AI papers — 2026-06-02

Problem: Existing scientific figure generation methods generalize poorly across figure types and cannot produce editable vector outputs.

Method: Crafter uses a four-role multi-agent orchestration loop (Designer, Executor, Verifier, Reviser) around an evolving figure specification, with diversity-driven plan exploration (K parallel candidates), a structured corrective layer that accumulates typed edits rather than free text, and a verify-then-refine loop with directive critics providing per-dimension diagnostics over up to T=3 rounds. CraftEditor applies the same harness pattern to raster-to-vector conversion via extraction, processing, and SVG composition phases.

Key results:

• Crafter scores 50.34% on PaperBanana-Bench vs. 11.13% for the Nano Banana 2 baseline (+39.21 pts)
• Crafter scores 50.20% on CraftBench vs. 19.90% baseline (+30.30 pts)
• Outperforms prior SOTA PaperBanana by 16.61 pts on PaperBanana-Bench and 22.20 pts on CraftBench
• CraftEditor scores 8.04/10 on editable output quality vs. AutoFigure-Edit (6.91) and Edit-Banana (3.69)
• Ablation: removing the corrective layer costs −8.90 pts; removing plan exploration costs −8.56 pts

Why it matters / caveats: The harness design shows that orchestration (not architectural changes) alone substantially improves structured visual output generation. Generalization is validated across four input conditions and three figure styles on the 279-sample CraftBench. The method depends on the quality of the underlying image generators, so gains from stronger future generators are expected but not yet measured.

Problem: Parameter-Efficient Fine-Tuning (PEFT) has been treated only as a budget substitute for full fine-tuning, leaving unexplored its potential as a substrate for persistent, personalized model instances at scale.

Method: The paper reframes small trainable adapters (e.g., LoRA) as "persistent local state" layered on top of a shared foundation model, treating the base model as shared competence and adapters as encoding per-user preferences, skills, tool habits, and memory. It proposes three scaling dimensions—scale up (stronger base models), scale down (minimum viable adapter size), and scale out (millions of concurrent instances)—supported by the MinT infrastructure system handling adapter identity, revision, provenance, evaluation, and serving residency.

Key results:

• Not stated

Why it matters / caveats: The framing opens a path toward trillion-parameter personalized models served at massive concurrency, which has broad implications for AI personalization infrastructure. The lack of available experimental results in the fetched content makes it difficult to assess empirical claims.

Problem: Existing agent benchmarks like τ²-Bench are saturating, with agents achieving high scores that reflect benchmark limitations rather than genuine capability.

Method: TASTE (Task Synthesis from Tool Sequence Evolution) uses an Adaptive Contrastive n-gram model trained on LLM-judged validity signals to sample valid tool sequences, applies clustering for representative sequence selection, and runs iterative difficulty evolution refinement to construct τ^c-Bench as an extension of τ²-Bench across three domains.

Key results:

• Gemini-3-Flash drops from 0.82–0.94 on original τ²-Bench to 0.28–0.61 on τ^c-Bench
• Generated tasks more than double unique tool combinations vs. the original benchmark
• Evaluated across 11 agent/user LLM pairs

Why it matters / caveats: Automated difficulty scaling addresses benchmark saturation, enabling continuous evaluation as agents improve. The reliance on LLM-judged validity signals for training the n-gram model may introduce biases from the judge model.

Problem: Web browsing agent benchmarks do not adequately evaluate performance in Korean-language contexts, obscuring large capability gaps between global and Korean-specific models.

Method: K-BrowseComp comprises 400 problems (300 human-validated, 100 synthetic) grounded in Korean contexts, with 53.3% multi-hop and 46.7% parallel-branching question formats. Evaluation uses the search_evals framework with a Perplexity Search backend, a 10-search-call budget per question, and pass@1 accuracy. Synthetic questions are generated by Claude Opus 4.7 and filtered through searchability, well-formedness, and adversarial difficulty checks (37.3% yield rate).

Key results:

• GPT-5.5: 45.67% accuracy (best overall)
• DeepSeek-V4-Pro and GLM-5.1: ~30%
• K-EXAONE-236B-A23B (Korean open-weight): 10.33%
• A.X-4.0 (Korean): 5.33%
• Synthetic split: 0–26% across all models
• Failed trials use 8–9.8 search calls vs. 5–8 for successful ones

Why it matters / caveats: The benchmark exposes a severe performance gap for Korean-specialized open-weight models despite language-specific training, pointing to deficits in multi-hop trajectory management and constraint tracking. The 10-call budget and single-run metric may underestimate best-case capabilities.

Problem: Draft models for speculative decoding trained via supervised fine-tuning on target-generated trajectories plateau quickly due to a mismatch between the offline training distribution and the on-policy inference distribution.

Method: Draft-OPD applies on-policy distillation with error-position replay: it runs speculative decoding to collect rollouts while recording anchor positions where draft blocks begin, then replays from anchors to compute student/teacher log-probabilities on draft-generated prefixes. The distillation objective applies forward KL divergence for accepted tokens and reverse KL for rejected tokens, with exponential position weighting (γ=0.8) to emphasize early failures. The approach extends EAGLE/DFlash-style training-based draft models.

Key results:

• Qwen3-4B: 5.31× speedup, τ=5.96 (vs. DFlash: 4.51×, τ=5.51)
• Qwen3-8B: 5.36× speedup, τ=5.73 (vs. DFlash: 4.67×, τ=5.19)
• Average 23% improvement over EAGLE-3 and 13% over DFlash
• SGLang deployment: 11.2% average acceptance length improvement, up to 17% throughput at concurrency 32
• Evaluated on GSM8K, MATH-500, AIME25, MBPP, HumanEval, SWE-Lite, MT-Bench

Why it matters / caveats: The method bridges the train/inference distribution gap for speculative decoding draft models with practical gains in real serving engines. Results are currently limited to Qwen3 models; generalization to other base model families is not yet demonstrated.

Problem: Video Generation Models (VGMs) used as reasoners are limited by their intrinsic capabilities, and existing approaches that use VLMs as solvers fail to leverage VGMs' visual generation strengths.

Method: A VLM teacher synthesizes task-adaptive supervision queries (goal achievement and process constraints) and formulates them as differentiable binary reward signals. A VGM reasoner equipped with a LoRA module (rank 16) is then optimized at test time using flow matching with a differentiable VQA loss. Step distillation compresses the VGM to a 4-step generator, and a lightweight surrogate decoder enables efficient optimization with loss-based early stopping (threshold τ=0.1).

Key results:

• VBVR-Bench (symbolic reasoning): 0.781 vs. baseline 0.666 (+0.115)
• RULER-Bench (general-purpose): 68.2 vs. baseline 46.4 (+21.8)
• Outperforms VLM-as-Solver by 0.4 pts and Best-of-N scaling by 2.2 pts
• Average 16 optimization steps with early stopping; covers all 30 RULER-Bench task categories

Why it matters / caveats: The VLM-as-Teacher paradigm effectively separates high-level logic (VLM) from visual execution (VGM), enabling reasoning beyond either model's individual limits. The approach adds per-instance test-time optimization cost and currently requires task-specific reward formulation by the VLM teacher.

Problem: Current Multimodal Large Language Models (MLLMs) have not been systematically evaluated on concurrent multi-stream video understanding, where multiple simultaneous video streams must be integrated in real time.

Method: X-Stream is a benchmark of 4,220 QA pairs across 932 videos (857 hours of raw multi-stream data from 20+ sources, 8 domains, 2–5 concurrent streams), built with a dual-verification process by 31 expert annotators. Three multiplexing strategies are evaluated: Spatial Division (pixel concatenation), Temporal Division (sequential frame sampling with temporal embedding alignment), and Semantic Division (token selection via Determinantal Point Process). Online inference uses 1-second intervals with a sliding memory window and a 250-token/second budget.

Key results:

• Best proprietary model Gemini 3 Pro: 73.38% instant, 72.23% backward, 20.77% forward, 82.04% comprehensive accuracy
• Best open-source model Qwen3-Omni-30B-A3B: 34.28% overall
• Streaming-specialized VideoLLM-online: 8.48%
• Temporal Division optimal for dual-stream; Semantic Division superior for 3+ streams

Why it matters / caveats: The benchmark reveals a critical gap between proprietary and open-source models on real-world multi-stream tasks, with forward (future) reasoning being the hardest. The 50% accuracy plateau for state-of-the-art models suggests current architectures are not designed for true multiplexed stream integration.

Problem: Autoregressive video diffusion models face prohibitive KV cache memory costs at minute-scale generation due to dense per-head key-value storage.

Method: VideoMLA applies Multi-Head Latent Attention (MLA) to the Wan-1.3B transformer backbone, replacing dense per-head keys/values with a shared low-rank content latent (down-projected to dc=192 dimensions) and a decoupled 3D-RoPE positional key (32 channels). The per-token cache stores 224 scalars versus 3,072 for standard MHA, achieved through a joint down-projection operator. The effective rank is determined by the MLA bottleneck rather than the pretrained spectral structure.

Key results:

• 92.7% reduction in per-token KV memory across 30 transformer layers (13.7× ratio)
• VBench overall score: 0.859 for both 30-second and 60-second generation (best among evaluated methods)
• Throughput: 23.96 FPS on B200 (1.23× improvement)
• CLIP-T: 0.3278 (best among chunk-wise models)
• 8.0× batch scaling headroom vs. dense MHA under fixed B200 memory

Why it matters / caveats: The work demonstrates that memory compression for long-form video generation does not require sacrificing quality if the architectural bottleneck is correctly identified. Results are on the Wan-1.3B backbone; applicability to larger video diffusion architectures is not directly shown.

Problem: LLM agents relying on trajectory-level reflection for skill updates cannot precisely attribute failures to specific skills, leading to unstable and inaccurate skill maintenance.

Method: SkillAdaptor is a training-free framework that performs step-level failure attribution through three stages: (1) a Localizer that identifies the earliest accountable fault step and failure description, (2) a Linker that estimates responsibility weights across retrieved skills to route to revision or generation, and (3) a Qualification phase that re-executes tasks and accepts skill updates only when performance is non-negative (Δ≥0). Skill retrieval uses Qwen3-Embedding-8B with cosine similarity, top-10 candidates reranked by the backbone LLM.

Key results:

• WebShop: +1.7 pp success rate vs. EvoSkill baseline
• PinchBench: +1.5 pts on Avg Score% (GLM-5)
• Claw-Eval: +1.8 on Avg Score (Kimi-K2.5)
• Removing the Localizer+Linker drops WebShop success from 33.0% to 28.6%
• Tested on three backbone LLMs: Kimi-K2.5, GLM-5, GPT-5.2

Why it matters / caveats: Step-level attribution is more auditable and stable than trajectory-level approaches and requires no backbone fine-tuning. Gains are modest (+1–2 pts) and concentrated in early adaptation rounds; the method may plateau after major procedural deficiencies are corrected.

Problem: Foundation models have not been rigorously tested on embodied spatial navigation tasks requiring translation control to match a target viewpoint in 3D environments.

Method: TVRBench is a 500-task indoor simulation benchmark (120 iTHOR single-room + 120 ProcTHOR multi-room scenes) where an agent selects from 9 discrete actions (translations, rotations, head pitches, stop). Post-training uses expert-trajectory SFT, chain-of-thought supervision via MiMo-V2.5 annotations, single-turn GRPO with per-step action-matching rewards, and multi-turn GRPO with live simulator rollouts on a Qwen3.5-9B backbone.

Key results:

• Best open-source baseline (Qwen3.5-27B): 7.8% success (action-only memory)
• Best closed-source baseline (Gemini-3.1-Pro): 12.0% success
• Human reference: 93.0%
• Removing translation actions raises an untrained 9B model from 2.8% to 80.5% (body translation is dominant failure mode)
• VA-SFT: 50.8% overall success (82.4% single-easy, 68.8% single-hard)
• VA-SFT + Multi-turn GRPO: 51.4% (+7.2 pp on multi-room easy tasks)

Why it matters / caveats: The 85-point human–model gap isolates body translation control as the core bottleneck in embodied spatial reasoning, distinct from visual recognition. CoT supervision hurts performance under visual-action memory, suggesting that verbalized reasoning chains are misaligned with low-level motor control.

Problem: It is unclear whether VLM or video generation model (VGM) pretraining produces better representations for different spatial intelligence tasks (semantics, instance grouping, 3D geometry).

Method: Frozen-feature probing extracts intermediate layer representations from VLMs (InternVL3, InternVL3.5, Qwen2.5-VL, Qwen3-VL) and VGMs (WAN, CogVideoX, OpenSora-2.0, Aether) using a unified transformer backbone with alternating frame and global attention. Task-specific lightweight readout heads are trained for semantic tagging (asymmetric loss), instance grouping (multi-view contrastive pull-push loss with HDBSCAN), and 3D geometry (confidence-weighted regression) on ScanNet20 and DL3DV-10K.

Key results:

• Semantic tagging: VLM avg mAP 92.08 vs. VGM avg 69.89
• Instance grouping: VLM avg T-mIoU 22.66 vs. VGM avg 13.24
• 3D geometry depth: VGM avg AbsRel 0.072 vs. VLM avg 0.113
• 3D camera AUC@30: VGM avg 0.527 vs. VLM avg 0.330
• Feature fusion (WAN2.1 + Qwen3-VL-8B): mAP 92.30, depth AbsRel 0.042, camera AUC@30 0.615

Why it matters / caveats: The complementarity finding has direct implications for architecture design: fusing VLM and VGM features yields the best of both worlds. The probing protocol uses frozen weights, so results may not reflect fine-tuned performance; the feature fusion is simple concatenation rather than learned integration.

Problem: Standard next-token prediction (NTP) pre-training causes representation degeneration (anisotropy, rank collapse) because it provides no dense supervision signal in the continuous representation space.

Method: NITP adds an auxiliary cosine similarity loss on top of NTP, where the prediction target is the stop-gradient hidden representation of the next token extracted from shallow layers (~20% of total depth). The total loss is ℒ_total = ℒ_NTP + λ·ℒ_NITP (λ=1.0), adding approximately 2% additional training FLOPs with zero inference overhead. The method is applied to both dense transformers and MoE architectures.

Key results:

• 9B MoE: +5.71% on MMLU-Pro (15.29%→21.00%), +6.36% on C3, +4.26% on CommonsenseQA
• 3B MoE: +2.1 pt average across 13 benchmarks
• 2B Dense: +1.79 average pts (+2.18 MMLU, +4.16 C3)
• MTEB evaluation: +2.33 pts overall (39.24→41.56), improving 23/25 tasks
• Ablation: cosine similarity > Smooth L1 > MSE > KL divergence as loss function
• Scales from 0.5B to 45B parameters

Why it matters / caveats: The method provides a theoretically grounded (via null-space analysis) and practically cheap way to improve representation quality during pre-training, with consistent gains across architectures and scales. Results are based on training from scratch; impact when continuing pre-training from an existing checkpoint is not explored.

Problem: Dense self-attention in video diffusion transformers scales quadratically with frame count, making minute-scale generation computationally prohibitive or causing out-of-memory errors.

Method: LVSA restricts each query frame's attention to a union of equidistant global anchor frames (period T_per) and a local sliding temporal window W(t), reducing complexity from O(N²d) to O(TCP²d) asymptotically linear in frame count. Expanded window bounds handle global/local overlap, and a rotating periodic global frame mechanism shifts anchor positions by denoising step modulo T_per to eliminate fixed-grid bias. The method requires no retraining and is applied to Wan 2.1 (1.3B and 14B), HunyuanVideo 1.5, and Wan 2.2.

Key results:

• Wan 2.1 1.3B: 3.17× speedup at 6× horizon (481 frames)
• Wan 2.1 14B: 2.98× speedup at 6× horizon
• HunyuanVideo 1.5: 3.33× speedup at 1.5× horizon
• Enables 2× horizon generation on HunyuanVideo (257 frames) where dense attention causes OOM
• VQeval composite advantage: +4.7 to +12.1 over dense attention at extended horizons
• Outperforms RIFLEx (2.41×) and UltraViCo (3.27×) speedups

Why it matters / caveats: The training-free nature and cross-architecture applicability (single-stream, dual-stream, various RoPE configurations) make LVSA broadly deployable. The paper introduces VQeval to address VBench-Long's bias toward static outputs, which is an important methodological contribution to video evaluation.

Problem: There is no benchmark for evaluating LLM agents on real-world personalized MCP (Model Context Protocol) tools, where agents must handle user-specific context, implicit constraints, and cross-tool coordination.

Method: MCP-Persona uses three components: Tool-Traverse (a traverse-then-simulate paradigm collecting authentic function calls from 24 real MCP servers and synthesizing Python simulation kernels via adversarial perturbation), Context-Tree (a hierarchical user profile structure populated via enumeration, LLM generation, random synthesis, and sanitized real data), and Persona-Gen (a two-stage task generation pipeline producing 173 human-verified tasks with instruction fuzzification). Evaluation covers 12 personalized MCP servers across social media, enterprise platforms, and content management systems.

Key results:

• Best model Claude-Sonnet-4.5: 41.50% overall accuracy
• GPT-5: 41.45% accuracy
• No model exceeds 50% on either checkpoint accuracy or execution accuracy
• Tool-Traverse: 94.0% behavioral alignment vs. 58.0% for documentation-only baseline
• Human-LLM judge correlation: 91.5% alignment
• Context distractors reduce performance by 4–5 points

Why it matters / caveats: The benchmark identifies three systematic failure modes (under-exploration, skipped dependencies, over-long context degradation) that current agent architectures do not address. Limited to 173 tasks across 12 servers, which may not cover the full diversity of real-world MCP deployments.

Problem: RL training of LLMs with PPO wastes compute on post-failure reasoning tokens after a model commits an early wrong step, slowing learning and increasing cost.

Method: ESPO detects trajectory failures during rollout by combining a stepwise deviation signal (regret = greedy log-prob minus sampled action log-prob), EMA normalization, and a value-gated termination condition (cumulative regret exceeds β times critic value estimate). Truncated rollouts are treated as absorbing failure states with terminal reward r_fail=−1.0, and critic warmup disables stopping until the value function stabilizes. The method runs on top of PPO without requiring additional reward models.

Key results:

• DeepSeek-R1-Distill-Qwen-7B on AIME 2024: 46.28% (vs. PPO 45.25%, DAPO 45.57%)
• AMC 2023: 85.83% (vs. PPO 82.94%, DAPO 83.76%)
• MATH-500: 87.42% (vs. PPO 85.43%, DAPO 85.95%)
• Token savings: 839M cumulative (22% fewer than PPO's 1,072M)
• 1.5B model average accuracy: 59.09% (vs. PPO 57.03%, DAPO 58.29%)
• Random truncation baseline: 42.4% (confirms gains come from where, not just how much, truncation occurs)

Why it matters / caveats: ESPO improves both efficiency and performance simultaneously, with the critical insight that detecting early failure location (not just length) drives gains. The method is evaluated only on mathematical reasoning benchmarks; generalization to coding or other RL-trained tasks is not demonstrated.

Problem: Answering visual questions directly from fMRI brain signals is unexplored, requiring both neural decoding of visual content and language-grounded question answering from brain activity.

Method: The Brain-IT-VQA model (BIT-L) extends the Brain Interaction Transformer with dual pathways: a CLIP-aligned pathway producing CLIP-compatible representations via cross-attention, and a direct conditioning pathway generating soft prompts for a frozen InstructBLIP (Q-Former + Flan-T5-XL). fMRI voxels are organized into 128 functional clusters as "Brain Tokens." Training uses a two-stage approach: Stage 1 pretraining with MSE loss targeting CLIP tokens and Q-Former outputs, then Stage 2 end-to-end LoRA fine-tuning on caption/VQA tasks, augmented with ~120k synthetic fMRI responses.

Key results:

• Captioning BLEU-4: 24.81 vs. MindLLM 21.24; CIDEr: 0.683 vs. 0.659
• VQA-v2 accuracy: 56.95% vs. MindLLM 52.14%
• NSD-VQA accuracy: 73.78% ± 0.92 (averaged across subjects)
• Category range: binary questions 79–93%, fine-grained color 47.84%, scene-level 93.00%
• NSD-VQA dataset: 73K images, ~1.46M QA pairs across 20 question categories

Why it matters / caveats: The work bridges brain signal decoding and multimodal question answering, with interpretability analysis showing region-specific contributions to different question types—useful for neuroscience research. The model is trained and evaluated on fMRI from the Natural Scenes Dataset (4 subjects), so generalization to other neuroimaging paradigms or subject populations is unknown.

Problem: Generating synchronized long-horizon character audio-video streams in real time is limited by coherence drift across chunks and insufficient inference speed for interactive use.

Method: StreamChar decouples long-horizon orchestration from short-window denoising: an LLM orchestrator reads transcripts and historical context to produce frame-aligned audio conditions via a Progress-Aware Pointer (PAP), while a joint audio-video DiT performs bidirectional denoising within chunks using a sink-frame mechanism (first chunk as persistent long-range anchor). A Modality-Aware MoE routes to separate FFN experts for cross-modal interaction vs. modality-specific processing, with modality-aware RoPE (1/4 frequency ratio) handling 4:1 video-to-audio token density. Two-stage distillation compresses the DiT from 50 to 4 denoising steps via distribution matching then online rollout simulation.

Key results:

• Word Error Rate: 3.65% post-distillation
• Sync-C (audio-visual sync): 8.126
• FID: 18.963; FVD: 289.091
• Human Anatomy: 0.941; Human Identity: 0.924
• Quality Drift: 0.0067 (negligible long-horizon degradation)
• Per-chunk latency: ~1.34 seconds (33 frames at 24 fps) on a single H100 GPU

Why it matters / caveats: The decoupled architecture achieves real-time streaming character generation on a single H100, with negligible quality drift over long horizons—a practical result for interactive character applications. The evaluation dataset SpeakerVid-5M and the EMTD benchmark are less established than standard benchmarks, making cross-paper comparison difficult.