AI papers — 2026-06-22

Problem: Autoregressive multimodal LLMs must caption multiple image regions sequentially, causing inference latency to grow linearly with region count.

Method: PerceptionDLM is a multimodal discrete diffusion language model (DLM) built on LLaDA-8B with a SigLIP-2 vision encoder and two-layer MLP connector. The base model (PerceptionDLM-Base) is trained in four stages via visual instruction tuning on up to 22M samples (LLaVA-OneVision-1.5) plus a high-quality SFT refinement stage on Honey-Data-15M. Parallel region perception is enabled by two additions: learnable per-region prompt embeddings (spatially fused with visual tokens per mask) and a structured block-wise attention masking mechanism that allows all N region captions to denoise simultaneously while preventing inter-region interference. A new benchmark, ParaDLC-Bench (2,345 manually verified questions), evaluates concurrent multi-region captioning quality and efficiency jointly using GPT-5.2 as judge.

Key results:

• ParaDLC-Bench accuracy: 62.4% vs LLaDA-V baseline 35.2%
• Throughput at 5 masks/image: up to 3.5× speedup over autoregressive sequential baseline
• Constant-workload parallelism (4 masks): 3.44× TPS at degree-4 vs degree-1
• PerceptionDLM-Base beats LLaDA-V on 15/16 multimodal benchmarks (e.g., ChartQA 91.6, DocVQA 89.9, MMVP 82.0, MMBench 85.0)
• Competitive with Qwen2.5-VL-7B and InternVL3-8B on fine-grained perception while trailing on heavy reasoning (MMMU 47.2 vs 51.3)

Why it matters / caveats: Diffusion LMs' intrinsic parallel decoding is finally exploited for practical multi-region perception, breaking the linear-latency bottleneck of AR models with no extra forward passes. The parallel decoding paradigm still fundamentally limits sequential reasoning traces; the paper notes DLMs underperform AR models on MMMU and MathVista and calls out RL as a needed next step.

Problem: All existing LLM-agent memory benchmarks assume single-user settings; shared-memory agents in hospitals, enterprises, and households must simultaneously serve utility, enforce access-control boundaries, and honor deletion requests — a governance challenge that has not been benchmarked.

Method: GATEMEM spans 91 long-form multi-party episodes and 2,218 hidden checkpoints across four domains (medical, office, education, household). Each episode contains an evolving principal-to-role permission structure; checkpoints test three governance pillars scored as: Effective Utility U, Access-Control Violation rate A, and Active Forgetting Failure rate F, combined into a multiplicative Memory Governance Score MGS = U × (1−A) × (1−F). Six LLM backbones (GPT-5.4, Deepseek-V4-Pro, Llama-4-Maverick, GPT-5-mini, GPT-4o-mini, Gemini-2.5-Flash-Lite) are tested across five memory architectures: LONG-CONTEXT, RAG-NAIVE, RAG-POLICY, A-MEM, MEM0, and REMEM (graph-based).

Key results:

• No method achieves simultaneously high utility and low leakage across all domains in any backbone
• Best single-domain MGS: GPT-5.4 LONG-CONTEXT on Medical = 80.1; Office = 57.0 (RAG-POLICY), Education = 71.0 (Deepseek LONG-CONTEXT), Household = 68.5 (Deepseek LONG-CONTEXT)
• LONG-CONTEXT still exceeds 20% A or F in multiple backbone-domain combos despite achieving best MGS
• REMEM (graph-based) cuts to ~1k tokens/ckpt but incurs up to 260 s/ckpt latency vs 4.22 s/ckpt for LONG-CONTEXT on Medical
• Human–judge agreement: max 1.04 pp absolute difference across U/A/F/MGS; ≥97.7% field-level agreement

Why it matters / caveats: Demonstrates that scaling context alone does not solve memory governance — access control and deletion compliance require dedicated architectural treatment. GATEMEM is an interface-level benchmark and does not certify physical erasure from underlying stores; backbone quality has a large confounding effect on the governance trade-off.

Problem: Presentation-generation agents regenerate entire slide decks per revision turn and carry no persistent user preferences across sessions, forcing users to restate styling preferences on every job.

Method: MemSlides is a hierarchical memory agent combining three stores with different lifetimes: (1) user profile memory — intent-conditioned, organized by theme/layout/content/visual/template dimensions, persisting across jobs and routed into working memory at job start via intent-matching and conflict resolution; (2) tool memory — two granularities of reusable execution experience (round-scope task experience and operation-scope tool-chain experience for reliability during editing); (3) working memory — session-scoped state carrying active temporary preferences and edit-state records across revision rounds. A Plan–Act–Guard loop enforces scoped slide locality: each edit binds to the minimum affected slide region, with snapshot-hash-bound patch operations and a coverage gate that blocks premature finalization.

Key results:

• Persona-alignment (0–10 scale, averaged across 3 personas): MemSlides+GLM-5 scores 9.00 / 9.00 / 8.56 (Content / Structure / Visual) vs DeepPresenter 6.67/7.61/5.28 and SlideTailor 4.44/4.89/4.00
• Tool-memory injection: closed-loop completion 96.3% (vs 81.5% without) and first-correct-edit time 0.327× on 9 diagnostic pairs (overall across GPT-5 / GLM-5 / Gemini 3.1 Pro)
• User profile memory improves over SlideTailor by +2.73 / +2.95 / +2.79 / +3.08 on Content / Structure / Visual / Specificity (3-model-family average)
• Best general-quality Avg (1–5 scale): MemSlides+GPT-5 4.17 vs DeepPresenter+GPT-5 3.99 / SlideTailor+GPT-5 3.60

Why it matters / caveats: Cleanly separates what-to-generate (user profile) from how-to-execute (tool memory) from what-is-active (working memory), providing a practical template for personalized agentic systems. Evaluation relies on a small controlled profile bank with 3 personas; real-world preference diversity and longer revision horizons may expose additional failure modes.

Problem: 3D brain MRI latent-diffusion pipelines use reconstruction-optimized encoder–decoder tokenizers that sacrifice clinical information in encoder embeddings, making the latent space unsuitable for both downstream clinical tasks and controllable generation simultaneously.

Method: BrainG3N uses a two-phase MAE–CNN tokenizer. Phase 1 pretrains a fully volumetric 3D MAE encoder (12-layer ViT, 1152 hidden dim, 1200 tokens/volume) on 35,309 brain MRI volumes from 18 public cohorts across four modalities (T1/T2/FLAIR/T1c) with 70% random patch masking; only 360/1200 patches enter the encoder per step. Phase 2 freezes the encoder, trains a linear projection P ∈ R^(1152×32) and a 3D CNN decoder under l1 reconstruction loss. A conditional flow-matching DiT (12 blocks, adaLN-Zero modulation, CFG s=2.0) is trained on the projected d'=32 tokens for 6-axis conditional generation (disease 8-way, sex, modality 4-way, acquisition site, age continuous, IDH1 binary). A longitudinal variant uses a vanishing-endpoint Brownian-bridge interpolant for intra-patient temporal forecasting.

Key results:

• Frozen encoder outperforms BrainIAC / BrainSegFounder / MedicalNet on 21/23 linear-probing tasks
• IDH1 mutation AUC 0.937, tumor grade AUC 0.921, brain-age MAE 4.43 y, sex AUC 0.967
• Generation controllability: age Pearson r=0.93, HC vs AD COND agreement 0.99
• Longitudinal forecasting: Pearson r=0.716, slope 0.268 (recovers ~27% of true aging magnitude)
• Pooled 3D-FID gen-vs-recons 34.4, gen-vs-real 107.3 (vs prior range 40–120)
• Training: ~200 H100-hours total; probing on 16-core CPU ~2 h/task

Why it matters / caveats: A single MAE encoder serving both clinical prediction and controllable 3D generation could enable privacy-preserving cohort sharing, counterfactual patient simulation, and zero-GPU-inference downstream probing for resource-constrained clinical sites. The longitudinal model recovers only ~27% of true aging magnitude; all findings are brain-MRI-specific and transfer to CT or other 3D modalities is untested.

Problem: No existing benchmark tests both long-term memory and persistent physical world-state maintenance for embodied household agents operating under partial observability across multi-day interactions.

Method: WorldLines constructs multi-day household traces from Habitat/HSSD scenes containing interleaved dialogue, human activities, robot actions, device operations, and project-driven long-horizon plans. Traces feed an evidence-linked task constructor that produces cutoff-controlled Memory QA (310 samples across six categories) and Embodied Task Planning samples. The proposed ObsMem framework maintains four typed stores: an append-only Event Track (raw observed/reported atoms gated by robot observability), a State Track (current snapshot + history per entity-attribute pair), a Belief Track (epistemic confidence decay based on intervening events), and a Commitment Track (future obligations). At query time, a query plan routes evidence retrieval across relevant views rather than running undifferentiated similarity search.

Key results:

• ObsMem: Judge 0.713, Perfect 69%, Event R@5 0.879, StateMH-J 0.762, StateMH-E 0.452 vs A-mem (0.575 / 53% / 0.839 / 0.540 / 0.216) and Mem0 (0.554 / 53% / 0.823 / 0.598 / 0.264)
• Ablation: removing evidence selector drops Judge by -0.264; removing world-state by -0.102; removing episode consolidation by -0.145
• Hidden-observer questions (6 samples): ObsMem 0.563 Hidden Judge vs w/o Belief 0.000 and w/o Selector 0.000

Why it matters / caveats: Surfaces a fundamental gap: text-retrieval memory systems cannot reliably distinguish direct observations from reported claims, or current states from overwritten ones, in embodied settings. ObsMem's typed-provenance architecture is a practical reference design. The benchmark is simulation-only (Habitat/HSSD); real-world sensor noise and unstructured environments remain untested.

Problem: Hierarchical VLA planners suffer from monocular pose hallucination during single-view 3D reconstruction and accumulate unverified, conflicting, or stale manipulation memories that degrade planning reliability.

Method: Two improvements to the GeneralVLA robot stack: (1) GeoFuse-MV3D, a conservative multi-view reconstruction branch on top of MV-SAM3D. It takes 5-view RGB-D + masks + poses and applies a VGGT geometry-prior source with appearance affine calibration, a provider-free axis-wise compensation source, soft visual-hull shrink for low-support points, and geometry-only alpha-blending (preserving color/opacity/SH). (2) Governed KnowledgeBank, replacing the append-only predecessor with structured records carrying quality score, confidence, lifecycle state, conflict links, and verifier metadata. Precision-oriented retrieval: S = r_text + confidence + success_prior + recency - conflict_penalty - stale_penalty; consolidation manages add/merge/replace/discard/summarize/archive under a fixed active-memory budget.

Key results:

• GeoFuse-MV3D on GSO-30 vs MV-SAM3D baseline: CD -2.20%, LPIPS -2.02%, PSNR +2.36%, SSIM +1.03%
• Governed KnowledgeBank vs ReasoningBank: Terminal-Bench 2.0 SR +4.53%, AS -4.95%; SWE-Bench Verified resolve rate +3.73%, AS -5.65%
• RLBench (14 tasks, Franka Panda): GeneralVLA-2 succeeds on all 14 tasks vs Hamster 10/14, VoxPoser 9/14, CAP 7/14
• Real-world (Agilex-2.0 Piper, 4 tasks): success rates 63.3% / 40.0% / 53.3% / 83.3% vs CAP near-zero on most

Why it matters / caveats: Decoupling geometry prior from appearance and treating manipulation memory as governed knowledge rather than a text cache are defensible engineering choices that yield consistent, if modest, improvements. Reconstruction gains over MV-SAM3D are small; the system requires calibrated multi-view inputs and reliable verifier signals — failures in either component propagate directly to planning.

Problem: Mask Diffusion Models (MDMs) cannot actively revisit and revise already-denoised tokens during generation, preventing iterative self-correction that could leverage the model's own uncertainty signals.

Method: Reflective Masking (RM) reformulates denoising as a per-position three-way decision at each step: keep, re-mask (if P(MASK) > P(current token)), or reveal (argmax over vocabulary). Activated via lightweight post-training on synthetic trajectories: positions are corrupted with MASK or wrong tokens (top-k MDM predictions), and position-wise oracle labels (reveal/mask/keep) supervise a combined cross-entropy loss. History Reference (HR) augments the input with a parameter-free per-position accumulated embedding using sinusoidal rotary History Embedding Rotation (HER), compressing the denoising trajectory into an O(1) per-step update with no learnable parameters. Training: ~5 hours on 2x NVIDIA H100 80GB GPUs.

Key results:

• Image editing (Lumina-DiMOO base): edit precision 99.73% vs SFT 71.84%; PSNR 34.76 dB vs 23.90 dB; VQAScore 85.17 vs 81.61; user preference 68.2% vs 53.3%
• Sudoku (0.81M-param MDM): RM+HR+decay+HER exact accuracy 93.4% vs RM w/o HR 82.4%; replay mistakes 0.03% vs 0.57%
• Text reasoning (LLaDA base): MATH500 24.8% (+2.4pp vs vanilla SFT), MBPP 39.4% (+8.8pp), ARC-Challenge 86.1% (+4.8pp)

Why it matters / caveats: Provides a post-training pathway to elicit localized self-correction from MDMs without architectural changes, establishing a native test-time scaling axis distinct from AR chain-of-thought. Text-reasoning gains are modest (+2-9pp); only validated on relatively small MDMs and gains on frontier-scale diffusion LMs are unknown.

Problem: Standard few-shot ICL for classifying B2B sales call segments degrades severely as concatenated examples grow, because long multi-party transcripts exhaust context budgets and long-context performance declines with scale.

Method: Rather than including raw examples in the prompt, an LLM generates distilled knowledge in one of two forms: Criteria-Ex (explicit positive/negative classification criteria) or Description-Ex (a coherent natural-language task description). These replace example concatenation entirely, reducing token usage by 99%. Iterative variants Criteria-De and Description-Cr chain the two representations. The Call Playbook dataset is introduced: 5 binary classification tasks from 50 English B2B sales calls (30-90 min), annotated by 3 in-house annotators. Evaluated across GPT-4o, Claude Sonnet 3.7, Claude Haiku 3, Mistral Large, and Mistral Small.

Key results:

• Standard Examples F1: degrades from 71.5% (0-shot) to 60.7% (100-shot)
• Description-Ex: stable 71.5%-72.6% across 0-100 shots; Criteria-De: 69.3% to 72.2%
• Macro-average AUC: Criteria-Ex 76.2%, Description-Ex 75.6% vs Examples 69.3%
• Extractive compression (SC): drops 71.4% to 62.1% (-9.3pp); LLMLingua-2 stays below 50%
• Best model: Claude Sonnet 3.7 at 77% macro-avg F1, peaking at 25 shots
• Cross-model distillation: criteria from larger models help smaller models; descriptions from larger models do not transfer as well

Why it matters / caveats: Shows that knowledge extraction is the key lever for robust few-shot performance under long-context pressure, with interpretable and human-editable classification logic as a bonus. Evaluation is limited to 50 calls from one commercial platform; generalizability to other industries and languages is unverified.

Problem: Feed-forward 4D head avatar generators and per-subject refiners cannot share a Gaussian representation end-to-end because per-subject refinement's adaptive densification destroys the upstream spatial layout, and existing feed-forward models hard-code source count and inherit single-dataset domain bias.

Method: SpatialAvatar-0 uses a shared FLAME-mesh-bound 3D Gaussian representation (~58K Gaussians/identity via 256x256 UV grid). The feed-forward stage uses a frozen DINOv3 ViT-B/16 + trainable DPT head feeding a StyleUNet UV generator (StyleGAN2-based) with five parallel attribute heads and a parameter-free K-source mean-pool (K in {1,2,3,4}). A FLAME-conditioned FiLM-modulated residual head provides expression/pose corrections. Two-phase training: Phase 1 monocular-temporal on CelebV-HQ (200K iters), Phase 2 multi-view-spatial on NeRSemble v2 with L2-SP anchor and 25% cross-time mix to prevent identity-prior collapse. The per-subject refinement (10K iters) freezes FLAME binding and Gaussian count, replacing densification with three-component anti-spike regularization: scale-freeze warmup (500 iter), hard log-scale clamp, and screen-space anti-anisotropy penalty on projected 2D covariances.

Key results:

• VFHQ zero-shot: 23.34 dB PSNR, 0.875 SSIM, 0.077 LPIPS vs in-domain leader GAGAvatar 21.83/0.818/0.122 (+1.5 dB PSNR)
• HDTF zero-shot: 24.67 dB PSNR, 0.916 SSIM, 0.076 LPIPS vs GAGAvatar 23.13/0.863/0.103 (+1.5 dB PSNR)
• SplattingAvatar per-subject: surpasses GeoAvatar (300K iter, 4.90 h) by +1.3 dB PSNR at 10K iter (~2 min) — 30x fewer iterations
• Feed-forward inference: ~50 ms/avatar on RTX 3090; per-subject renders at 265.6 FPS at 512^2
• Training: single H100 NVL, 14 days total

Why it matters / caveats: Achieves state-of-the-art cross-domain generalization without training on either test domain and reduces per-subject refinement from hours to minutes — a meaningful step toward deployable digital-human pipelines. Evaluation is limited to near-frontal portrait scenarios; heavy occlusion, accessories, and deformable regions are not covered.

Generated automatically from HuggingFace Daily Papers API + arXiv PDFs on 2026-06-22.