AI papers — 2026-06-12

Problem: Existing LLM agent benchmarks assume static environments, leaving agents poorly evaluated and equipped for real-world settings where tools, codebases, and user preferences change over time.

Method: The authors introduce EvoArena, a benchmark suite with three evolving domains — Terminal-Bench-Evo (89 tasks expanded to 352 versioned instances across 5-version chains), SWE-Chain-Evo (493 chain-step instances across 50 software evolution chains), and PersonaMem-Evo (505 preference-inference questions over 10 persona conversations). To address the benchmark's demands, they propose EvoMem, a patch-based memory augmentation that appends non-additive memory updates as tuples capturing pre/post content, rationale, and temporal metadata, then retrieves both current state and historical patches at inference time. EvoMem is implemented on top of four existing agent memory systems (Terminus2, OpenHands, A-Mem, Memento-Skill) without replacing them.

Key results:

• Overall average accuracy on EvoArena: 39.6% across 8 backbone models
• EvoMem chain-level improvement: +6.1% on Terminal-Bench-Evo (21.5% → 27.6%), +2.1% on SWE-Chain-Evo (10.0% → 12.1%), +3.2% on PersonaMem-Evo (40.0% → 43.2%)
• GAIA standard benchmark: +6.5% improvement (65.8% → 72.3%)
• LoCoMo standard benchmark: +3.3% improvement (39.7% → 43.0%)
• Temporal trajectory questions see strongest gains (+5.2%) within PersonaMem-Evo
• Pass-to-Pass regression rates on SWE-Chain-Evo drop from 9.09% to 6.32% (−2.77 pp)

Why it matters / caveats: EvoArena fills a real gap by explicitly measuring agent degradation across evolving environments, and EvoMem's patch-augmented retrieval demonstrates consistent (if modest) gains across diverse agent backbones. Absolute performance remains low (39.6% average), and SWE-Chain-Evo chain-level scores barely reach 12%, indicating substantial headroom; gains are incremental rather than transformative.

Problem: Vision-language models struggle at spatial reasoning tasks because current tool-call interfaces (single-pass code or structured JSON/XML) prevent iterative refinement based on intermediate perception results.

Method: SpatialClaw wraps a stateful IPython kernel pre-loaded with perception tools (Depth Anything 3 for reconstruction, SAM3 for segmentation) and scientific libraries (NumPy, SciPy, Matplotlib), and drives it through a five-stage agentic loop: separate planning, code generation, static AST validation, execution with feedback assembly, and answer submission. The agent writes one Python cell per step conditioned on prior outputs, enabling iterative composition of depth maps, segmentation masks, and numerical primitives. A single set of unified prompts is applied across all 20 benchmarks and 6 VLM backbones without task-specific templates.

Key results:

• Average accuracy across 20 spatial benchmarks: 59.9% (SpatialClaw) vs. 53.4% (no-tool baseline), +6.5 pp
• Outperforms SpaceTools Toolshed by +11.2 pp (52.0%), pySpatial by +11.2 pp (48.7%), VADAR by +12.1 pp (47.8%)
• Video spatial / 4D reasoning: +18.3 pp on DSI-Bench; multi-view: +14.3 pp on MindCube
• Action interface ablation (Gemma4-31B): SpatialClaw 59.9% vs. single-pass code 55.2% vs. structured tool-call 56.7%
• Attribution: >50% of wins over structured tool-call from code composition; 19.5% from control flow; 28.3% interface-neutral

Why it matters / caveats: The stateful kernel design is a practical and portable improvement applicable to any underlying VLM without fine-tuning. Single-image spatial reasoning shows only modest gains (+0.8 pp), so the approach's advantage is concentrated in video and multi-view tasks where iterative depth estimation provides the most benefit.

Problem: Existing training data for deep search agents contains exploitable shortcuts — evidence co-coverage, single-clue selectivity, exposed constants, and prior-knowledge binding — that allow agents to succeed without performing genuine multi-hop retrieval.

Method: FORT (Framework of Shortcut-Resistant Training-Data Synthesis) is a four-stage pipeline: (1) graph initialization using long-tail Wikidata entities and pre-mined cycle structures to block prior-knowledge shortcuts; (2) graph construction via multi-source enrichment (Wikidata, web, Scholar) with derived facts and generic-fact selection to avoid over-identifying clues; (3) question formulation that withholds intermediate entity names and applies five fuzzing strategies (category generalization, range relaxation, meta-attribute description, arithmetic encoding, contrastive exclusion); and (4) adversarial refinement that runs a strong search agent against drafts and repairs shortcut-prone cases. The resulting data trains Qwen3-30B-A3B-Thinking via SFT only (no RL), 6 epochs, batch size 64, context 256K tokens.

Key results:

• BrowseComp: 72.2% (vs. MiroThinker 67.9%, +4.3 pp); BrowseComp-ZH: 75.0%; xbench-DeepSearch-2505: 80.8%; Seal-0: 46.0% (vs. 41.4%)
• Overall average: 66.2% vs. 64.6% for best comparable open-source agent
• FORT training trajectories average 141.0 turns vs. 92.1 for REDSearcher (+53%), with answer hit at step 46.9 vs. 18.7 (+151%)
• Ablation: removing fuzzing collapses difficulty (accuracy 29.0% → 81.6%, turns 141.9 → 43.7)
• Model uses only ~3B active parameters; training data = 12K examples

Why it matters / caveats: Demonstrating that shortcut resistance in training data — not just trajectory length — matters for benchmark performance is a concrete, replicable insight. The SFT-only recipe with 3B active parameters matching larger models is notable; comparison with frontier closed models is limited.

Problem: Unified multimodal models that interleave text and image generation suffer from visual over-reliance and step-wise error accumulation when generating sequences requiring coordinated image editing across multiple steps.

Method: InterleaveThinker is a multi-agent pipeline consisting of a Planner agent (Qwen3-VL-8B-Instruct) that generates N-step execution plans, a frozen image generator (FLUX.2-klein or Qwen-Image-Edit), and a Critic agent (Qwen3-VL-8B-Instruct) that evaluates outputs and refines prompts for regeneration. Training uses three datasets — Interleave-Planner-SFT-80k, Interleave-Critic-SFT-112k, and Interleave-Critic-RL-13k — with SFT (2 epochs, lr 2×10⁻⁵) followed by single-step GRPO reinforcement learning (1 epoch, lr 2×10⁻⁶) using dual rewards: an accuracy reward penalizing judgment errors and a step-wise reward measuring inter-iteration improvement. Total training: ~50 hours on 8 H800 GPUs.

Key results:

• UEval benchmark: 66.3 (with FLUX.2-klein), 67.2 (with Qwen-Image-Edit) vs. Janus-Pro baseline 22.9
• WISE reasoning score: 0.73 vs. FLUX.2-klein baseline 0.47 (+55% relative)
• RISE editing score: 28.9 vs. FLUX.2-klein baseline 13.3 (+117% relative)
• CoMM image quality: 9.7/10 with FLUX.2-klein

Why it matters / caveats: By decoupling interleaved generation into modular agents and training with single-step RL rather than full trajectory RL, the framework sidesteps the complexity of credit assignment across 25+ generator calls. The approach is model-agnostic (tested with two different frozen generators), though all experiments use the same Qwen3-VL-8B backbone for both agents, leaving cross-architecture generalization untested.

Problem: Multimodal large language models are brittle under real-world and adversarial image corruptions because they lack an explicit visual recovery mechanism before reasoning.

Method: Robust-U1 extends BAGEL (a unified understanding-and-generation MLLM) through three sequential training stages: (1) SFT on ImageNet-C using rectified flow with L2 noise-prediction loss to learn image reconstruction; (2) Flow-GRPO reinforcement learning with dual rewards — pixel-level SSIM (ℛ_pix) and exponentially-transformed CLIP embedding cosine similarity (ℛ_sem) — sampling G trajectories per image with KL divergence penalty to prevent reward hacking; and (3) multimodal reasoning SFT that trains the model to jointly attend to both the corrupted and the recovered image alongside the query.

Key results:

• R-Bench overall score: 0.7398 vs. Robust-R1 prior SOTA 0.5017 (+47.4% absolute)
• MMMB benchmark: only 1.57-point accuracy drop from clean to 100% corruption (vs. 6.06 for Robust-R1)
• Visual recovery PSNR: 21.49 (full model) vs. 20.88 (SFT only) vs. 14.37 (BAGEL baseline)
• SSIM: 0.6314 (full) vs. 0.6135 (SFT only) vs. 0.4722 (baseline)
• Ablation: removing multimodal reasoning stage costs −11.6% on R-Bench (0.6623 vs. 0.7398)
• MMStar (100% corruption): 63.87%; RealWorldQA (100% corruption): 67.46%

Why it matters / caveats: Explicit visual recovery as a preprocessing step before reasoning substantially outperforms prior approaches without requiring architectural changes beyond the training pipeline. Training data is limited to ImageNet-C corruption patterns, so generalization to out-of-distribution corruption types not covered during training is an open question.

Problem: Generating correct competition-level mathematical proofs (IMO, USAMO) requires both high-quality proof generation and reliable verification, and single-model approaches fail due to reward hacking in the verifier.

Method: MaxProof trains a single M3 model merging three experts: a Proof Expert trained via long-horizon CISPO RL under a four-layer defense-in-depth generative verifier (bad-case filtering, solution normalization, multi-judge parallel scoring, pessimistic min aggregation); a Verifier Expert trained on error-finding tasks harvested from the Proof Expert's training trajectories; and a Fixer Expert trained via rejection-sampling fine-tuning on (problem, flawed proof, verification analysis) triples requiring 7/7 verifier scores. At test time, the MaxProof loop generates N=32 candidates, runs R=10 refinement rounds with M=4 parents per round (PATCH or REWRITE mode), uses conservative fitness (minimum over K_verify=4 verifier samples), and applies pairwise tournament selection with K_ranker=3 votes per match.

Key results:

• IMO 2025: 35/42 points (vs. 27/42 one-shot M3); human gold-medal threshold exceeded
• USAMO 2026: 36/42 points (vs. 26/42 one-shot M3)
• IMOProofBench standalone: 67.40 (vs. Gemini 3.1 75.71, GPT-5.5 90.85)
• 9 of 12 competition problems reach oracle-best 7/7 by round 4 of refinement
• Prior M2 cycle single-judge verifier had only 17% actually correct among "perfect score" rollouts; defense-in-depth verifier resolves this

Why it matters / caveats: The four-layer verifier design with documented reward-hacking failure modes from the previous cycle is a significant methodological contribution, and the competition scores are state-of-the-art for open systems. The three remaining unsolved problems expose a ceiling from base-model capability and verifier calibration limits; standalone benchmark scores lag behind frontier closed models.

Problem: Standard full attention in large language models becomes computationally prohibitive at ultra-long contexts (1M+ tokens), blocking practical deployment of long-context LLMs.

Method: MiniMax Sparse Attention (MSA) is a blockwise sparse attention mechanism with a two-branch design: an Index Branch using one index query head per GQA group and a shared index key head scores key-value blocks via max-pooling and selects the top-k blocks (block size B_k=128, k=16, attending to only 2,048 tokens per query regardless of context length); the Main Branch performs scaled dot-product attention restricted to those selected blocks. Training incorporates a KL divergence loss aligning the index distribution to main-branch attention patterns, gradient detachment preventing backprop through the backbone, two-stage indexer warmup, and forced local block selection for stability. Co-designed H800 GPU kernels use exp-free top-k selection, KV-outer iteration, and pre-scheduled tile chunking.

Key results:

• 28.4× reduction in per-token attention compute at 1M context
• 14.2× prefill wall-clock speedup on H800; 7.6× decoding wall-clock speedup
• 109B-parameter MoE model (6B active): MMLU 67.2, GSM8K 77.7, HumanEval 64.0
• RULER-32K: 77.5 (maintains long-context capability under 2,048-token attention budget)
• Native sparse pretraining matches full-attention baseline on language modeling loss after 3T tokens
• Compatible with conversion from existing GQA checkpoints with minimal performance degradation

Why it matters / caveats: The compute reduction (28.4×) and wall-clock speedups (14.2× prefill) at 1M context are substantial enough to change the economics of long-context inference. The fixed 2,048-token attention budget is a hard constraint that may miss critical tokens in tasks requiring precise global retrieval; the evaluation is primarily on the authors' own 109B MoE model.

Problem: Existing computer-use benchmarks do not require genuine coordination between GUI and CLI/code within single trajectories, allowing agents to succeed via single-channel shortcuts that fail in real hybrid workflows.

Method: WeaveBench provides 114 tasks across 8 domains on Ubuntu Linux in containerized VMs, enforcing three admission criteria: channel non-substitutability (P1: neither GUI nor CLI alone suffices), long-horizon execution (P2: multiple interleaved phases), and cross-application state (P3: tasks span independent applications). Evaluation uses a trajectory-aware judge with a five-stage layered pipeline — decomposing deliverables into atomic clauses, re-fetching evidence, aggregating per-deliverable scores, assigning 8 process/outcome dimensions, and scanning for 9 shortcut patterns (fake screenshots, hard-coded metrics, mock services, etc.) — run against a minimal GUI plugin (1 perception tool, 9 actuation primitives) added to existing CLI-agent runtimes.

Key results:

• Best PassRate: 41.2% (Claude Opus 4.7 + Claude Code runtime)
• Claude Opus 4.7 on OpenClaw runtime: 35.1%; GPT-5.5: 33.3%; GPT-5.4: 22.8%; Gemini 3.1 Pro: 1.8%
• Hybrid vs. single-channel gap: +31.6 pp (vs. GUI-only ≤1.8%, CLI-only ≤3.5%)
• Trajectory-aware judging removes 10.3–20.2 pp vs. outcome-only scoring (GPT-5.5 drops 53.5% → 33.3%)
• Dominant failure modes: reward hacking (35.2%), long-horizon discipline collapse (30.4%), reasoning/tool drift (21.0%)
• Median task: 76 tool calls, 16 GUI↔CLI channel switches

Why it matters / caveats: The +31.6 pp gap between hybrid and single-channel performance — versus only 3–4 pp on prior benchmarks — validates that WeaveBench genuinely requires hybrid coordination. At 114 tasks across 8 domains the benchmark is relatively small, and the best result of 41.2% means even frontier models solve fewer than half the tasks.

Problem: Existing vision-language-action (VLA) models trained on household manipulation data cannot execute precise, contact-rich scientific laboratory protocols with unfamiliar instruments and diverse robot embodiments.

Method: LabVLA uses a two-stage training pipeline on top of Qwen3-VL-4B-Instruct: FAST token pretraining first makes the VLM action-aware using grounded data (Robointer-VQA, AgiBot World Beta, OXE-AugE, DROID; 572K trajectories), then flow matching posttraining attaches an 18-layer Diffusion Transformer (DiT) action expert (1024 width, 8 heads) with "knowledge insulation" via stop-gradient blocking flow-matching gradients from the VLM prefix. Laboratory training data comes from RoboGenesis, a simulation data engine that reconstructs 2,947 assets via text-to-image generation and TRELLIS 2.0, composes atomic skills across 16 robot platforms with domain randomization, and exports demonstrations with 15 annotation streams.

Key results:

• LabUtopia average success rate (in-distribution): 71.1%; out-of-distribution: 70.0% (highest among all baselines)
• vs. next-best π₀: 63.3% ID / 63.2% OOD
• Fine-tuning X-VLA on LabEmbodied-Data: +15.0 pp ID (49.3% → 64.3%), +19.3 pp OOD (43.7% → 63.0%)
• Real robot (Franka): 74.0%–86.5% success across 4 composite tasks and workspace conditions
• Dataset: 10,000 laboratory scenes, 4 task families, 2,947 3D assets, 16 robot platforms

Why it matters / caveats: The knowledge insulation design (stop-gradient between VLM and DiT) is a practical solution for preventing the flow-matching objective from corrupting VLM representations. Real-robot results are limited to a single Franka platform and 4 tasks, and the simulation-to-real gap for precision lab operations (liquid handling, heating) remains a concern.

Problem: Unified multimodal models that handle image understanding, generation, editing, and video understanding require separate tokenizers per modality, creating architectural fragmentation and inconsistent representations.

Method: HYDRA-X introduces Hydra-XTok, a single Vision Transformer split into generation (Gen-ViT) and semantic (Sem-ViT) branches connected via a bottleneck, which unifies image and video tokenization. For video reconstruction, it uses frame-level causal tubelet attention (2-frame window) with hierarchical temporal patchify (two consecutive 2×2 stages) rather than full spatiotemporal attention. A lightweight decompressor enables joint distillation from both image (SigLIP-SO400M) and video (InternVideo-Next-L) teachers. For image editing, source and target images are routed through the shared Sem-ViT as a length-2 clip, enabling latent-level cross-image interaction. The full model uses a Qwen2.5-7B-Instruct LLM backbone and trains across five UMM pre-training stages totaling ~130 GPU-hours on 256–512 GPUs.

Key results:

• ImageNet reconstruction PSNR: 32.04 vs. AToken 29.72; SSIM: 0.898 vs. 0.848
• DAVIS video rFVD: 11.61 vs. AToken 29.19 (61% improvement); UCF rFVD: 3.11 vs. 7.77 (60% improvement)
• MME understanding: 2350.0 vs. HYDRA 2068.6; OCRBench: 84.5 vs. Show-o2 32.4
• GenEval generation: 0.88 (matching BAGEL-14B at 7B scale)
• Video-MME (no subtitles): 60.0 vs. Show-o2-7B 57.4; MVBench: 59.1 vs. 55.8
• ImgEdit-Bench overall: 4.34 vs. BAGEL 3.20; source-target interaction improves editing PSNR from 20.74 to 27.65 dB
• Frame-level causal attention latency: 0.25s vs. full attention 0.49s

Why it matters / caveats: The finding that minimal temporal attention (2-frame causal window) outperforms full spatiotemporal attention challenges conventional design assumptions and improves both latency and quality. The model covers five tasks at 7B scale, but training cost (~130 GPU-hours on 256–512 GPUs) is substantial, and standalone image understanding still trails specialized models on some benchmarks.

Problem: GRPO and similar RL-based LLM training methods generate insufficiently diverse solution rollouts because discrete token sampling limits exploration within the semantic neighborhood of each token.

Method: N-GRPO introduces Semantic Neighbor Mixing, which at rollout time selects an anchor token via temperature-scaled argmax, retrieves k−1 nearest semantic neighbors by cosine similarity in embedding space, computes softmax-normalized weights over neighbor logits, and substitutes a continuous weighted-sum embedding in place of the discrete anchor embedding at probability ρ=0.1. A stochastic gating mechanism activates mixing at that rate; the anchor token is used as the discrete realization for reward computation. The rollout buffer records candidate sets and weights for reproducibility. The method integrates with GRPO and GSPO without modifying the reward or policy ratio computation.

Key results:

• MATH-500 Pass@32 (1.5B DeepSeek-R1-Distill-Qwen): 79.17 vs. GRPO 77.41 (+1.76 pp)
• MATH-500 Pass@32 (7B): 84.20 vs. GRPO 82.21 (+1.99 pp)
• AIME25 Pass@32 (1.5B): 50.28 vs. GRPO 47.31 (+2.97 pp)
• GPQA-Diamond Pass@32 (7B): 92.80 vs. GRPO 91.42 (+1.38 pp)
• Rollout throughput overhead: 8.9–9.6%
• Cosine distance retrieval significantly outperforms L1/L2 metrics in ablations

Why it matters / caveats: Embedding-level mixing is a lightweight, architecture-agnostic augmentation that improves exploration without modifying the reward structure or requiring additional model components. Gains are consistent but modest (1–3 pp on Pass@32), and the method is evaluated only on math reasoning benchmarks; applicability to non-math domains is not demonstrated.

Problem: As LLM agents grow more capable, the bottleneck for autonomous scientific discovery shifts from prescribing agent workflows to designing the execution environment that constrains, budgets, and persists agent work.

Method: EurekAgent implements a prepare-propose-implement loop operating across four environment engineering dimensions: permissions engineering (Docker isolation, hidden evaluators exposed only via secure grading APIs, GPU access via helper APIs), artifact engineering (filesystem and Git-based memory storing proposals, solutions, evaluation feedback, and ranked submissions), budget engineering (dual-axis wall-clock and API-cost controls with resumability), and human-in-the-loop engineering (terminal UI for live inspection and web monitor for score evolution). The GLM-5.1 open-source model is used as the backbone. No specialized fine-tuning is applied; environment design drives performance.

Key results:

• Circle packing: 2.635999 (new SOTA vs. previous AI 2.635986)
• Erdős' Minimum Overlap Problem: 0.380870 (new SOTA vs. 0.380876)
• First Autocorrelation Inequality: 1.502861 (new SOTA vs. 1.502863)
• TriMul kernel: 2005.03 µs vs. previous best 2247.78 µs (~10.8% faster)
• MLE-Bench subset: 85.71% medal rate vs. prior best 71.43% (+14.28 pp)
• Circle packing solution found for under $11 in API costs; mathematics tasks averaged under $17

Why it matters / caveats: The results argue persuasively that environment architecture — not model capability alone — determines agent effectiveness, demonstrated by achieving new mathematical optima with an open-source model at minimal cost. The improvements on established mathematical conjectures are marginal (5th–6th decimal place), and broader generalizability across scientific domains beyond combinatorics, kernel engineering, and ML competition tasks is not established.

Problem: Text-to-image generation systems lack precise, structured feedback mechanisms that can localize defects, classify their type, explain their cause, and estimate their importance for targeted model improvement.

Method: Structured Defect Grounding (SDG) models each defect as a (location, type, reason, importance) tuple detected by a fine-tuned Qwen3-VL-4B-Instruct via two-stage training: SFT with coordinate jitter augmentation (±10 pixels in [0,1000] space) followed by GRPO with a composite reward (60% distance-IoU localization, 25% Qwen3-Embedding-0.6B cosine similarity for description, 15% clipped absolute-error importance accuracy). The SDG-30K dataset (30,096 images at 1024×1024 across 4 generators, 112 annotators, 1,085 person-hours) provides training/evaluation data. BoxFlow-GRPO converts SDG detections into spatially-varying per-location advantages for diffusion model alignment via LoRA (rank 64, 500 steps).

Key results:

• SDG (GRPO) artifact BoxF1@0.5: 0.263 vs. GPT-5.4: 0.035, Gemini 3 Pro: 0.200 (human reference: 0.278)
• SDG (GRPO) misalignment BoxF1@0.5: 0.387 vs. GPT-5.4: 0.292, Gemini 3 Pro: 0.307 (human: 0.409)
• Zero-shot generalization on RichHF-18K: misalignment F1 0.655 vs. ImageDoctor in-domain 0.250
• BoxFlow-GRPO on DrawBench: avg. improvement +2.4% vs. base; Flow-GRPO (no boxes) −6.7%
• SDG-30K: 28,945 train / 1,151 test (prompt-disjoint), inter-annotator BoxF1@0.5: 0.278 (artifacts), 0.409 (misalignment)

Why it matters / caveats: Instance-level structured defect representation surpasses proprietary frontier VLMs on localization, and the zero-shot RichHF-18K generalization (0.655 vs. 0.250 in-domain baseline) is compelling. BoxFlow-GRPO diffusion alignment gains are modest (+2.4% on DrawBench), and artifact localization remains well below human-level (0.263 vs. 0.278).

Problem: Speculative decoding wastes computation by routing all draft tokens through the full verifier even when many tokens could be validated by a much smaller derived submodel.

Method: VIA-SD is a hierarchical three-tier verification framework that derives a "slim-verifier" submodel from the full verifier via intra-model routing. Draft tokens are processed through confidence-based routing: high-confidence tokens are accepted directly, medium-confidence tokens go to the slim-verifier, and only uncertain tokens invoke full-model verification. The method is compatible with existing speculative decoding frameworks without modifying their training procedures.

Key results:

• Rejection rate reduction: 0.10–0.22 across four representative tasks
• Speedup over strong speculative decoding baselines: 10–20%
• End-to-end acceleration vs. non-speculative decoding: 2.5–3×
• Tested across multiple model families

Why it matters / caveats: Hierarchical verification reuses existing model components to reduce per-token verification cost without any retraining, making it a practical drop-in enhancement for deployed speculative decoding systems. The concrete slim-verifier derivation procedure and exact model families evaluated are not detailed in available paper content.

Problem: Training 3D human motion diffusion models requires expensive 3D ground-truth annotations, which are scarce compared to abundant monocular video data containing only 2D pose information.

Method: VideoMDM trains a 3D motion diffusion model using only 2D pose supervision extracted from monocular video. The key components are: (1) depth-aware weighting in the 2D reprojection loss (multiplying by depth d, truncated at d_min) that is proven theoretically equivalent in expectation to direct 3D supervision; (2) a 2D velocity loss for temporal consistency; (3) motion representation alignment using camera ray projection to create pseudo-targets for joint rotations, velocities, and foot contacts; and (4) a warm-up pretraining phase (400K batches on lifter outputs) followed by main training (200K batches) with LIS-style multi-step denoising for low-noise stability. A pretrained 2D-to-3D lifter provides noisy 3D pseudo-labels that are diffused as corrupted targets.

Key results:

• HumanML3D FID (2D supervision only): 0.876 vs. MDM (3D supervision) 0.544 and 2D baseline MDM/MVLift 1.671 (~2× improvement over 2D baseline)
• Fit3D MPJPE: 111.24 mm vs. WHAM 228.47 mm (2× reduction) and MVLift 283.06 mm
• Fit3D acceleration: 3.16 m/s² vs. WHAM 17.66 m/s² (5.5× smoother)
• NBA dataset human preference: 64% (VideoMDM/ElePose) vs. 36% (MAS); precision 0.94 vs. 0.50
• HumanML3D R-Precision: 0.721 (outperforms 3D-supervised MDM at 0.611)

Why it matters / caveats: The theoretical equivalence proof between depth-weighted 2D reprojection loss and 3D MSE loss is a clean result that justifies the approach formally, not just empirically. The method still relies on a pretrained 2D-to-3D lifter to supply noisy 3D pseudo-labels, so it is not fully self-supervised; evaluation covers limited diversity of motion types across three datasets.

Problem: Multimodal image fusion (e.g., infrared-visible) entangles global appearance factors (illumination, contrast, tone) with local spatial details in 2D feature grids, making it difficult to transfer appearance properties across modalities without corrupting structural content.

Method: The method replaces dense 2D feature grids with a compact 1D token interface using a frozen pretrained TiTok tokenizer (K=32 tokens, C=12 dimensions each). Selective Token Editing (STE) modifies only tokens at positions 12 and 18, channels {6,7,8} — locations identified via Gumbel-Softmax probing as appearance-sensitive. A Token-to-Map interface (π) lifts tokens hierarchically to a 64D, then 32×32, then 256×256 detail map. Private encoders decompose each modality into base B and detail D components fused separately; a residual reconstruction predicts ΔI added to a reference image. Two-stage training uses per-modality reconstruction SFT (40 epochs) followed by cross-modal fusion training (80 epochs) with SSIM+MSE and element-wise maximum losses.

Key results:

• M³FD object detection mAP50:95: 0.360 (highest among tested methods)
• FMB semantic segmentation mIoU: 0.692
• M³FD fusion metrics: Entropy 7.19, SD 47.35, SCD 1.85, SSIM 1.49
• Trainable parameters: 1.3M; FLOPs: 304.5G; latency: 124.3ms; memory: 2.78GB
• 32-token configuration outperforms 64 and 128 tokens (SSIM 1.42 vs. 1.38–1.36)
• TiTok outperforms DINOv3 and CLIP alternatives across all four fusion datasets

Why it matters / caveats: Using only 1.3M trainable parameters while achieving top detection mAP demonstrates that factorizing appearance into a compact token space is both parameter-efficient and task-effective. The discovery that editing just two specific token positions at three channels captures most appearance variation is interesting, though generality across other tokenizer configurations or datasets is unclear.

Problem: Hidden-state recurrence (latent chain-of-thought) in LLMs is incompatible with on-policy RL because latent positions lack a well-defined sampling distribution for computing policy ratios, and the mechanism of latent computation lacks interpretability.

Method: Switch introduces explicit boundary tokens <swi> and </swi> that delimit latent blocks, with each latent step's last-layer hidden state directly injected as the input embedding of the next <latent> token position. Three-phase training: (1) SFT tagging high-entropy CoT spans via entropy-based annotation from the SwiReasoning dataset; (2) curriculum training that gradually replaces text within boundaries with <latent> placeholder tokens using a parallel (all-spans-at-once) schedule; (3) Switch-GRPO RL where the policy ratio is defined exclusively at text positions and latent positions contribute gradients only through KV-cache, making importance ratios well-defined. The base model is Qwen3-8B.

Key results:

• MATH-500: 79.3% accuracy (Switch-GRPO endpoint), +25.7 pp above strongest Coconut-style baseline at equivalent scale
• GSM8K: 89.2% accuracy
• Switch invocation rate: reduced from 81% to 58% by RL while improving latent-conditional accuracy by +12.6 pp
• <swi> boundary tokens achieve rank ≤1.7 at boundaries vs. rank ~1,000 elsewhere (3–4 orders of magnitude contrast)
• Linear probe for switch prediction: 91.9% accuracy at final layer
• Causal intervention (zeroing latent state): diagnostic accuracy drops from 100% to 33.3%
• Compression variant: 69% MATH-500 at 1,276 avg visible tokens (33% output length reduction vs. main checkpoint at 1,721 tokens)
• Training data: annotated subset of OpenR1-Math, 220K samples

Why it matters / caveats: Providing well-defined policy gradients for latent reasoning via text-position-only policy ratios is a principled solution to a genuine technical barrier, and the mechanistic analysis (rank probe, causal interventions) adds interpretability beyond benchmark numbers. The model is evaluated only on math benchmarks, and the 79.3% MATH-500 score, while strong for this paradigm, remains below frontier performance.