AI papers — 2026-06-17

Problem: Determining the optimal loop count for Parallel Loop Transformers (PLT) is unclear, as standard sequential looping incurs latency and memory costs proportional to loop count, making loop selection critical for practical deployment.

Method: LoopCoder-v2 is a 7B-parameter code model with 14 layers trained on 18 trillion tokens (1:1 text-to-code ratio) using a Parallel Loop Transformer (PLT) architecture. PLT employs Gated Sliding-Window Attention (G-SWA), which shares the first-loop KV cache across all subsequent loops to maintain O(L·S·d) memory regardless of loop count, and Cross-Loop Parallelism (CLP), which right-shifts previous-loop hidden states by one token position before adding to embeddings, eliminating sequential inter-loop dependencies for near-single-pass latency. Models are trained and instruction-tuned at loop counts R ∈ {1, 2, 3, 4}.

Key results:

• R=2 vs. R=1: SWE-bench Verified improves from 43.0% to 64.4%; Multi-SWE improves from 14.0% to 31.0%
• R=3: SWE-bench Verified drops to 27.6%; R=4: drops further to 22.4%
• Loop 2 achieves peak effective rank (174.6), largest output-distribution shift (Δp=1.75), and highest attention divergence; Loops 3+ show declining effective rank (172.5, 158.2) and oscillatory updates (cos θ < 0)
• Thinking variants (explicit chain-of-thought) improve LiveCodeBench by 26.9 points over instruction-tuned baselines

Why it matters / caveats: The gain–cost trade-off explains why looping exactly twice is optimal: Loop 2 provides primary productive refinement while the fixed CLP-induced positional mismatch dominates over diminishing representational gains beyond that. Analysis is limited to a single 7B architecture, and adaptive or dynamic loop allocation remains unexplored.

Problem: Small language models trained via knowledge distillation suffer mode-seeking and poor generalization, while reinforcement learning fails on hard questions where all student rollouts are incorrect and yield zero advantage signals.

Method: ZPPO (Zone of Proximal Policy Optimization) is built on GRPO with DAPO components (clip-higher, token-level loss) and keeps teacher knowledge in prompts rather than gradients via three mechanisms: Binary Candidate-included Questions (BCQ) pair one correct teacher response with one incorrect student response, prompting the student to reason between them on-policy; Negative Candidate-included Questions (NCQ) aggregate all student wrong rollouts from a batch, label them as incorrect, and prompt the student to find the right answer; and a Prompt Replay Buffer stores only questions (not responses) where student mean accuracy is below 50%, graduating them upon reaching 50% accuracy, ensuring practice within the learner's zone of proximal development. Training uses I=4 iterations per step with batch-level advantage normalization.

Key results:

• Tested on Qwen3.5 at 0.8B–9B with a 27B teacher across 31 benchmarks (16 VLM, 10 LLM, 5 Video)
• VLM training-aligned benchmarks: ZPPO gains +9.3 pp at 0.8B, +5.2 pp at 2B, +4.0 pp at 4B, +2.8 pp at 9B vs. GRPO†
• LLM + Video generalization benchmarks: distillation degrades by −2.5 pp to −0.3 pp; ZPPO improves by +6.8 pp to +2.7 pp
• 28% of zero-accuracy questions graduate with ZPPO vs. 4% with GRPO†

Why it matters / caveats: ZPPO successfully transfers teacher knowledge without imitating logits while improving generalization where traditional distillation fails. BCQ requires teacher success on hard questions; when both models fail, only NCQ applies, creating a structural performance ceiling bounded by teacher capability.

Problem: Vision-Language-Action (VLA) models require large-scale diverse embodied data but combining human egocentric videos with multi-robot data introduces representation heterogeneity (different coordinate frames and kinematic structures) and supervision-quality mismatch (noisy pseudo-actions from human videos vs. precise robot sensor logs).

Method: ACE-Ego-0 unifies 6,013.7+ hours of heterogeneous data (4,534.8+ robot hours, 1,478.9 human video hours) through three alignment mechanisms: camera-space action standardization to head-camera coordinate frames via calibrated extrinsics; morphology conditioning using URDF-graph message-passing neural networks for robot platforms and learned surrogate embeddings for human sources, both projected into shared token space; and time-aligned action chunking defined by physical duration rather than frame count. A Reliability-Aware Training Objective applies standard flow-matching on robot data and weighted Huber regression on human pseudo-actions (Wt,j = ρj · wdata · wstep), concentrating supervision on reliable position channels while downweighting noisy rotations and gripper signals.

Key results:

• RoboCasa GR1 TableTop (24 tasks): 72.8% average success, outperforming DIAL (70.2%), JoyAI-RA (63.2%), ABot-M0 (58.3%)
• RoboTwin 2.0 (50 tasks): Easy/Clean 91.12%, Hard/Randomized 90.62%, surpassing JoyAI-RA by +0.64% and +1.34%
• Real-world ARX bimanual (6 tasks): 78.3% average success, outperforming fine-tuned π0.5 (71.7%) by 6.6% and GR00T-N1.7 (35.6%) by 42.7%
• Data source ablation: robot-only 68.3% vs. robot + human 72.8% (+4.5%)
• Limited-data fine-tuning on Sweep Cubes: robot-only 10% vs. robot + 419 human video episodes 40% (4× improvement)

Why it matters / caveats: Reliability-aware weighting enables noisy pseudo-actions to complement rather than corrupt high-fidelity robot supervision, with human video providing substantial behavioral coverage. Evaluation is limited to tabletop manipulation; mobile manipulation, dexterous hands, and force/torque sensing are excluded.

Problem: Existing benchmarks for AI game generation evaluate only partial aspects—code correctness or visual plausibility—without jointly assessing whether agents can deliver functioning, playable interactive systems in a real game engine.

Method: GameCraft-Bench is a benchmark of 140 tasks across 15 game families (platformers, strategy, tycoon, visual novels, etc.) in Godot 4, evaluated through a five-stage pipeline: task packaging (specification + hidden rubric), agent generation (project creation + demonstration traces), a build gate (launchability verification), replay (recorded interaction at 1280×720 resolution), and multimodal judge scoring across four categories—Core Mechanics (15% weight), Content Depth (35%), Functional Visuals (15%), and Art and Presentation (35%).

Key results:

• Strongest agent (Claude Opus-4.7): 41.46% overall
• GPT-5.5 high: 39.49%; Kimi-K2.6: 30.65%
• Category breakdown for Opus-4.7: Core Mechanics 55.34%, Content Depth 39.48%, Functional Visuals 42.78%, Art and Presentation 36.86%
• Multimodal judge shows stable scores across repeated evaluations (std dev ~0.004), calibrated within ±3.32% of human annotators

Why it matters / caveats: Current frontier agents remain far from reliable end-to-end game generation, capable of runnable prototypes but struggling with content depth and polished presentation. Evaluation is restricted to 2D Godot games, ignores audio, and measures specification adherence rather than subjective fun or commercial viability.

Problem: Existing AI-assisted learning systems focus on what to teach rather than how to teach, lacking integration of adaptive content generation with embodied instructional delivery tailored to individual learners.

Method: LectūraAgents is a hierarchical multi-agent system using a "Swarm-of-Ranks" group chat where a ProfessorAgent (Rank 1) coordinates a LecturePlanner (Rank 2) and executor agents—ResearchAgent, SlideAgent, ScriptAgent, SpeechAgent, TasaAgent (Rank 3)—across two sessions: Lecture Preparation (iterative planning, research, slide/script/speech generation conditioned on learner profiles) and Lecture Delivery (pedagogically motivated teaching actions including rough notations and handwriting synchronized with speech). The Teaching Action-Speech Alignment (TASA) algorithm uses LLM-based salience heuristics and temporal semantic segmentation to generate coherent action sequences labeled as Pedagogical, Personalized, Salient, Adaptive, or Assessment.

Key results:

• Best model (Gemini 3 Pro): 80.4% overall (Lecture Content Quality 80.2%, Personalization Quality 83.3%, Assessment Quality 81.6%, Teaching Action Quality 76.5%)
• Performance range across 7 frontier models (280 lectures): 64.1%–80.4%
• Comparative: LectūraAgents 71.6% overall vs. Learn Your Way 60.5% vs. GenMentor 54.0% (20-lecture evaluation)
• Student study (45 students): 95% felt "adequately prepared" vs. 80% (Learn Your Way) and 72% (Adobe Reader)

Why it matters / caveats: LectūraAgents provides the first multi-agent framework with end-to-end personalization combining embodied teaching actions with adaptive content. The TASA module relies on offline heuristics with limited supported teaching actions, and the system inherits common LLM failure modes including factual errors.

Problem: Clinical risk prediction systems either produce opaque predictions without explanation or generate natural language reasoning that polarizes risk estimates toward near-certainty (>99.98%), eliminating cross-patient comparability.

Method: TRIAGE applies a two-stage pipeline to Qwen3-4B-Base. Stage 1 (Dialectical Reasoning Supervision) uses GPT-5.1 to generate separate outcome-conditioned rationales for each candidate outcome, then concatenates them with ground-truth labels to create balanced training traces that avoid confirmation bias. Stage 2 (Self-Refinement via GRPO) uses Group Relative Policy Optimization with a batch-level reward function that compares each sample's log-odds against opposite-class batch averages; cross-entropy loss supervises only the final decision token while preceding reasoning uses GRPO. Risk scores are extracted from implicit token probabilities at a fixed "## Final Decision" header to preserve continuous probability distributions.

Key results:

• AUROC on P12: 87.3%, P19: 89.3%, MIMIC-III: 86.7%
• Average rank 1.58 across all metrics; 3.3% relative AUPRC improvement over strongest baseline (GRU-D)
• Mean Expected Calibration Error reduced by 80% post-RL; mean Brier score reduced by 49%
• LLM-as-judge explainability score: 7.744 vs. 6.474 for baseline (STraTS + Integrated Gradients + GPT interpretation), +0.902 on interpretive summary
• Outperforms zero-shot gpt-oss-120b on calibration (ECE: 0.03 vs. 0.23 on MIMIC-III)

Why it matters / caveats: TRIAGE decouples reasoning quality from risk polarization through dialectical deliberation on a small 4B LLM without sacrificing predictive performance. Evaluation is restricted to binary classification and relies on LLM judges rather than clinical expert assessment; the system is a research prototype, not a clinical deployment tool.

Problem: Existing on-policy self-distillation (OPSD) methods are designed for autoregressive LLMs and are incompatible with diffusion LLMs' (dLLMs) arbitrary-order, iterative denoising generation pattern.

Method: d-OPSD adapts on-policy self-distillation to diffusion LLMs through two key modifications. Rather than appending privileged information as prefixes (as in autoregressive OPSD), teacher conditioning uses self-generated answers as suffix conditioning: at denoising step t, the teacher receives partially revealed tokens from the final generated response while the student sees only the current noisy sequence. Instead of token-level KL divergence, the method computes step-level divergence over denoising steps, supervising only the top-k most confident tokens among currently masked positions. The training objective minimizes step-level KL divergence across student-sampled trajectories using reverse KL with pointwise clipping for stability.

Key results:

• GSM8K: 81.0% (d-OPSD) vs. 79.8% (diffu-GRPO baseline)
• MATH500: 37.2% (matched with baseline)
• Countdown: 37.9% vs. 33.2%
• Sudoku: 23.9% vs. 18.4%
• Sample efficiency: d-OPSD requires ~10% of optimization steps vs. RLVR (converging in ~425 steps on GSM8K vs. 7,700 for diffu-GRPO)

Why it matters / caveats: d-OPSD enables single-model self-improvement for diffusion LLMs with dramatically improved sample efficiency over RLVR baselines. The method shares RLVR's vulnerability to policy collapse after reaching peak performance, and a robust mitigation strategy remains unresolved.

Problem: Memory-augmented agents can store trajectories and retrieve reflections but lack the holistic competence to select useful experience, act on it, write reusable knowledge, and maintain a growing repository—meaning they retain experience without genuinely learning how to evolve through it.

Method: OPD-Evolver is a slow-fast co-evolution framework using on-policy self-distillation. The fast loop interacts with a four-level memory hierarchy for reading, using, writing, and maintaining experience during test-time evolution. The slow loop applies outcome-calibrated memory attribution and privileged hindsight distillation to internalize the four core memory competencies into a deployable policy, transferring capabilities from larger to smaller models.

Key results:

• Surpasses ReasoningBank memory systems by up to 11.5%
• Outperforms training-based method Skill0 by ~5.8%
• OPD-Evolver-9B challenges Qwen3.5-397B-A17B and Step-3.5-Flash across multi-domain benchmarks

Why it matters / caveats: The approach demonstrates that smaller agents can internalize high-value experience and memory management, moving beyond memory-augmented systems toward genuinely self-evolving agents. Specific benchmark names and absolute accuracy numbers are not stated in the available text; performance claims are relative comparisons only.

Problem: Pixel-space diffusion models process full-bandwidth noisy images at every timestep, but at each timestep a per-band data-to-noise ratio defines a moving boundary between signal-bearing low frequencies and noise-dominated high frequencies, forcing networks to discover this boundary implicitly and wasting capacity on frequency-time regions where predictions collapse to deterministic baselines.

Method: Spectral Forcing (SF) is a parameter-free, time-conditional 2D-DCT low-pass filter applied before patch embedding, defined as IDCT(DCT(z) ⊙ M(t)) where the cutoff radius c(t) expands monotonically with diffusion time. The cutoff schedule is derived from the per-band data-to-noise contour k*(t) = (1-t)^(-2/α), with an empirical linear schedule f(t) = t performing best on ImageNet-256. SF adds ~0.5% compute overhead, requires no parameter changes, and composes with any rectified-flow recipe without modifying loss, sampler, or classifier-free guidance.

Key results:

• JiT-700M/32 (64 tokens) at 60 epochs: FID reduced from 24.19 to 20.68 (−14.5%); Inception Score improved from 83.28 to 93.96 (+13%)
• JiT-130M/32: +3.9% FID gain at 60-epoch budget
• Reaches prior published FID reference (~15.24) in substantially fewer epochs
• Transfer to SenseNova-U1 text-to-image model: wins 99 of 113 DPG-Bench subcategories

Why it matters / caveats: Making the implicit bandwidth boundary explicit frees model capacity in coarse-tokenization regimes without modifying any learned parameters. The method is regime-dependent—it is neutral to slightly harmful at fine tokenization (256+ tokens) where high-frequency content carries genuine signal rather than noise.

Problem: Modern hybrid language models combine full attention with efficient attention modules (sliding-window attention, recurrent mixers) to handle long contexts, but the mechanistic roles these modules play in enabling long-context capability remain poorly understood.

Method: The paper conducts systematic scaling-law analysis across model sizes (15M–477M parameters) and training budgets (100N–1000N tokens), comparing seven architectures: full attention, three sliding-window attention (SWA) variants (128, 512, 2048 token windows), and three recurrent mixers (Lightning Attention, Mamba-2, Gated DeltaNet). Evaluation uses validation loss for short-context performance and log(LongPPL) for long-context capability, fitting power-law curves L(N,D) = aN^(-α) + bD^(-β). Receptive-field constraint experiments and layer-wise probing isolate the contribution of full vs. efficient attention layers.

Key results:

• All hybrids overlap nearly perfectly on validation loss curves, but log(LongPPL) shows substantial gaps at early training that shrink with sufficient data
• Long-range information concentrates in full-attention layers; restricting efficient-attention receptive fields causes minimal LongPPL change
• Large-window SWA (2048 tokens) delays formation of retrieval heads in full attention (large-window laziness phenomenon)
• Applying NoPE (no positional encoding) exclusively to full-attention layers in small-window SWA hybrids: +16–32 point improvement on RULER NIAH benchmarks at 16K context while maintaining short-context performance

Why it matters / caveats: Efficient attention functions as an "optimization prior" shaping how rapidly full attention develops long-range retrieval, reframing hybrid design around full-attention activation rather than increasing efficient-attention intrinsic strength. Experiments use sub-billion-parameter models on at most ~100B tokens, and findings may not transfer directly to frontier-scale systems.

Problem: Large language models that perform well on English logical reasoning benchmarks may fail when processing equivalent logical structures expressed in diverse Chinese surface realizations such as ellipsis, rhetorical devices, idiomatic language, and pragmatic indirection.

Method: ChLogic is an English-Chinese aligned benchmark built from formal logical templates with three components: a General aligned set (60 propositions across 9 logical families, 3,000 items), a Difficult aligned set (40 mathematical logic problems, 2,000 items), and a Chinese-only set (15 language-specific phenomenon types, 1,500 items). Each template generates one English reference plus five Chinese surface variants—standard, natural written, colloquial, rhetorical-question, and perturbed forms. DeepSeek-V3 generates surface realizations; Yi-1.5-34B-Chat verifies semantic alignment; human reviewers make final acceptance decisions. Back-translation diagnostics isolate whether failures stem from formal reasoning deficits or Chinese surface normalization difficulties.

Key results:

• GLM-5.1: 98.30% on English General questions, drops to 78.89% on rhetorical Chinese variants
• Qwen3-32B: 96.05% on difficult English, drops to 69.35% on difficult rhetorical Chinese
• Qwen3 series scaling: average Chinese accuracy rises from 79.61% (8B) to 95.06% (32B) with persistent gaps remaining
• Ministral-8B: propositional equivalence drops from 94.67% (English) to 8.00% (rhetorical Chinese)
• Back-translation: Qwen3-32B improves from 95.73% to 99.30% on General set, but mixed/negative effects on harder problems

Why it matters / caveats: High English logical accuracy does not guarantee robust multilingual reasoning; failures concentrate in surface-to-logical-form normalization rather than abstract inference. The benchmark covers template-driven items rather than discourse-level reasoning and evaluates only contemporary model families.

Problem: Reinforcement learning with verifiable rewards (RLVR) improves language-model reasoning but GRPO-style optimization remains prone to training collapse, and the mechanism underlying this instability is not well understood.

Method: The paper analyzes instability through token-level gradient dynamics, deriving a taxonomy that predicts how updates affect next-token probabilities and entropy based on advantage sign and token distribution under the current policy. Motivated by this analysis, the authors propose Winner Advantage Policy Optimization (WAPO), a simple online clipped policy-gradient objective that updates only on positive-advantage completions, filtering out negative-advantage signal that contributes to collapse.

Key results:

• WAPO evaluated across mathematical reasoning and multi-hop QA benchmarks on multiple model families
• WAPO improves training stability and matches or outperforms GRPO-style baselines across all evaluated settings
• Specific accuracy percentages and benchmark names are not stated in the available text

Why it matters / caveats: The gradient-taxonomy framework provides actionable theoretical grounding for why selective positive-advantage updates enhance RLVR stability. Detailed quantitative results are not disclosed in the abstract, limiting direct comparison with prior work.

Problem: Existing unified multimodal systems rely on two separate visual tokenizers for understanding and generation, fragmenting the representation space and preventing models from directly interpreting their own generated visual tokens without additional re-encoding.

Method: UniAR is a unified autoregressive framework where a single discrete visual tokenizer bridges understanding and generation. The tokenizer adapts a pretrained vision encoder with multi-level feature fusion and lookup-free bitwise quantization, preserving both high-level semantics and low-level details while scaling the effective visual vocabulary at minimal cost. The unified autoregressive model employs parallel-bitwise-prediction to jointly predict spatially grouped, multi-level visual codes, substantially reducing visual sequence length and accelerating generation. A diffusion-based visual decoder then decodes discrete visual tokens to high-fidelity images. The training pipeline proceeds through large-scale pre-training, supervised fine-tuning, and reinforcement learning. Accepted at ICML 2026.

Key results:

• Achieves state-of-the-art performance on image generation and image editing benchmarks
• Remains competitive on multimodal understanding benchmarks
• Specific numerical scores (FID, accuracy %) are not stated in the available text

Why it matters / caveats: A single shared tokenizer is demonstrated to be sufficient for both visual understanding and generation within one autoregressive model, simplifying architecture and enabling direct self-interpretation of generated tokens. Concrete benchmark numbers are not available in the abstract; the paper requires access to the full text for quantitative evaluation.

Problem: Text-only image editing instructions lack spatial precision, while visual-only prompts (e.g., drag-based methods) lack semantic grounding, making it difficult for users to express both where and what to edit without ambiguity.

Method: TV-Edit is a plug-and-play architecture operating on pretrained editing backbones (Qwen-Image-Edit, FLUX.1 Kontext) with two branches: a main branch (VAE encoder, text encoder, Multi-Modal Diffusion Transformer) and a control branch (sparse point encoder plus Content-Aware Spatial Controller). The controller performs early fusion of spatial cues, image content, and textual conditions through convolutional layers that concatenate with image/noise latents; a time-modulated inject layer uses timestep-dependent scaling coefficients to regulate injection intensity. Training uses Z₀-prediction loss with Beta distribution timestep sampling (Beta(5,2) annealed from Beta(20,2)). TV-Edit-23K is a paired dataset of 23,000 samples from video frames with point trajectories, text prompts, source and target images, constructed via optical flow tracking, MLLM-guided semantic annotation, and quality filtering.

Key results:

• Dense matching distance (MDd): 0.0462 (TV-Edit-Qwen) vs. 0.0648 (GoodDrag), a 28.7% reduction
• Prompt following (PF): 0.93 vs. 0.86 (Qwen baseline), +7%
• Global DINO score: 0.9134 vs. 0.9096 (NanoBanana), +0.4%
• DragBench: mean displacement 17.31 vs. 23.77 (best drag baseline); image fidelity 0.86

Why it matters / caveats: Co-instructing with text and visual point trajectories reduces editing ambiguity and achieves strong spatial control while maintaining semantic coherence. Inference speed precludes real-time interactive editing, and the method handles limited 3D manipulations involving out-of-plane rotations.

Problem: Interactive world models focus on navigation while neglecting object-level interaction, suffering from a data bottleneck (lack of human-object interaction datasets with dense labels) and a memory bottleneck (recency-biased history compression discards causally important contact and manipulation frames, causing action-forgetting).

Method: ActWorld addresses these bottlenecks through per-chunk Chain-of-Thought annotation (33-frame video segments described by comparing consecutive keyframe pairs and assigning interaction phases), keyboard/mouse control conditioning via dual-branch Plücker-ray FiLM modulation plus symbolic text-camera embeddings for 81 discrete command combinations, and Hierarchical Action-Aware Memory comprising Event-Aware Frame Re-assignment (EAFR, importance-scored frame selection prioritizing interaction-critical frames), Action-Conditioned History Amplification (ACHA, an MLP sharpening attention on causally relevant history frames per action class), and a persistent action-aware memory bank (Ktot=16 tokens) carrying event and DINOv3-anchored object tokens that survive beyond the latent buffer's eviction horizon. A three-stage training procedure uses multi-resolution flow-matching and DMD-style distillation reducing a 50-step teacher to a 3-step student.

Key results:

• I-Bench (300 prompts, navigation + object interaction): VBench Subject Consistency 0.871 vs. 0.856 (HY-World); VLM-Action-Judge Success Rate 57.8% vs. 20.12% (Yume 1.5)
• Ablation: Event Memory alone improved subject consistency from 0.844 to 0.871 and success rate from 54.0% to 57.8%
• User study: ActWorld ranked first in action following (4.05 vs. 2.68 next best), key/mouse following (3.69 vs. 2.35), overall quality (3.92 vs. 3.15)
• DINOv3 overhead ~15%; 3-step distilled inference enables real-time generation

Why it matters / caveats: Unifying navigation and fine-grained object interaction within a single real-time world model is demonstrated by treating memory design as the primary bottleneck. Action vocabulary is restricted to 40 predefined classes, and the 57.8% success rate on semantic instruction-following indicates substantial room for improvement.

Problem: World models face a fundamental tension between faithful long-horizon simulation requiring deep computation and the exponential error accumulation of multi-step rollouts in conventional architectures.

Method: LoopWM introduces looped transformer architectures for world modeling using a parameter-shared transformer block applied iteratively (T iterations) with the update rule h^(t+1) = Ā·h^(t) + B̄·e + R̄(h^(t), e). Spectral stability is enforced by constraining the state-retention matrix Ā to eigenvalues in (0,1) via Ā = exp(Δ·A) with A := diag(-exp(a)), bounding latent evolution across arbitrary rollout lengths. Adaptive Early Exit uses lightweight exit gates halting inner-loop iterations upon convergence, allocating more compute to complex transitions. Deferred Decoding eliminates intermediate observation decoding during multi-step rollouts, performing K action-conditioned latent transitions before a single terminal decode.

Key results:

• ScienceWorld: LoopWM 68.4% exact match (EM) vs. claude-opus-4-6-max 47.2% EM (+21.2 pp), using ~1B parameters (100× smaller)
• ScienceWorld Lifespan task: 0% to 100% with LoopWM
• AlfWorld: 51.6% EM, 80.4% F1, 71.6% BLEU (second on EM/F1, first on BLEU)
• Adaptive exit reduces inference FLOPs by ~25× for simple transitions
• Deferred decoding: 54.5%–113.8% relative EM gains on 5-step rollouts vs. Gemini baseline

Why it matters / caveats: Iterative latent depth is established as a new compute-scaling axis for world simulation orthogonal to model size and training data, with provably stable rollouts via spectral constraints. The paper is intentionally selective in disclosure scope, validation is concentrated in specific task domains, and scaling laws across compute ranges are incompletely characterized.

Problem: Computer-use agents require large-scale diverse trajectory training data, but the largest existing human-annotated resource (AgentNet, 22.5K trajectories) causes negative transfer when used for SFT, reducing UI-TARS 7B's OSWorld performance from 26.3% to 8–10%.

Method: ProCUA-SFT is a synthetic trajectory generation pipeline built on four principles: Grounded Task Synthesis, where a single VLM (Kimi-K2.5) generates goals alongside binary preconditions verified by an independent judge policy against desktop screenshots with iterative retry on failure; Real-World Content Seeding, initializing desktops with authentic resources (912 spreadsheets from SpreadsheetBench, ~10,000 presentations from Zenodo10K, OSWorld configurations); Single-Model Rollouts, where Kimi-K2.5 serves as goal generator, verifier, and executor using different prompts, self-signaling completion via a reserved terminate() function; and Step-Prefix Expansion, where each T-step trajectory generates T training samples with growing screenshot-action history context matching inference layout. The infrastructure uses decoupled inference and environment with pluggable VM backends (local KVM via Singularity, serverless NVCF).

Key results:

• OSWorld success rate: 45.0% (ProCUA-SFT on UI-TARS 7B) vs. 26.3% (base UI-TARS 7B, +18.7 pp) vs. 8–10% (AgentNet-trained)
• Dataset: 3.1M step-level samples from 93K synthetic trajectories across 2,484 application combinations
• ProCUA trajectories: mean length 29.7 steps vs. AgentNet's 18.6 steps (1.8×)
• Diversity ablation: round-robin app-combination sampling achieves 30.9% vs. non-diversified 27.3%
• Action distribution: 41% keyboard-based (vs. AgentNet's 63% clicks), reducing brittle menu-navigation failures

Why it matters / caveats: Grounded synthetic trajectories with feasibility-verified tasks substantially outperform human-annotated data for computer-use agent training, demonstrating that data quality and feasibility guarantees matter more than human origin. All roles depend on Kimi-K2.5's capabilities, evaluation is OSWorld-centric, and complex multi-application workflows remain underrepresented (0.9% involve 4+ applications).

Paper Links:

• [2606.18023] https://arxiv.org/abs/2606.18023
• [2606.18216] https://arxiv.org/abs/2606.18216
• [2606.17200] https://arxiv.org/abs/2606.17200
• [2606.17861] https://arxiv.org/abs/2606.17861
• [2606.16428] https://arxiv.org/abs/2606.16428
• [2606.09030] https://arxiv.org/abs/2606.09030
• [2606.18195] https://arxiv.org/abs/2606.18195
• [2606.17628] https://arxiv.org/abs/2606.17628
• [2606.15236] https://arxiv.org/abs/2606.15236
• [2606.15378] https://arxiv.org/abs/2606.15378
• [2606.17905] https://arxiv.org/abs/2606.17905
• [2606.16154] https://arxiv.org/abs/2606.16154
• [2606.18249] https://arxiv.org/abs/2606.18249
• [2606.16767] https://arxiv.org/abs/2606.16767
• [2606.17730] https://arxiv.org/abs/2606.17730
• [2606.18208] https://arxiv.org/abs/2606.18208
• [2606.17321] https://arxiv.org/abs/2606.17321