ViPE 0.4395 (+28%)
• Average APD₃D: 0.6817 vs. 0.6184 (+10%)
• Average OA: 0.9258 vs. 0.8144 (+14%)
• With DA3 geometry: AJ 0.6785 vs. DELTAv2+DA3 0.4975
• Speed: 3.91s vs. DELTAv2's 5.00s (1.3× faster) on 12-frame 448×448 clips
• Memory: 7.63 GB vs. DELTAv2's 35.46 GB (4.6× less)
• Temporal RoPE Alignment ablation: removal drops AJ from 0.5609 to 0.4450

Why it matters / caveats: This is the first method to successfully convert a video DiT's generative paradigm into reference-anchored dense 3D tracking without external heads, demonstrating strong transfer of spatio-temporal priors to structured correspondence. A key limitation is dependency on input geometry quality from foundation models (ViPE or Depth Anything 3); ground-truth geometry yields AJ of 0.7649, indicating meaningful headroom.

Problem: Standard many-shot in-context learning rules that reliably improve non-reasoning tasks do not transfer to chain-of-thought (CoT) reasoning tasks, where scaling demonstrations introduces instability, order sensitivity, and semantic-similarity-based retrieval failures.

Method: The paper reframes many-shot CoT-ICL as "in-context test-time learning" governed by two principles: Ease of Understanding (demonstrations must align with the model's comprehension; self-generated CoTs outperform dataset-provided ones for weaker models) and Smoothness of Information Flow (demonstrations should follow gradual conceptual progression). Curvilinear Demonstration Selection (CDS) operationalizes these principles using TSP-based optimization to minimize total curvature of demonstration sequences in embedding space while maintaining spatial proximity. Curvature is measured via embedding trajectory analysis (average r = −0.547 across tasks between curvature and performance).

Key results:

• Non-reasoning LLMs show negative gains on math reasoning even at 70B scale; reasoning LLMs (QwQ-32B, DeepSeek-R1) improve near-monotonically with more demonstrations
• At n=128, procedural CoT corruption degrades performance by 1.31–2.51 percentage points
• CDS geometry gain: up to 5.42 percentage points with 64 demonstrations
• CDS DetectiveQA gain: 2.29 points at 64 demonstrations (with GPT-5.2)
• Curvature-performance correlation: r = −0.545 (Geometry), r = −0.628 (Counting & Probability)

Why it matters / caveats: The work reframes the long context window from a retrieval buffer into a structured curriculum, challenging the assumption that demonstration ordering becomes irrelevant at scale for reasoning tasks. Limitations include analysis capped at n≤128 (CoT context constraints), TSP-based approximation being heuristic rather than globally optimal, and evaluation focused on mathematical and narrative reasoning without broader domain coverage.

Problem: Vision-Language-Action (VLA) model training treats all frames in robot demonstration trajectories equally, diluting policy learning by oversampling redundant segments (approaching objects, maintaining grasps) relative to manipulation-critical moments (contact, grasping, release).

Method: FrameSkip is a data-layer frame selection framework that operates during training without modifying VLA architecture, action heads, or inference. Frame importance is scored by combining four signals: Action Variation Importance (AVI, measuring action magnitude changes), Visual-Action Coherence (VAC, using DINOv2 features to capture visually grounded transitions), Task Progress Importance (TPI, encoding task-stage structure via Gaussian mixture models fitted to annotated manipulation-critical stages), and Gripper-Transition Preservation. Ratio-Aware Pruning selects frames above importance thresholds to achieve a target retention ratio (default 20%). Full-frame anchor batches (1 per 5 pruned mini-batches) stabilize learning. The framework is tested with StarVLA (Qwen3-4B-VL backbone, Diffusion Transformer action head).

Key results:

• RoboCasa-GR1 (24 tasks): 47.8% → 59.5% (+11.7%)
• SimplerEnv (4 tasks): 55.2% → 71.55% (+16.35%)
• LIBERO (4 suites): 96.5% → 97.4% (+0.9%)
• Macro-average across all benchmarks: 66.50% → 76.15% (+9.65%)
• Random frame selection yields only 66.83% macro-average, confirming principled selection drives gains
• Performance peaks at 50% retention; 20% retention still achieves strong gains

Why it matters / caveats: FrameSkip is positioned as the first frame-level supervision allocation study for VLA training, identifying temporal imbalance as an underexplored lever for embodied model training. A limitation is that the Task Progress Importance signal requires stage annotations for dataset-adaptive use, though a fallback Gaussian option is available.

Problem: Existing World-Action Models for autonomous driving treat world prediction and action generation as decoupled processes, whereas effective interactive decision-making requires that future scene evolution and planned maneuvers co-evolve during inference.

Method: DAWN (Denoising Actions and World iNteractive model) implements World-Action Interactive Models (WAIMs) through five components: Student/Teacher V-JEPA 2 Large vision encoders (512×256 resolution), an Auto-Encoder Resampler compressing visual tokens into 16 compact latent tokens, a causal Transformer World Predictor rolling out future latent states over ~4 seconds, a DiT-style World-Conditioned Action Denoiser generating trajectories conditioned on predicted futures, and recursive interaction over 4 rounds where predicted worlds refine action hypotheses and updated actions reshape world predictions. Training uses a four-stage pipeline: vision pretraining on OpenScene, DrivingDojo, and CoVLA; Auto-Encoder Resampler training; World Predictor supervised on future latent targets; and joint world-action training with shared denoiser weights.

Key results:

• NAVSIM v1 (perception-free): PDMS 89.1 (best among perception-free approaches), TTC 96.0 (best), No-at-fault Collisions 98.7
• nuScenes end-to-end planning: average L2 error 0.33m (vs. 0.47m prior best), average collision rate 0.11% (vs. 0.15% prior best)
• Ablation: World Predictor alone adds +2.4 PDMS points; enabling interaction adds further +2.7 points
• Performance plateaus after 4 interactive rounds

Why it matters / caveats: DAWN shows that explicit but limited future reasoning (4-second latent rollout) combined with tight world-action coupling improves both trajectory accuracy and safety metrics, suggesting interactive generation as a principled path toward actionable world models. Stated limitations include no formal convergence or safety guarantees for recursive interaction, short latent rollouts potentially insufficient for long-range multi-agent scenarios, and reduced interpretability from compact latent representations.

Problem: Standard flow-based image generation models predict velocities combining structured data with unstructured high-dimensional noise, requiring networks to carry excessive noise through internal states and causing numerical instability at high resolutions.

Method: AsymFlow introduces a rank-asymmetric velocity parameterization where the standard target u = ε − x₀ is replaced by u_A = Pε − x₀, with P an orthogonal projector onto a rank-r subspace applied patch-wise following Vision Transformer conventions, while full-rank data prediction is preserved. An analytical recovery mechanism converts asymmetric predictions back to full-rank velocity via orthogonal component decomposition. For latent-to-pixel finetuning of pretrained models (e.g., FLUX.2 klein 9B), the method uses Procrustes-based linear lift from latent to pixel space, trajectory coupling for semantic preservation, variance-reduced finetuning loss with control variates, and LPIPS perceptual correction for low-noise approximation errors. PCA initializes the subspace for training-from-scratch; Procrustes alignment is used for finetuning.

Key results:

• ImageNet 256×256 from-scratch: 1.57 FID with REPA loss (state-of-the-art among practical pixel diffusion models); 1.76 FID without REPA (vs. JiT baseline's 1.90)
• ~40% faster convergence than JiT baseline
• Numerical stability: FID degrades only 0.52 points when disabling low-noise clamping vs. JiT's 1.37 point degradation
• AsymFLUX.2 klein (finetuned FLUX.2 klein 9B at 1024×1024): HPSv3 10.66 (vs. latent base 9.50), DPG-Bench 86.8 (vs. 85.2), GenEval 0.82 (matching latent base)
• Optimal patch rank r=8; PCA-based subspace substantially outperforms random projections

Why it matters / caveats: AsymFlow provides the first practical method for converting large-scale pretrained latent flow models into pixel-space generators while achieving state-of-the-art quality, requiring no architectural modifications. The stated limitation is that latent-to-pixel finetuning assumes a good patch-level linear lift and may not work when the pretrained latent space does not preserve pixel structure (e.g., RAE models).

Problem: Agentic reinforcement learning for LLMs fails when reward states fall outside the policy's reachable capability region—a "reachability barrier"—causing training signal collapse because the model cannot naturally explore toward successful outcomes.

Method: ActGuide-RL injects action trajectory data as plan-style reference guidance to overcome reachability barriers, then internalizes gains through mixed-policy optimization. Three key components: Reachability Barrier Analysis formalizes policy failure via reachability dynamics measuring effective state-visiting mass (Ψ(s)); a Minimal Intervention Principle adaptively identifies the minimum guidance level k* needed per task via binary search, balancing barrier-repair benefit (Bk) against off-policy distribution shift risk (Rk); and Off-Policy Internalization trains guided and unguided rollouts jointly through modified importance ratios to transfer learned behaviors back to the unguided policy for inference. Evaluated on Qwen3-4B-Instruct across search-agent benchmarks.

Key results:

• GAIA: +10.68 pp over baseline RL (25.24% → 35.92%)
• WebWalkerQA: +27.79 pp (12.06% → 39.85%)
• XBench: +19.00 pp (18.00% → 37.00%)
• BC-ZH: +5.15 pp (15.26% → 20.41%)
• Performs on par with SFT+RL pipeline without cold-start initialization

Why it matters / caveats: ActGuide-RL reduces dependence on expensive supervised fine-tuning by leveraging abundant action trajectory data, offering an alternative to demonstration-based warm-starting. Stated limitations include sensitivity to action noise at 20% corruption rates (performance drops from 37.94% to 35.00% on WebWalker) and evaluation confined to search-agent tasks only.

Problem: Multi-hop question answering with RAG systems suffers from brittle reasoning when intermediate states are implicit in free-form natural language, lacking grounded error detection and producing opaque reasoning trajectories.

Method: PyRAG reformulates multi-hop RAG as program synthesis and execution using three specialized agents: a Decomposition Agent that breaks questions into atomic sub-queries, a Planning Agent that generates executable Python programs using two primitives (retrieve(query, topk) and answer(query, docs)) with explicit variable bindings for data dependencies, and an Answer Agent that produces short answers from retrieved evidence. Two execution-guided refinement mechanisms improve robustness: Compiler-Grounded Self-Repair uses runtime exceptions as deterministic signals for program revision, and Execution-Driven Adaptive Retrieval automatically increases retrieval scope when intermediate answers signal insufficient evidence. A reinforcement learning variant (PyRAG-RL) is also trained on Qwen2.5-7B.

Key results:

• Training-free Qwen2.5-7B: average EM 30.8% across five benchmarks vs. 19.0% Vanilla RAG (+11.8 points)
• Bamboogle: 41.5% (+25.5 vs. baseline); 2WikiMultihopQA: 33.4% (+14.5)
• PyRAG-RL: average EM 39.2% (comparable to ReSearch at 38.9%); 2WikiMultihopQA: 49.4% (highest among 7B models)
• Generalizes to Qwen3-4B (36.3% avg) and LLaMA-3.1-8B (40.9% avg)
• Achieves highest EM with only 3.1 average LLM calls per query

Why it matters / caveats: PyRAG demonstrates that explicitly formulating reasoning as program synthesis yields inspectable, deterministic feedback loops unavailable in natural language chains. Stated limitations include retrieval recall bottleneck (missing retrieval accounts for ~50% of errors), silent error propagation from sentinel-based adaptive retrieval, and planning agents sometimes bypassing multi-hop chains with single-retrieve operations that evade compiler detection.

Problem: Existing presentation generation systems require complete source documents (papers, reports) as input, whereas a generalist agent should be able to produce structured presentation videos from a bare natural language query alone.

Method: PresentAgent-2 is a pipeline comprising four components: a Deep Research Module that searches presentation-friendly web sources (tutorials, demos, webpages) biased toward rich media, filtering by content completeness and extracting text, images, GIFs, and videos; a Presentation Generation Backbone organizing retrieved resources into slide sequences with titles, bullets, and explanatory scripts while preserving playable dynamic media; three independent delivery modes (Single Presentation with single-speaker narration, Discussion with multi-speaker dialogue across complementary roles, and Interaction with audience Q&A grounded in slides and retrieved evidence); and Video Composition integrating slide visuals, synthesized narration audio, and dynamic media into final presentation video. Evaluated on PresentEval, a benchmark of 60 curated query-reference video pairs using VLM-based quiz scoring (0–5) and human subjective scoring (1–5).

Key results:

• Single Presentation mode: Quiz 4.84/5, Subjective 4.47/5
• Discussion mode: Quiz 4.85/5, Subjective 4.37/5
• Interaction mode: Quiz 4.85/5, Subjective 4.52/5
• Performance comparable to human-created reference videos (4.40–4.46 subjective scores)
• Other tested models show 4.18–4.47 subjective score range

Why it matters / caveats: PresentAgent-2 extends presentation generation from document-dependent slide creation toward query-driven, research-grounded video synthesis supporting multiple delivery formats. Key limitations include output quality dependence on available public multimodal resources, error propagation through the pipeline in Interaction mode, and a benchmark of only 60 examples that does not cover all presentation domains or audience types.