AI papers — 2026-05-19

Problem: Generating long videos autoregressively at high quality and real-time speed requires overcoming memory, compute, and parallelism bottlenecks that standard diffusion training and inference pipelines cannot address.

Method: LongLive-2.0 introduces Balanced Sequence-Parallel (SP) autoregressive training that pairs clean-history and noisy-target temporal chunks across distributed ranks, directly fine-tuning diffusion models into multi-shot autoregressive models without ODE initialization or distribution-matching distillation. At inference, the system uses W4A4 NVFP4 quantization (including KV cache quantization) on Blackwell GPUs, SP inference for non-Blackwell hardware, SP-aware chunked VAE encoding, and asynchronous streaming VAE decoding. Standalone LoRA weights enable conversion to real-time generation mode.

Key results:

• 2.15× training speedup
• 1.84× inference speedup
• 45.7 FPS inference on the LongLive-2.0-5B model

Why it matters / caveats: This is reported as the first NVFP4 system for long video generation, achieving real-time inference on 5B-parameter models. The system is tied to Blackwell GPU availability for full NVFP4 benefits; non-Blackwell deployments rely on sequence-parallel inference only.

Problem: Existing unified multimodal models require large-scale capacity increases to jointly handle image and video understanding, generation, and editing without one capability degrading another.

Method: Lance uses a dual-stream Mixture-of-Experts architecture operating on shared interleaved multimodal sequences, with decoupled capability pathways and modality-aware Rotary Positional Encoding to reduce cross-modal token interference. Training follows a staged multi-task paradigm with capability-oriented objectives and adaptive data scheduling that balances semantic comprehension against visual generation quality.

Key results:

• Substantially outperforms existing open-source unified models on image and video generation benchmarks (specific numbers not stated in the abstract)
• Retains strong multimodal understanding capabilities alongside generation

Why it matters / caveats: Lance demonstrates that collaborative multi-task training with architectural decoupling can replace brute-force model scaling for unified multimodal systems. Specific benchmark numbers are not reported in the abstract, limiting direct numerical comparison.

Problem: Reusable agent skill libraries extracted from LLM trajectories suffer from redundancy, uneven quality, environment sensitivity, and harmful updates that degrade agent performance over time.

Method: SkillsVote governs skills across three lifecycle phases: collection and profiling over a million-scale open-source corpus (assessing environment requirements, quality, and verifiability), agentic library search over a structured skill library for recommendation before execution, and evidence-gated post-execution evolution that attributes trajectory outcomes to specific skills and admits only verified successful updates. Skills couple executable scripts with non-executable procedural guidance.

Key results:

• Offline evolution improves GPT-5.2 on Terminal-Bench 2.0 by up to 7.9 percentage points
• Online evolution improves SWE-Bench Pro by up to 2.6 percentage points

Why it matters / caveats: SkillsVote shows that governed external skill libraries can improve frozen agents without model weight updates, a practically important result for deployment. Challenges around environment sensitivity and optimal skill-update thresholds are noted as ongoing.

Problem: Practitioners lack a systematic, evidence-based account of where AI-assisted research is reliable versus unreliable across the full research lifecycle.

Method: The paper conducts an end-to-end analysis of AI across four epistemological research phases — Creation (idea generation, literature review, coding and experiments, tables and figures), Writing, Validation (peer review, rebuttal and revision), and Dissemination — producing a structured taxonomy, benchmark suite, tool inventory, cross-stage design principles, and a practitioner-oriented playbook.

Key results:

• Fully automated systems can generate research papers for as little as $15
• AI excels at structured, retrieval-grounded, and tool-mediated tasks but remains fragile for genuinely novel ideas
• Frontier LLMs still fabricate results, miss hidden errors, and fail to judge novelty reliably
• Autonomous systems have not consistently achieved major-venue paper acceptance

Why it matters / caveats: The paper provides actionable guidance on a sharp stage-dependent boundary between reliable AI assistance and unreliable autonomy, directly relevant to any research team integrating AI tools. Greater automation can obscure rather than eliminate failure modes, cautioning against uncritical deployment.

Problem: Aligning streaming autoregressive video generators with human preferences is difficult because existing GRPO approaches add stochastic noise to ODE-based samplers, violating their deterministic dynamics and yielding low-quality exploration on the data manifold.

Method: KVPO replaces noise-based exploration with Causal-Semantic Exploration, which stochastically routes historical KV cache entries to produce semantically diverse generation branches while staying on the data manifold. Branch likelihood is estimated via Trajectory Velocity Energy (TVE) measured in the flow-matching velocity field, yielding a reward-weighted contrastive objective fully consistent with the ODE formulation. Online Group Relative Policy Optimization (GRPO) is then applied using this ODE-native surrogate policy.

Key results:

• Consistent gains in visual quality, motion quality, and text-video alignment across single-prompt short-video and multi-prompt long-video settings (specific numeric values not stated in the abstract)

Why it matters / caveats: KVPO addresses long-horizon coherence by manipulating high-level semantic progression rather than low-level appearance noise, which is the correct inductive bias for ODE-based video generators. Specific quantitative comparisons against baselines are not reported in the abstract.

Problem: Generating spatially accurate 3D indoor scenes from top-down view images is hindered by text-based methods lacking spatial precision and agent-based pipelines prone to infinite looping and instability.

Method: Code-as-Room uses a multimodal large language model (MLLM) within a structured execution harness to parse top-down room images and synthesize executable Blender code that sequentially specifies geometry, materials, and lighting. A cross-stage memory module persists context across pipeline stages to mitigate forgetting. The system represents 3D space in Blender code rather than generating geometry directly.

Key results:

• A dedicated evaluation benchmark for code-based 3D synthesis is introduced with various evaluation protocols (specific quantitative results not stated in the abstract)

Why it matters / caveats: Representing 3D scenes as executable code makes the generation process interpretable and editable, with direct application to interior design, VR, and embodied AI. Detailed numerical comparisons are referenced in the paper but not available from the abstract.

Problem: Formal theorem provers in Lean 4 fail to generalize because they do not leverage failed proof attempts and compiler feedback during training, relying instead on static proof corpora.

Method: OProver uses a two-phase iterative post-training loop: (1) continued pretraining on a 1.77M-statement, 6.86M-proof OProofs dataset of compiler-verified Lean 4 proofs, then (2) iterative cycles of agentic proving, memory indexing of retrieved context and failed attempts, supervised fine-tuning on repair trajectories, and reinforcement learning on unresolved cases. The 32B-parameter variant is the primary evaluated model.

Key results:

• MiniF2F: 93.3% Pass@32 (state-of-the-art)
• ProverBench: 58.2% Pass@32 (state-of-the-art)
• PutnamBench: 11.3% Pass@32 (state-of-the-art)
• MathOlympiad: 22.8% Pass@32 (second place)
• ProofNet: 33.2% Pass@32 (second place)

Why it matters / caveats: Integrating agentic proving into the training loop — rather than only at inference — allows the model to learn from its own failures in a principled way. Specific limitations and failure modes for the framework are not detailed in the abstract.

Problem: Static Mixture-of-Experts models activate the same number of experts for every token regardless of token difficulty, wasting compute on easy tokens that could be handled with fewer experts.

Method: Zero-Expert Self-Distillation Adaptation (ZEDA) injects parameter-free zero-output experts into each MoE layer of a fully trained static model, then applies two-stage self-distillation using the original frozen model as a teacher, combined with a group-level balancing loss. This converts static expert routing to dynamic routing without retraining from scratch.

Key results:

• Eliminates over 50% of expert FLOPs with minimal accuracy degradation
• Outperforms the strongest dynamic MoE baseline by 6.1 points on Qwen3-30B-A3B and 4.0 points on GLM-4.7-Flash
• Approximately 1.20× end-to-end inference speedup
• Evaluated on 11 benchmarks spanning math, code, and instruction-following

Why it matters / caveats: ZEDA enables practical conversion of already-deployed MoE models without retraining, directly lowering inference costs. The paper does not detail adaptation overhead, failure cases, or performance on specialized domains beyond the tested benchmarks.

Problem: Scaling Video LLMs to longer video content is bottlenecked by the per-frame compute cost of large vision encoders, not just the downstream LLM token count.

Method: LiteFrame trains a compact student vision encoder via Compressed Token Distillation (CTD), where the student learns to produce information-dense, spatiotemporally compressed representations from a larger teacher encoder. Language Model Adaptation (LMA) further refines the compact encoder for LLM compatibility. The resulting LiteFrame backbone replaces the full-size encoder in Video LLM pipelines.

Key results:

• 35% reduction in end-to-end latency compared to InternVL3-8B
• Processes 8× more frames within the same compute budget
• Improved average video understanding accuracy across multiple benchmarks

Why it matters / caveats: Shifting the optimization target from post-hoc token reduction to encoder-level compression unlocks a new scaling axis for video understanding. The abstract does not discuss performance trade-offs on fine-grained or domain-specific video tasks.

Problem: Large Reasoning Models generate excessively long chain-of-thought sequences because they lack a mechanism to detect when continued reasoning no longer adds new information.

Method: PUMA combines a lightweight Redundancy Detector that identifies steps where successive reasoning steps revisit established conclusions without novel progress, and an Answer-level Verification stage that confirms exit safety before stopping generation. The framework is plug-and-play and requires no retraining of the base reasoning model.

Key results:

• 26.2% average token reduction across five large reasoning models and five benchmarks
• Maintains answer accuracy throughout

Why it matters / caveats: PUMA offers a training-free way to cut reasoning token budgets substantially, with direct latency and cost benefits. The paper notes that answer-level signals alone may trigger early exit prematurely, which the detector mitigates but does not fully eliminate.

Problem: The maximum activation magnitudes in modern LLMs — a key factor for safe low-bit quantization — vary widely across model families and training stages but have not been systematically measured.

Method: The authors run a unified measurement pipeline on 27 checkpoints from 8 model families, using a 5,000-sample multi-domain corpus with family-specific tokenization and identical hooks across embeddings, hidden states, attention, MLP/MoE sublayers, SwiGLU gates, and final norms to record global and layerwise maximum activations.

Key results:

• Global maxima span nearly four orders of magnitude at comparable parameter counts
• Qwen3.5 and MoE checkpoints: 10² to 10³ range
• Gemma3-27B-it: approximately 7 × 10⁵
• MoE checkpoints exhibit 14.0–23.4× lower peaks than matched-scale dense counterparts
• The residual stream carries the global maximum in 22 of 24 checkpoints examined

Why it matters / caveats: Maximum activation magnitude is a model property tied to family, architecture, and training stage — not a simple function of parameter count — meaning empirical measurement is necessary before low-bit deployment rather than assumptions inherited from prior models. Study covers 8 families; generalization to all architectures requires further work.

Problem: Vision-Language-Action models degrade significantly when encountering visual disturbances at inference time that were absent from training data, limiting real-world robot deployment.

Method: StableVLA adds an Information Bottleneck Adapter (IB-Adapter), a lightweight module grounded in information-bottleneck theory that selectively filters visual input noise, to a 0.5B-parameter VLA backbone. The IB-Adapter adds fewer than 10M parameters and requires no additional training data or augmentation; it does not require pretraining on the Open X-Embodiment dataset.

Key results:

• Approximately 30% average performance gain over the baseline VLA
• Matches robustness of 7B-scale VLAs despite being 14× smaller
• Surpasses OpenPI under both synthetic and physical visual corruptions
• Fewer than 10M additional parameters

Why it matters / caveats: IB-Adapter demonstrates that robustness to visual corruption can be achieved through architectural filtering rather than data augmentation, enabling deployment on resource-constrained robots. The robustness improvements are focused on visual corruption specifically and may not transfer to other distribution shifts.

Problem: Extending LLM context windows to long lengths typically requires expensive training on full-length sequences, which is quadratically costly.

Method: EndPrompt constructs two-segment training sequences: the original short context is kept intact as the first segment, and a brief terminal prompt with positional indices set near the target long context length is appended as the second segment. This exposes the model to long-range relative distances within RoPE (Rotary Position Embedding) without requiring full-length inputs, with convergence guarantees derived from the Bernstein inequality applied to position interpolation's smoothness constraints on the attention function.

Key results:

• EndPrompt RULER score: 76.03 (highest among compared methods at 64K context on LLaMA)
• LCEG: 72.24
• LongLoRA: 72.95
• Full-length fine-tuning: 69.23
• Highest score on LongBench among compared approaches

Why it matters / caveats: EndPrompt challenges the assumption that dense long-sequence training is necessary for context window extension, achieving state-of-the-art RULER scores with only short training sequences. Evaluation is limited to specific LLaMA model families; broader applicability to other architectures and position encoding schemes requires further study.

Problem: Hybrid diffusion-autoregressive language models underperform because diffusion is applied in hidden-state spaces that are geometrically ill-suited for denoising and token recovery.

Method: DiHAL uses geometry-based proxy metrics to score each transformer layer and identify the hidden-state interface most amenable to diffusion. A diffusion bridge module then replaces the lower transformer layers at that interface, while upper layers and the original LM head are preserved. Diffusion operates on hidden states rather than on discrete tokens directly, sidestepping continuous-to-discrete recovery problems. The system reuses pretrained transformer weights for the preserved upper layers.

Key results:

• Geometry scores successfully predict effective diffusion insertion layers under fixed training protocols on 8B-scale model backbones
• Hidden-state recovery outperforms continuous diffusion baselines in diagnostic comparisons (specific metrics not stated in the abstract)

Why it matters / caveats: DiHAL provides a principled, geometry-driven methodology for designing diffusion-LM hybrids rather than relying on heuristic layer selection. Current evaluation is limited to diagnostic comparisons at 8B scale; full-scale generation quality benchmarks are not reported.

Problem: LLMs frequently mis-invoke or fail to invoke external tools not because they lack knowledge of when tools are needed, but because of a gap between internal recognition and executed behavior.

Method: The paper defines tool necessity relative to each model's actual task capabilities (model-adaptive), then applies a two-stage decomposition: probing LLM hidden states at late layers for linearly decodable tool-necessity signals (internal cognition stage), and measuring whether that recognition translates into actual tool-call behavior (execution stage). Four models are analyzed on arithmetic and factual QA tasks.

Key results:

• Arithmetic task mismatch rate: 26.5–54.0% across models
• Factual QA mismatch rate: 30.8–41.8% across models
• Most mismatches occur at the cognition-to-action transition, not in cognition itself
• Tool-necessity signals are linearly decodable from late-layer, last-token hidden states

Why it matters / caveats: The knowing-doing gap finding implies that improving tool-use reliability requires targeting the representational-to-behavioral translation step, not just improving tool-necessity recognition. The study covers only four models and two task categories, limiting generalizability.

Problem: Chunked prefill attention in long-context LLMs is slow because existing sparse attention approaches either lose efficiency at limited query lengths or require costly explicit KV cache compaction.

Method: CompactAttention treats 2D block-sparse attention masks as KV-selection signals rather than direct sparse-kernel execution plans. It applies Q-block union and intra-group union operations under GQA (Grouped Query Attention) to convert sparse masks into per-group KV block tables that enable in-place KV block access in paged execution, eliminating explicit compaction.

Key results:

• Up to 2.72× attention speedup at 128K context length under chunked prefill on LLaMA-3.1-8B-Instruct
• Maintains accuracy close to dense attention on the RULER benchmark

Why it matters / caveats: CompactAttention provides a practical, accuracy-preserving method for accelerating long-context LLM serving without retraining or modifying the model weights. Evaluation is on a single model (LLaMA-3.1-8B-Instruct) and one benchmark; performance on other architectures and real serving workloads is not characterized.