AI papers — 2026-05-23

Problem: Current AI systems cannot reliably forecast whether specific scientific advances will occur or when they will happen, even when provided with relevant prior knowledge.

Method: The authors introduce CUSP (Cutoff-conditioned Unseen Scientific Progress), a temporally grounded benchmark of 4,760 verifiable scientific milestones (Jan 2024–Mar 2026) drawn from Nature, Science, Cell, and community AI repositories. Each milestone generates four evaluation tasks: (1) binary feasibility prediction, (2) multiple-choice mechanistic reasoning over competing technical approaches, (3) free-response solution design, and (4) date prediction scored via exponential-decay metric. Free-response answers are evaluated by a GPT-5.4-mini LLM-as-judge augmented with web search that first detects temporal leakage then scores along four rubric dimensions (alignment, specificity, novelty, feasibility). Models evaluated include GPT-5.4, Claude S4.5, DeepSeek R1, LLaMA 3.3, GPT-4o, and GPT-OSS-20B.

Key Results:

• MCQ mechanistic reasoning: GPT-5.4 scores 0.819; all models above chance (0.25); AI sub-domain most predictable (0.461 date score), biology/chemistry lowest
• Binary feasibility: All models near chance (0.453–0.519, chance = 0.50), including the best frontier models
• Date prediction: All models show positive signed error (predict too late); LLaMA 3.3 achieves best score (0.500, median error +4 months); GPT-5.4 median error +14 months
• Free-response: GPT-5.4 achieves 5.04/10 and 60.3% pass rate; all other models ≤20% pass rate
• Performance is insensitive to pre- vs. post-training-cutoff events — models do not leverage memorized knowledge for forecasting
• With web-search access restricted to pre-cutoff information, gains are measurable but a large forecasting gap (∆fore) remains, growing with citation impact (∆fore = 0.875 for top-quartile papers)

Why It Matters / Caveats: CUSP reveals a fundamental gap between AI's ability to access and utilize scientific knowledge and its capacity for forward-looking prediction—more knowledge does not translate into better forecasting. The benchmark is continuously updated as new milestones are resolved, making it a living evaluation framework; however, evaluation of free-response tasks depends on a proprietary LLM judge (GPT-5.4-mini), and the benchmark currently skews toward English-language publications from high-impact venues.

Problem: Diffusion Transformers (DiTs) degrade when generating images beyond their training resolution because uniform RoPE attention scaling cannot simultaneously preserve global structure and fine-grained detail.

Method: SEGA (Spectral-Energy Guided Attention) is a training-free, parameter-free method that dynamically scales RoPE attention components per-dimension based on the spatial-frequency structure of the intermediate latent at each denoising step. At each step a 2D FFT of the channel-averaged latent yields axis-wise spectral profiles (height/width) and a radial profile; per-dimension corrections (standardized log-energy, zero-sum tanh redistribution) amplify scaling for under-resolved low-energy frequency bands and suppress it for high-energy bands. A global amplitude factor derived from spectral flatness (Wiener entropy) gates the overall correction strength—early in denoising (noisy latent, flat spectrum) the correction is near zero; later (structured content, peaked spectrum) it reaches full strength. The reference scale $m_\text{ref}$ follows a power-law in the resolution ratio. SEGA is applied via NTK RoPE extrapolation on both Flux and Qwen T2I architectures with no fine-tuning.

Key Results:

• Flux at 4096×4096: SEGA achieves best FID (150.05), ImageReward (1.26), CLIP Score (29.22), and CQA (0.725) vs. all baselines including multi-stage methods (HiFlow, I-Max, ScaleDiff) and direct methods (YaRN, DyPE, UltraImage)
• Qwen at 4096×4096: Best across all 8 metrics; IR 1.51, FID 148.26, MSQ 45.03, CQA 0.74 — at least +0.54 IR and −11 FID over strongest baseline
• Qwen at 2048×4096: IR 1.50 vs. next-best DyPE 1.07; FID 149.49 vs. DyPE 157.86
• Ablation confirms both axis-wise correction and global amplitude factor are necessary; NTK + fixed scaling is substantially worse
• Tested up to 6144×6144 (>36M pixels) successfully

Why It Matters / Caveats: SEGA resolves the long-standing coarse-vs-fine trade-off in training-free DiT extrapolation with zero learnable parameters and minimal code changes; the spectral-energy principle may extend to video generation and other bidirectional attention modalities. The hyperparameters κ=0.08 and γ=1.5 are fixed empirically and may require re-tuning for architectures with different RoPE configurations.

Problem: Process Reward Models (PRMs) require expensive step-level human annotations to train, limiting their scalability to new domains and model families.

Method: uPRM trains a PRM with no human supervision—neither step labels nor ground-truth final answers. The scoring function uses Qwen2.5-14B-Instruct's next-token probabilities over sequences that interleave reasoning steps with "+" (correct) and "−" (incorrect) markers: $S(j; s) = \log p^-_j + \sum_{t<j} \log p^+_t$ for candidate error position $j$. Multiple trajectories are scored jointly (concatenated in-context) so each trajectory is conditioned on preceding ones, leveraging in-context learning for more reliable assessments; a corner-budget correction term prevents degenerate all-same-label solutions. A LoRA-parameterized PRM (Qwen2.5-14B-Instruct base, two-layer MLP head, trainable [*] token embeddings) is trained via an entropy-regularized RL objective using an actor-critic gradient estimator with FlexAttention masks, on PRM800K reasoning trajectories without labels. Training takes ~5.5 hours on 8 H200 GPUs.

Key Results:

• ProcessBench F1 improvement over LLM-as-Judge baseline: +8.5% (GSM8K), +9.8% (MATH), +13.3% (OlympiadBench), +13.2% (Omni-MATH)
• Best-of-8 on MATH-500/Minerva/OlympiadBench: uPRM avg 60.1 vs. supervised baselines (Math-Shepherd, RLHFlow, Skywork, Qwen-Math-PRM) 60.0–60.8 — competitive despite zero supervision
• RL training (PURE framework): Qwen2.5-Math-1.5B trained with uPRM gains +4% average accuracy over verifiable outcome reward; Qwen2.5-Math-7B: +2.4% avg
• Reward hacking: Supervised PRM collapses in <50 iterations (Qwen2.5-Math-7B); uPRM completes training without hacking

Why It Matters / Caveats: uPRM eliminates the annotation bottleneck for process supervision and shows markedly better robustness to reward hacking during RL—a practical advantage independent of raw accuracy. The method requires a sufficiently capable and long-context base LLM for joint scoring, which limits the smallest usable models, and the precise mechanism behind uPRM's reward-hacking resistance is not yet fully explained.

Problem: Text-to-image reward models are either opaque scalar Bradley-Terry models requiring massive annotated datasets, or VLM rubric judges with manually designed criteria that don't reliably reflect human preferences, leaving both prone to reward hacking.

Method: AutoRubric-T2I is an iterative rubric learning framework using only 256 preference pairs. Starting from diversity-aware seed data selection (proxy reward margin + prompt clustering) and initial VLM-based CoT rubric generation (Gemini-3-Flash), the framework iterates: (1) score all candidate rubrics on training pairs using a VLM judge (yes-token probability); (2) solve an ℓ1-regularized logistic regression over pairwise rubric-score differences to select the top-N discriminative rubrics with positive weights; (3) mine curriculum-bucketed hard pairs (small-margin, large-margin, high-reward-wrong) to generate new failure-driven rubrics via VLM failure diagnosis. The final reward is a weighted sum of VLM binary yes-token probabilities. Downstream RL uses Flow-GRPO on SD-3.5-Medium with Qwen3-VL-8B as judge.

Key Results:

• MMRB2 (OOD): Qwen3-VL-8B pointwise baseline 26.5% → AutoRubric-T2I 62.5% (vs. HPSv3 59.4%, UnifiedReward 59.8%)
• Gemini-3-Flash judge on MMRB2: 71.4% overall, matching raw pairwise mode (70.8%)
• TIIF short prompts (RL): SD-3.5-Medium 65.3% → 71.6% (HPSv3 rubrics); 70.8% (PickScore rubrics)
• UniGenBench++ long prompts: 64.0% → 67.7% (PickScore rubrics)
• Human evaluation (30 annotators, 20 prompts): AutoRubric-T2I selection rate 44.8% vs. 25% random (p<0.001)
• Uses <0.01% of the annotated preference data compared to full reward model training

Why It Matters / Caveats: AutoRubric-T2I shows that a compact set of learned rubrics from as few as 256 preference pairs can match fully trained scalar reward models while remaining interpretable and less susceptible to reward hacking. The framework depends on a capable proprietary VLM for rubric generation (Gemini-3-Flash), and rubric quality may degrade for highly specialized visual domains not covered in the seed data.

Problem: Cross-entropy loss and direct downstream evaluation are both inadequate signals for LLM development decisions—the former is task-agnostic, the latter is expensive and uninformative for near-zero accuracy models at early training stages.

Method: Proxy metrics are computed in a single forward pass by running expert-written reasoning trajectories through the candidate model and extracting 80 combinations of 10 core token-level statistics (cross-entropy, top-k accuracy for k∈{1,2,3,5}, entropy, rank, reciprocal rank, margin, wrong-confidence) under 8 weighting schemes (uniform, probability, expert-disagreement, entropy, inverse entropy, frequency, inverse frequency, Gaussian-NLL). These are aggregated as weighted averages over the last 1,000 tokens per trajectory. For cross-family model ranking, a linear RankSVM trained on held-in tasks predicts model orderings. Expert trajectories come from Kimi-K2.5, MiniMax-M2.5, Qwen3-Next-80B. Evaluation spans 18 models (0.6B–70B) on AIME2025, HMMT, GPQA, USACO, MMLU-Pro, SuperGPQA.

Key Results:

• Cross-family model ranking: Linear RankSVM achieves Spearman ρ = 0.81 (leave-2-tasks-out CV) vs. ρ=0.36 for FineWeb CE loss, ρ=0.33 for rBridge
• Pretraining data selection (DataDecide, 25 corpora): Frequency-weighted top-5 accuracy reaches >0.85 decision accuracy at ~10⁻⁵ of target 1B compute — matching direct evaluation at 10,000× less compute and pushing beyond rBridge's Pareto frontier
• Training-time forecasting: Proxy metrics extrapolate downstream accuracy across an 18× compute horizon with ~50% less error than CE loss or compute
• Ranking signal concentrates on inverse-frequency-weighted top-1 accuracy and entropy-weighted entropy

Why It Matters / Caveats: Expert-trajectory proxies are architecture-agnostic, require no expert logprobs (unlike rBridge), and work with closed-weight frontier models or human experts as trajectory sources. The approach requires access to expert solutions for the target domain; for frontier research tasks without existing solutions, trajectory quality may be a bottleneck.

Problem: Foundation models and LoRA adapters deployed on edge devices are vulnerable to model recovery attacks (e.g., Spectral DeTuning), but existing protection methods require retraining or full plaintext decryption at inference time.

Method: LoREnc (Low-Rank Encryption) is a training-free, data-independent framework. For each FM weight matrix W, truncated SVD extracts the top-Δr singular components as a spectral key L = U_FM Σ_FM V_FM^T; the protected FM weight is W̃ = W − L. Authorized LoRA adapters absorb L via rank expansion: B̃_k = [B_k, U_FM Σ^{1/2}], Ã_k = [A_k, Σ^{1/2} V_FM^T], ensuring exact recovery W̃ + B̃_kÃ_k = W + B_kA_k. The expanded adapter is then re-factorized via SVD and split into encrypted weights (rank-r) and restoration keys, followed by orthogonal reparameterization with a random rotation matrix M_k to destroy the SVD structural fingerprint that would expose the protection scheme. Restoration keys are stored in a hardware TEE; authorized users reconstruct on-the-fly during the forward pass.

Key Results:

• Unauthorized access (SD 1.5, 5 downstream tasks): ΔCLIP ≈ −0.148–−0.155, LPIPS ≈ 0.827–0.870 (near-random outputs)
• Authorized access: ΔCLIP = 0.000, LPIPS = 0.000 (exact recovery on all tasks)
• GPT-2: Foundation perplexity 120 → downstream 0.000; LLaMA 3: 8793 → 0.000
• Fine-tuning resilience (100k samples): LoREnc CLIP 0.231 vs. baseline 0.267 and NNSplitter 0.251 — strongest protection at all data sizes
• Computational overhead (smartphone): <0.2% parameters, 0.40% GFLOPs, 0.22% inference latency

Why It Matters / Caveats: LoREnc is the first training-free, data-independent FM protection with exact authorized performance recovery, suitable for on-device deployment where full-model decryption is infeasible. It provides empirical rather than formal cryptographic guarantees; adaptive detectors specifically designed for orthogonal adapter structures are outside the current threat model.

Problem: Agentic LLMs trained end-to-end with RL increase reasoning token length dramatically during training without reliable accuracy improvements because planning is unstructured and unregulated.

Method: SR2AM (Self-Regulated Simulative Reasoning Agentic LLM) decomposes agentic reasoning into three systems: (I) reactive execution (free-form CoT + tool calls), (II) simulative planning (the LLM acting as its own world model, generating explicit plans of proposed actions and predicted future belief states), and (III) a configurator that decides per-turn whether to make a new plan, continue an existing one, or skip planning. Two instantiations: v0.1 uses multi-module prompted inference with o4-mini to generate training data; v1.0 reconstructs configurator decisions and structured plans from DeepSeek-V3.2 reasoning traces via an annotator LLM, preserving original free-form reasoning while adding planning structure. Both versions are trained via SFT then RLOO-based RL with a combined answer + format + plan-quality reward. SR2AM-v0.1-8B and SR2AM-v1.0-30B are the final models.

Key Results:

• SR2AM-v0.1-8B achieves Pass@1 competitive with systems at 120–355B parameters on math, science, tabular analysis, and web tasks
• SR2AM-v1.0-30B achieves Pass@1 competitive with systems at 685B–1T parameters
• SR2AM-v1.0-30B uses 25.8–95.3% fewer reasoning tokens than competitive agentic LLMs of similar scale
• RL increases average planning horizon by 22.8% while planning frequency grows only 2.0 pp — model learns to plan further ahead rather than more often

Why It Matters / Caveats: Structured self-regulation of when and how deeply to plan yields comparable accuracy to models 10–30× larger with a fraction of the token budget. The simulative planning structure can be reconstructed from existing reasoning traces without new architecture, making the approach compatible with any capable pretrained LLM. Performance and token savings depend on the quality of the world model (the LLM itself), which may fail on highly novel interactive environments.

Problem: Translating natural language chip design rules into executable DRC scripts is labor-intensive but lacks large-scale benchmarks with execution-based evaluation, and existing methods ignore DRC execution feedback or require labeled test layouts as input.

Method: Rule2DRC provides 1,000 NL-to-DRC script tasks: 310 rules derived from the SkyWater130 PDK paired with reference scripts and corner-case layouts, plus 690 synthetic multi-constraint rules covering complex modern-node scenarios. Evaluation uses 13,921 chip layouts in GDSII format run through the KLayout DRC engine; correctness is measured by comparing violation outputs against ground-truth scripts—not code similarity. SplitTester is a tester agent that (1) generates initial test layouts, (2) clusters candidate scripts by identical execution output patterns, (3) adaptively generates additional tests to split previously indistinguishable clusters, and (4) uses a judge LLM to select among top-3 candidates. This contrasts with prior work (CodeMonkey: top-3 filter only; S*: samples from already-separated clusters) by explicitly targeting the largest indistinguishable groups.

Key Results:

• Rule2DRC is >10× larger than all prior DRC script benchmarks (<200 tasks)
• Execution-based scoring correctly identifies equivalent implementations that surface-level code similarity fails to equate (e.g., separation vs. sized vs. sep aliases)
• SplitTester outperforms CodeMonkey and S* under Best-of-N selection on Rule2DRC
• Full benchmark, pipeline, and code released open-source

Why It Matters / Caveats: Rule2DRC enables rigorous functional evaluation of DRC script synthesis agents and demonstrates that execution feedback substantially improves candidate selection via SplitTester. The benchmark currently uses KLayout DSL and SkyWater130-derived rules; coverage of proprietary process nodes and other EDA tools (SVRF, Calibre) remains future work.

Problem: Existing methods for attributing AI contributions in human-AI collaboration focus only on final artifacts (text watermarking, stylometric analysis), missing how AI shapes goals, requirements, and design decisions throughout the collaboration process.

Method: COTRACE is an automated goal-level attribution pipeline decomposing collaboration logs into hierarchical goals → requirements → atomic actions. Four LLM-as-judge stages: (1) outcome and action extraction (roles: SHAPER, EXECUTOR, OTHER); (2) requirement extraction with CREATE/REVISE/DELETE versioned history; (3) indirect influence labeling via embedding-similarity pre-filtering + LLM judge (DIRECT, IMPLICIT, NO CONNECTION); (4) contribution aggregation into speaker × role matrices. Applied to 638 real-world ShareChat logs across Computer Programming, Data Analysis, Writing, and Planning. A COTRACE-viewer exposes analyses to users in a 10-participant study.

Key Results:

• Models account for 11–26% of goal-shaping (SHAPER mass) but 96–99% of execution — humans set direction, models execute
• LLM contributions to goal shaping increase with specificity: larger role in child goals and individual requirements than parent goals
• In technical tasks (Comp. Prog., Data Analysis), LLM requirement generation eventually surpasses users'
• Exposing users to COTRACE analysis shifts perceived AI contribution by nearly 2 points on a 5-point scale
• Interaction design (e.g., requiring AI communication before action) significantly amplifies model goal-shaping
• 11 recurring indirect influence patterns identified; manual validation achieves >90% accuracy

Why It Matters / Caveats: COTRACE provides the first principled, quantitative measurement of goal-level AI contributions across dialogue turns, with direct implications for educational integrity, creative attribution, and AI collaboration governance. The framework relies on LLM judges for decomposition and influence labeling, and validation was limited to 10 participants in the user study, leaving larger-scale calibration open.

Problem: Existing benchmarks for proactive streaming video understanding rely only on visual signals, use polling or fixed-timestamp protocols rather than true proactive evaluation, and cover too few task types to differentiate omni-modal models.

Method: OMNIPRO is a benchmark comprising 2,700 human-verified samples across 9 sub-tasks (Alert, Monitoring, Grounding, Counting, Narration, Prediction, and compound variants) organized into 3 cognitive levels covering 6 basic video understanding capabilities. 84% of samples require audio (speech or non-speech sound); each sample carries modality-isolation labels for ablation. Two evaluation modes: Probe (queries model before/after ground-truth trigger timestamps, measuring content understanding without requiring streaming capability) and Online (model autonomously decides when to respond in continuous streaming input). 11 representative models evaluated including open-source and proprietary systems.

Key Results:

• Audio provides +2.4 to +11.1 accuracy gains across models, but utilization is highly variable
• Models retain on average only 37% of early-segment performance at later video positions (long-horizon degradation)
• Non-speech sound perception is the weakest dimension across all 11 models
• OMNIPRO covers 6/6 basic capabilities vs. 1–2/6 for prior benchmarks (StreamingBench-Pro, OVO-Bench-Pro, OmniMMI-Pro)
• 2,700 samples across 1,262 videos (avg 189s), with avg 3.4 responses per question

Why It Matters / Caveats: OMNIPRO is the first benchmark to jointly require omni-modal perception, autonomous proactive response timing, and diverse video tasks, revealing long-horizon temporal degradation and weak non-speech audio as key open problems. The benchmark is limited to pre-collected static videos rather than real-time sensor streams, and the Online mode evaluation protocol may be sensitive to timing measurement precision.

Problem: Audio diffusion music models use bidirectional attention incompatible with KV caching, making them computationally unable to run as interactive streaming generators on consumer hardware.

Method: LMDMs (Live Music Diffusion Models) modify block-wise outpainting diffusion by routing clean context history and noisy target frames through separate projections with dedicated causal attention masks, enabling noise-level KV caching over diffusion steps. Two variants: Encoder-Decoder (KV-caches clean context across denoising steps) and Block-Causal (additionally caches across time, achieving strictly lower FLOPs than encoder-decoder LMMs). Fine-tuned from existing open-source audio diffusion models in <8 GPU hours. ARC-Forcing combines Self-Forcing (rollout-based training on real multi-block generation sequences) with ARC adversarial post-training (global distribution-matching loss on rollouts) to provide stable long-horizon supervision without RL or reward models. Deployed as ONNX + C++/JUCE on a consumer gaming laptop with sketch-based and text-based controls.

Key Results:

• Recovers and surpasses the exact inference complexity of discrete-AR Live Music Models (LMMs)
• Enables stable minute-long musical rollouts vs. error accumulation in standard bidirectional diffusion
• Fine-tuning cost: <8 GPU hours from an existing checkpoint
• Demonstrated in real artist-AI collaboration as a "generative delay" (live performance deployment)
• Supports text, gestural sketch, and accompaniment conditioning inherited from offline diffusion

Why It Matters / Caveats: LMDMs are the first demonstration that open-source audio diffusion models can match discrete-AR inference efficiency for live performance with minimal training cost; ARC-Forcing avoids RL overhead while stabilizing long rollouts. Absolute audio quality relative to the largest discrete-AR LMMs (requiring 40+ GB VRAM) was not directly compared; consumer-hardware performance depends on model scale.

Problem: Existing 3D animal reconstruction methods focus on single-animal settings with pre-cropped inputs and cannot handle multi-animal scenes with mutual occlusion and complex interactions.

Method: SAM 3D Animal is a promptable multi-instance 3D reconstruction framework built on the SMAL+ parametric animal model (145-dimensional shape space covering 145+ species, 35 joints). A ViT-Huge encoder feeds a SAM-style promptable Transformer decoder that simultaneously predicts SMAL+ parameters, cameras, and bounding boxes for up to P=30 animal instances via DETR-style bipartite Hungarian matching (no NMS). Query tokens include pose, bounding box, 2D/3D keypoint, and prompt tokens; a layer-wise keypoint feedback loop iteratively refines geometric predictions via positional embedding injection. Prompts (keypoints or masks) enable explicit disambiguation in crowded scenes. Herd3D (>5K multi-animal images with per-instance SMAL+ ground-truth meshes, diverse species/interactions/occlusions) is introduced to provide 3D training supervision for multi-animal scenes.

Key Results:

• Animal3D: 5.2 PA-MPJPE improvement over the strongest baseline
• Animal Kingdom (OOD): up to 54% AP gain and 80% mAP gain over strongest baseline with keypoint prompts
• APTv2: competitive or superior even without prompts
• Keypoint prompts dominate over mask prompts; performance scales monotonically with number of keypoints
• Herd3D brings consistent improvements, especially on multi-animal benchmarks

Why It Matters / Caveats: SAM 3D Animal is the first promptable multi-animal 3D reconstruction system, and Herd3D fills a critical data gap for multi-instance 3D supervision. The SMAL+ model currently covers quadruped species, limiting applicability to birds, fish, or reptiles; Herd3D uses synthetic generation (GenZoo pipeline), so sim-to-real gaps may affect out-of-distribution species.

Problem: The Platonic Representation Hypothesis demonstrates convergence in AI systems, but it is unknown whether human brains exhibit analogous shared neural geometry recoverable without paired cross-subject data or model representations as anchors.

Method: Subject-specific fMRI encoders are trained on the Natural Scenes Dataset (NSD) using repeated stimulus presentations as self-supervision: voxel reliability weighting → PCA → multi-view CCA (MCCA) → residual nonlinear refinement (MLP) trained with InfoNCE + cosine pull loss over view pairs. Pairwise brain-to-brain translation uses an unsupervised orthogonal rotation (Procrustes problem on geometry-derived pseudo-correspondences from K-means centroid matching + iterative closest point refinement) restricted to O(d) to preserve distances. Pairwise rotations are synchronized into a single shared latent space via global alignment. No shared stimuli, paired cross-subject data, or external model features are used during training; evaluation uses held-out shared images.

Key Results:

• Independently learned subject fMRI embeddings are approximately isometric: simple orthogonal rotations recover accurate instance-level cross-subject correspondences
• Synchronizing pairwise rotations into a shared space further improves cross-subject retrieval, confirming mutual compatibility with a common coordinate system
• Demonstrates the existence of a shared neural geometry in the human visual cortex recoverable from fMRI data alone

Why It Matters / Caveats: This work extends the Strong Platonic Representation Hypothesis to biological neural systems, suggesting that human visual cortex representations across individuals share an approximately isometric geometry translatable by pure geometric transformations. Results are limited to the NSD dataset (8 subjects, natural images) and to visual cortex; generalization to other modalities, brain regions, or larger populations remains to be explored.

Problem: Wearable IMU signals are tightly coupled to sensing setup (body location, mounting, orientation, hardware, sampling rate), preventing motion representations from transferring across devices and datasets.

Method: AnyMo is a geometry-aware, setup-agnostic wearable motion framework with four components: (1) Physics-grounded geometry-aware IMU simulation over dense body-surface placements on the Nymeria body mesh — for each anatomical segment, candidate surface vertices define sensor frames (tangent/binormal/normal) with in-plane rotation sampling to model mounting variation; virtual IMU trajectories are computed from body kinematics; (2) Setup-agnostic graph encoder pre-training via paired synthetic placement views and masked partial observations (contrastive learning); (3) Full-body IMU tokenizer mapping multi-position IMU observations from 23 anatomical segments into compact full-body motion tokens; (4) IMU-language alignment via LLM for open-vocabulary zero-shot recognition, cross-modal retrieval, and motion captioning. Evaluated on 14 unseen downstream datasets (zero-shot HAR), Nymeria sim-to-real retrieval, and EgoExo4D zero-shot captioning.

Key Results:

• Zero-shot HAR (14 unseen datasets): avg +11.7% Accuracy, +11.6% F1, +22.6% R@2
• IMU-to-text retrieval MRR: +15.9%; text-to-IMU retrieval MRR: +28.6%
• Zero-shot captioning BERT-F1: +18.8% on EgoExo4D

Why It Matters / Caveats: AnyMo is the first generalist wearable motion model combining physics-based surface simulation, setup-agnostic pre-training, full-body tokenization, and LLM alignment. Simulation uses the Nymeria body model, so rare body shapes or very unusual mounting positions may not be well-covered; large-scale real-world IMU data remains fragmented across datasets with heterogeneous annotation formats.

Problem: LLM-generated Lean proofs are correct but verbose and brittle across Lean/Mathlib version updates, and prior refactoring work ignores compilation cost and version compatibility as co-equal objectives.

Method: Lean Refactor is a retrieval-augmented agentic framework with a densely annotated strategy bank of >200K long-short proof pairs and 9K refactoring strategies (sourced from NuminaMath-1.5, FineLeanCorpus, Mathlib, ATLAS; proofs synthesized via Goedel-Prover-V2-32B and GPT-OSS-120B). Each strategy entry is annotated with: expected compilation-cost reduction (profiled via lake env lean --profile) and Lean/Mathlib version compatibility (v4.14, v4.16, v4.22, v4.24). A retrieval + multi-objective rerank pipeline filters strategies by user-specified trade-offs. A frozen agentic LLM (Gemini, Claude, or GPT) runs a planner-refactorer-debugger loop without any fine-tuning, enabling plug-and-play use across LLM families.

Key Results:

• >70% token-level compression on competition benchmarks (miniF2F, PutnamBench, Putnam2025)
• >20% compression on research repositories (PFR, PhysLean, Analysis, FLT)
• Up to 60% compilation-time reduction (>30% consistent)
• Outperforms Claude Code and all prior Lean refactoring methods
• Refactored miniF2F proofs type-check on more future Lean/Mathlib releases than unrefactored counterparts

Why It Matters / Caveats: Lean Refactor addresses three simultaneously neglected challenges—multi-objective optimization, version compatibility, and scarce paired data—without LLM fine-tuning, making it maintenance-friendly as Lean/Mathlib evolves. The strategy bank must be periodically re-profiled against new Mathlib releases; research-level compression (>20%) lags behind competition-level (>70%) due to dense intra-project dependencies.

Problem: Representation Autoencoders (RAEs) using frozen Vision Foundation Models (VFMs) as encoders lose fine-grained pixel details (color, texture) because VFMs are trained for semantic invariance, but injecting low-level signals via fine-tuning or feature concatenation degrades generative quality.

Method: DecQ adds a small set of learnable detail-condensing queries that attend to intermediate VFM layer features via cross-attention condenser modules (multiple residual cross-attention blocks), extracting low-level reconstruction information complementary to the VFM semantic space while keeping the VFM fully frozen. The condenser queries are appended to the ViT decoder alongside the standard patch tokens, and crucially, they are jointly denoised with patch tokens during flow-matching diffusion training (the generative model predicts both patch tokens and detail queries). A key finding: shallow-layer VFM queries primarily benefit reconstruction while deep-layer queries primarily benefit generation. With DINOv2 backbone + DiTDH-S on ImageNet 256×256.

Key Results:

• 8 additional queries, 3.9% extra compute: PSNR improves from 19.13 dB → 22.76 dB over frozen RAE baseline
• FID without guidance: 1.41; with guidance: 1.05 (vs. RAE baseline)
• 3.3× faster convergence than RAE
• Unlike VFM-finetune / VFM-distill / VFM-feat-concat (all trade reconstruction vs. generation), DecQ improves both simultaneously

Why It Matters / Caveats: DecQ resolves the reconstruction-generation trade-off in RAEs with minimal overhead by decoupling low-level detail extraction from the frozen semantic space; jointly denoising detail queries with patches is the key insight that benefits generation as well. Results are demonstrated on DINOv2 at 256×256; scaling to larger resolutions and other VFM backbones remains to be validated.

Problem: In class-imbalanced CT body composition segmentation, the apparent benefits of episodic sampling over standard sampling strategies are confounded by uncontrolled differences in training iteration budgets.

Method: Three sampling strategies are compared under matched conditions on 9 tissue classes (SAROS dataset, 210 CT scans, T1–L4 vertebral range): random (uniform), weighted (inverse rare-class frequency per slice), and episodic (class-structured mini-batches from few-shot learning, adapted to fully supervised training). Architecture, loss function, and optimizer are held fixed across conditions. Two regimes: full-data (210 scans) and low-data (21 scans, 10% subsampling). Training iteration budgets are explicitly matched to isolate sampling effect from scheduling artifacts. Network: nnU-Net-style architecture with DiceCE loss.

Key Results:

• Full-data: Episodic 0.882 Dice vs. random 0.878, weighted 0.878 (marginal improvement)
• Low-data (epoch-matched): Episodic 0.787 vs. random 0.758, weighted 0.762 — driven by a 12-fold difference in training iterations per epoch
• Low-data (iteration-matched): Random/weighted overfit earlier; episodic improves for ~3× more iterations before plateauing
• Residual advantage of episodic sampling after iteration matching suggests implicit regularization from class-balanced batch composition

Why It Matters / Caveats: Identifies training iteration budget as an under-recognized confound in sampling strategy comparisons; the implicit regularization from episodic sampling offers a low-cost, model-agnostic benefit for small medical imaging datasets with severe class imbalance. Results are specific to CT body composition segmentation; generalization to other anatomical tasks or modalities requires further validation.

Problem: Fashion image retrieval requires a single system handling diverse query formats (image, text, sketch, video, multimodal) and retrieval intents (similarity, compatibility, attribute-specific), but existing models are task-specific and suffer from feature interference and training imbalance across heterogeneous datasets.

Method: FashionLens is an MLLM-based unified retrieval framework trained on U-FIRE — a new benchmark consolidating 15 fragmented fashion datasets into 400K+ samples across 9 existing retrieval tasks plus 2 newly proposed OOD tasks. Two novel components: (1) Proposal-Guided Spherical Query Calibrator (PGSQC) performs adaptive spherical linear interpolation (slerp) to rotate query representations into task-aligned metric spaces, resolving feature interference between similarity-oriented and compatibility-oriented matching objectives without task-specific heads; (2) Gradient-Guided Adaptive Sampling (GGAS) dynamically re-weights tasks during training based on real-time gradient norm estimates (learning difficulty proxy) combined with dataset size priors, preventing under-optimization of rare or difficult tasks.

Key Results:

• State-of-the-art across diverse retrieval scenarios on U-FIRE
• Robust generalization to unseen OOD retrieval tasks (2 new test splits)
• PGSQC resolves negative transfer from divergent matching objectives
• GGAS prevents under-optimization of data-sparse or complex tasks
• Handles image, text, sketch, video, and multimodal query inputs in a single model

Why It Matters / Caveats: FashionLens is the first truly versatile fashion retrieval framework handling the full spectrum of realistic user search demands in a single model, and U-FIRE consolidates a fragmented benchmark landscape. Performance on very rare retrieval tasks with minimal training data may still be limited; the MLLM backbone introduces substantial inference cost compared to lightweight specialized retrievers.

Report generated from HuggingFace daily papers API + arxiv PDF extraction · 18 papers · UTC 2026-05-24