Friday, October 30, 2026

Web agents need diverse, executable, and resettable website environments, but existing routes trade off automation (human-assisted mirrors), source fidelity (generative construction), and interaction coverage (capture-and-replay). WebAnything automatically reconstructs real URLs into local, resettable environments grounded in browser-observable evidence, with an independent verifier that routes each discrepancy to either local repair or targeted recapture of the source.

• An Explorer-Builder records page structures, resources, and before/after interaction outcomes for each visited state and implements them as local routes and state transitions, after which an Independent Verifier re-renders and re-operates the frozen candidate in a separate context, returning evidence-supported failures for local repair and turning unresolved evidence gaps into bounded recapture requests.
• The pipeline produced 244 website artifacts across 14 domains and 53 categories supporting 12,984 executed and accepted tasks, with no recorded per-site manual implementation or repair after initial setup and authorization on an 85-site historical subset and a recorded model-call cost of $220.62 across 84 sites (median $1.60 per site), stated as a lower bound.
• Against the human-assisted WebHarbor on five matched sites and 14 historical states, WebAnything scores higher on Pixel Similarity (0.924 vs 0.808) and SSIM (0.790 vs 0.626) and takes 12 wins, two ties, and no losses in a blind fidelity audit, but WebHarbor retains broader tested interaction coverage at 80% vs 75% capability coverage and 76.7% vs 63.3% all-test pass.
• Fine-tuning Qwen3.5-9B on 5,856 interaction trajectories lifts REAL success from 24.2 to 34.1 (+9.9pp), Odysseys-Easy Rubric and Task averages by +6.25pp and +6.32pp, and WebArena-Verified Overall from 16.41 to 29.74 (+13.33pp), driven by Navigate (+18.49pp) and Mutate (+31.50pp) while Retrieve falls 8.75pp under strict scoring.
• Training utility is validated only at the 9B scale, mirrors freeze a site at capture time with no study of staleness or maintenance as live sources change, the fidelity and interaction comparisons rest on five-site subsets, and copyright and terms-of-service constraints limit redistribution of the reconstructed mirrors.

arXiv:2610.37786 PDF

Standard multi-token prediction (MTP) heads are trained with cross-entropy against the data, but the gradient of that loss decomposes exactly into a KL self-distillation term that pulls each head toward the model's own next-token distribution and a noisy residual equal to the next-token cross-entropy gradient. The paper shows the residual has low signal-to-noise ratio and interferes with learning, then removes it with a one-line change that recovers the KL gradient without ever materializing vocabulary-sized logits.

• The KL gradient-recovery trick adds a zero-valued straight-through term to the NTP hidden states so that the NTP loss backpropagates the negative residual into each MTP head, which combined with the ordinary MTP cross-entropy gradient leaves exactly the KL gradient reaching the MTP heads and shared trunk while the shared output head still sees cross-entropy.
• The residual carries roughly 80% of the MTP gradient reaching the trunk yet has about 4x lower SNR than the KL component, with the gap widening from 4.1x at 167M to 4.6x at 1.2B, and a step along it consistently increases NTP loss.
• On LLaMA-style models from 167M to 1.2B trained on 31.5B FineWeb-Edu tokens with DeepSeek-style MTP, Soft-MTP beats standard MTP on NTP perplexity, speculative-decoding acceptance length (+1.8 to 2.2%) and speedup at every scale, raises average zero-shot accuracy by about 0.5 points at 689M and 1.2B where it also surpasses the NTP baseline, and after NTP-only fine-tuning at 1.2B leads MTP in seven of eight math and code cells, including +3.1 points on HumanEval pass@64 and +1.8 on MATH-500.
• It matches or exceeds explicit KL-MTP on most metrics at identical FLOPs, MFU and fused-cross-entropy peak memory to standard MTP, whereas KL-MTP lowers throughput by 5.6% and adds 21 GB of peak memory; keeping cross-entropy at the output head also preserves the readout's effective rank (437 vs. 131 for standard MTP and 496 for NTP).
• Limitations are that the KL gradient is biased toward the NTP distribution rather than the data, the largest model is only 1.2B so frontier-scale gains are extrapolated, sequence-sharded parallelism needs a small extra backward-pass exchange, and the trick delivers the exact KL gradient only to the heads and trunk, not the shared output head.

arXiv:2610.38331 PDF

Existing finance benchmarks filter retrieval by publication date only, so post-cutoff information still leaks into retrieval, computation, and scoring, and they test isolated skills rather than the full analyst chain from evidence to forecast to decision. ILE is a buy-side-expert-authored equity research benchmark built on a versioned corpus where every document version carries a validated availability timestamp enforced across the entire toolchain, with rubrics tied to sources, calculations, and assumptions through financial dependency graphs, and it scores research quality, forecast accuracy, and simulated investment returns jointly.

• The corpus holds 3.43 million document versions across nine source categories (1.52 TB), the cutoff applies to search, read, download, and table tools with retrieved results frozen against later revisions, and 86 expert-authored tasks span seven deliverable categories, graded by deterministic checks for numeric criteria and GPT-6 Astra for semantic ones.
• GPT-6 Astra leads overall at 68.1 and Claude Opus 5.5 follows at 66.9, but Astra's 78.9 on Research & Evidence and 77.6 on Evaluate & Decide fall to 52.5 on Forecast & Model, with Opus essentially tied at 52.6, so forecast construction is the shared weak point.
• Predictive criteria reorder the field: DeepSeek V4.1 Flash leads forecast-outcome scoring at 40.3 versus 37.7 for Astra and 37.6 for Opus, and in the March to August 2026 trading simulation Hunyuan 4 Preview returns 8.50%, Flash 7.98%, Astra 7.18%, and Kimi K3 7.15%, ahead of Opus at 6.39% and an always-long baseline of 1.18%, while Flash costs $0.36 per task against $7.96 for Opus.
• Astra scores 68.3 on public-web search versus 68.1 on the versioned database, yet 57% of its public-web trajectories contain post-cutoff sources such as realized Microsoft revenue and Tesla deliveries despite dated queries, so comparable scores do not certify that the information boundary held.
• The top-scoring model also serves as the semantic grader, which the authors acknowledge invites self-enhancement bias, pretraining memorization of historical outcomes remains uncontrolled since only retrieval leakage is blocked, the trading sample covers just 20 assignments and 11 companies, and strict submission mode drops some systems sharply, with Step 5 Preview falling from 55.1 to 34.5.

arXiv:2610.38342 PDF

Turning a fixed narration script into a long-form video requires interdependent decisions across planning, retrieval, speech synthesis, and timeline assembly, and prior agentic editors drive multi-stage workflows with general-purpose models without showing whether a compact, task-trained policy can match that quality at a fraction of the token cost. WeVisCut answers this with a single 9B policy that serves as orchestrator, generator, and critic inside a stateful tool harness, trained on executable trajectories reconstructed from finished videos rather than on final-video preferences alone.

• Because finished videos expose only narration and shot structure, a four-stage pipeline recovers script-to-shot intent through editing inversion, re-executes it with real production tools under verified replay, applies bounded observe-reflect-repair, and materializes the recorded calls, returns, and failures into 134,771 role-specific examples from 3,750 Bilibili projects.
• Training proceeds in three phases on a Qwen3.5-9B backbone with LoRA: supervised fine-tuning plus runtime-verified rejection sampling, then Guided-OPD distillation with teacher intervention phased out to zero, then RCF-DAPO, a variant of DAPO that compares rewards only within the same role and production context.
• In a shared tool environment on the authors' 100-task benchmark, the policy raises the Overall Score from 74.5 to 90.5, a 16.0-point gain over its base, and matches Claude Opus 5 at 89.9 while generating 2,377 versus 78,763 tokens per successful output minute, roughly 33× fewer, though Opus still edges it on visual-narration alignment at 74.8 versus 74.6.
• Against native end-to-end editors on the same tasks, it scores 90.5 versus 84.9 for Crayotter and 82.5 for VideoAgent, costs an estimated $0.09 per output minute versus $0.97 and $0.58, and wins a blinded three-rater human preference at 4.56 versus 3.48 and 3.46.
• The benchmark, metrics, and source videos are all the authors' own and share a single platform with the training data, the comparison models receive no task-specific tuning, the human ratings carry no significance test, the stage ablation is cumulative rather than factorial, and the authors name limited template coverage and retrieval robustness as open gaps.

arXiv:2610.38352 PDF

Zero-data recursive self-improvement loops tend to keep spending updates on a self-generated task pool after it stops yielding useful gradient, and when training targets come from majority voting the model's own errors become its next labels. A-RSI separates the two problems: it decides when to swap the pool purely from training logs, never from validation or test scores, and it gets answer correctness from a frozen library of model-written programs that render the image, question, and reference answer from one shared configuration.

• The Solver trains with GRPO in windows of 100 updates on a 400-task pool, using a reward of correctness plus a format bonus and a truncation penalty, and after at least three windows the controller replaces the whole pool with 400 unused random tasks under fixed family quotas (geometry, charts, tables, counting) if the count of distinct prompts with nonzero within-group advantage drops to at most 85% of its value two windows earlier or training accuracy rises by at least 10 points, carrying the model weights forward.
• On an eleven-benchmark suite (MathVision, MathVerse, MMMU-Pro, ChartQA, PhyX, and others), Qwen3-VL-8B-Instruct improves from 57.77 to 61.05 macro average over eight continuation windows with only 800 fresh tasks (the public base scores 56.65 before initialization), while MiMo-VL-7B-SFT goes from 59.08 to 60.71.
• At matched update budgets, refreshing every window reaches 60.64 using 3,200 new tasks, periodic refresh reaches 59.89 with 1,200, and never refreshing scores 58.35 at window 6 versus 60.40 for A-RSI before its training health gate fails and halts the run.
• Difficulty-aware replacement strategies (frontier, failure-prioritized, learnable-band, ladder controllers) all trail plain random draws by 0.05 to 2.63 points, and the gain over base retains 80% when the output limit doubles from 4096 to 8192 tokens, compared with 59% for MM-Zero v5 (3.25 versus 2.45 points above base).
• Each arm is a single run with no replications, the detector is not shown to beat a fixed three-window schedule (with frontier selection held fixed, log-triggered refresh trails periodic refresh by 0.30 points and the first two triggers coincided with a three-window cadence), and the +2.05-point edge over no-refresh decomposes into +2.66 from responses that now finish within the token budget and −0.60 on answers both arms completed, so most of the gain comes from completing rather than correcting answers.

arXiv:2610.38568 PDF

Proactive streaming-video models can only wait or answer, so when the information needed is missing from the video or the user's request they either stall indefinitely or hallucinate. AskStream adds a third action, ask, and benchmarks whether models know when and what to ask before answering under both video and user uncertainty.

• The benchmark contains 438 videos and 1,401 questions across eight subtasks (blurred key evidence, missing viewpoint, missing personal object or action concepts, ambiguous or false-premise questions, missing or outdated preferences), with 2,714 manually annotated frame-level wait, ask, and answer windows scored by Ask-F1, which rewards both the timing and the content of asks and answers.
• Offline models are run through StreamHarness, a hierarchical memory that keeps the last 10 frames at 1 fps and compresses older frames into mid-term and long-term summaries, with one wait, ask, or answer decision allowed per second.
• In the multiple-choice setting the best model, Gemini-3.8 Flash, reaches only 52.18% Ask-F1, while the best native streaming proactive model, AURA, scores 7.22% and several streaming models score zero on every subtask.
• Asking too early is the dominant error, and giving Gemini-3.8 Flash an oracle for when to ask on the blurred-evidence subtask lifts its Ask-F1 from 37.17% to 58.30%, whereas cross-turn memory adds only about 4 to 5 points and solely on user-uncertainty tasks.
• The open-ended results are judged by Gemini-3.8 Flash, the same model that tops the leaderboard, the main setting allows just one ask per turn, and offline models are evaluated through the authors' own harness, so the offline-versus-streaming gap partly reflects that wrapper.

arXiv:2610.38678 PDF

Small vision-language models trained for UI navigation with supervised fine-tuning tend to memorize trajectories and cannot recover from errors, while frontier models that navigate well cost dollars per task. The authors train an 8B planner with online reinforcement learning directly on hundreds of live iOS devices, pairing it with a frozen grounding model and rewarding every step with an LLM judge in addition to a terminal task-completion reward.

• The planner is warm-started from Qwen3-VL-8B-Instruct and trained with PPO using step-level advantages, where a Gemini-3.5-Flash judge scores each step as success, fail, or overcomplete and a Gemini-3.1-Pro judge scores the final outcome, with dense rewards discounted at γ=0.5 and the terminal reward left undiscounted to stop the agent from padding trajectories for step credit.
• GiGPPO stabilizes the critic by pairing each state with the closest screen by 64-bit perceptual hash among sibling rollouts and feeding that neighbor's return and Hamming distance into the value estimate, which on UINavBench lets training at steps 75 and 168 match standard PPO at steps 168 and 336, while dense rewards alone let PPO at one epoch match terminal-only PPO at two.
• On UINavBench, a benchmark of 116 tasks across 36 real iOS apps, PPO at step 336 reaches 32.4% versus 16.7% for the SFT baseline, tying Gemini-3.1-Pro and beating Gemini-3.5-Flash at 28.4% and Qwen3.8-27B at 19.6%, and on the out-of-distribution AndroidWorld it scores 32.8% versus 4.7% for SFT, the paper's "nearly sevenfold" claim.
• The infrastructure runs 512 concurrent trajectories on cloud-hosted iOS virtual machines, with episodes reset by snapshot rollback in roughly 10 seconds rather than by tearing down sessions.
• Gains on the synthetic-app iOSWorld benchmark are small and noisy, with the best RL checkpoint at 9.8% against 39.5% for Gemini-3.1-Pro and one PPO checkpoint dropping below the SFT baseline, frontier planners still lead on AndroidWorld by roughly 25 to 28 points, and the comparison carries caveats since frontier models were evaluated only as planners inside the authors' own harness, two of three benchmarks use Gemini-3.5-Flash as the grounder, and the training signal itself depends on Gemini judges.

arXiv:2610.39264 PDF

One- and few-step text-to-image generators must reproduce fine prompt detail, with long legible text as the stress test, despite having almost no refinement steps. IDM keeps the APEX-style internal distribution matching distillation (DMD), which estimates the score difference from two velocity branches of the same generator rather than a separate fake-score network, and adds an explicit loss matching feature means and per-coordinate standard deviations in the generator's own exponential-moving-average (EMA) feature space, so neither a discriminator nor an external encoder is needed.

• Decomposing the reverse-KL score difference into a Gaussian moment component (means and covariances) plus a non-Gaussian residual motivates keeping the full DMD correction while adding a moment constraint, implemented as a diagonal-Gaussian 2-Wasserstein loss on mean-pooled last-layer EMA activations at high noise, compared against detached queues of 16,384 model and reference features, on top of an RCGM base loss with weights 1 for the score term and 0.5 for the moment term.
• On Z-Image-Turbo 6B at 2 network function evaluations (NFE), IDM reaches 94.25% LongText-CN and 90.20% LongText-EN, the best among compared methods at that budget, beating Z-Image-Turbo++ by 2.63 and 0.32 points and APEX by 3.87 and 1.44 points, though the 8-NFE Z-Image-Turbo baseline still scores higher on LongText-EN at 91.74.
• On Qwen-Image 20B with full-parameter tuning at 2 NFE, IDM reaches 0.91 GenEval and FID 6.34 versus APEX at 0.90 and 6.44 at essentially the same latency (0.44 vs 0.45 s per sample), while rank-64 LoRA training of only 0.4B parameters retains 0.90 GenEval.
• Controlled ablations show the moment loss lifts the internal-DMD control on all five aggregate metrics (GenEval 77.46 to 77.81, LongText-CN 93.87 to 94.25, LongText-EN 89.79 to 90.20), removing DMD costs more than removing the moment loss (0.91 vs 0.35 GenEval points), EMA features perform on par with an external DINOv2 encoder (94.25 vs 94.13 LongText-CN) and better than teacher or live self-features, and equal-weight velocity-plus-endpoint regression beats either term alone.
• The objective's own contribution is small (tenths of a point), the larger system-level gaps over APEX are confounded by a far bigger text-rendering training set (5M synthesized samples versus APEX's 200K), APEX still leads on DPG-Bench (88.64 vs 86.82), the Z-Image results start from the already-distilled Turbo checkpoint rather than the base model, and the moment loss matches only pooled diagonal statistics across prompts rather than per-condition distributions.

arXiv:2610.39633 PDF

Retrieval-augmented generation pipelines usually serve a dedicated reranker next to the answer model, so every candidate passage is encoded once for scoring and again for generation on two resident backbones. Glance instead gives the generator a rank-16 LoRA adapter plus a 256-dimensional pointer head, letting one resident 9B backbone score every candidate from a single shared prefill and then answer with the adapter switched off.

• The candidates are prefilled once as a shared state, a short relevance probe of roughly 30 tokens branches per passage from the cached context (including the recurrent DeltaNet state on the hybrid MiMo-V2.6-Distill-Qwen-9B backbone), and the pointer head reads one score per passage without ever prefilling a separate query-passage pair, at about 0.1 GiB of extra memory per decision skill.
• After 39 minutes of MS MARCO fine-tuning on 4,000 queries, the packed_all format reaches nDCG@10 of 0.741 on SciFact and 0.345 on NFCorpus, against 0.721 and 0.326 for the bge-reranker-v2-m3 cross-encoder, with every zero-shot variant already beating BM25 before any retrieval training.
• On 300 HotpotQA distractor questions, the trained adapter's top-3 context raises answer F1 by +0.045 (0.684 to 0.729, paired bootstrap 95% CI 0.014 to 0.076) over cross-encoder-selected context under the same base generator, and an activated-adapter variant lets the answer continue directly from the passage cache instead of re-prefilling.
• Scaling the shared prefill helps rather than hurts once trained: quality keeps rising to 0.717 nDCG@10 at 200 candidates (43k tokens) versus 0.659 for the cross-encoder, while the zero-shot adapter peaks at 20 candidates and collapses at 200.
• The cross-encoder stays far faster per query (31 to 60 ms versus roughly 400 ms or more), evaluation covers only two BEIR datasets, one QA set and one backbone family with no significance tests on the BEIR gaps, memory grows with candidates times context length (56 GiB at 100 candidates), scores shift somewhat with passage order, and in the cache-reuse variant the generator sees all passages rather than the reranked subset.

arXiv:2610.40454 PDF

Pixel-space generative Transformers spend one token per image patch, so finer output detail means a longer global-attention sequence and quadratic cost growth. SpecSlot breaks that coupling by expanding each spatial anchor into K slots that each own a disjoint radial band of discrete cosine transform (DCT) coefficients, and routing only the lowest-frequency slot per anchor through global attention while all slots mix locally within the anchor.

• The model reads W anchors from non-overlapping input cells, expands each into K slots, runs a global-global-local-local (GGLL) schedule three times where slot 0 is the only cross-anchor carrier, and decodes with a 50%-lapped DCT synthesis plus a zero-initialized spatial correction head, all under a pixel-space rectified-flow loss; at W256/K4 on 256px images the global sequence holds 288 tokens instead of 1,056.
• On ImageNet-256, W256/K4 reaches FID 6.59 against 6.68 for the FLOPs-matched JiT-CM control (a statistical tie that converges faster), using 40.9% fewer FLOPs than JiT-B/8, which still leads the panel at FID 4.34.
• On ImageNet-512, it scores FID 9.14 versus 11.98 for JiT-CM (2.84 better at 3.9% fewer FLOPs) and lands within 0.69 FID of JiT-B/16 at 38.2% lower cost and 1.6× higher throughput, with the authors attributing the larger margin to each anchor covering four times more pixels while the global sequence stays fixed.
• Transferred to text-to-image, the same architecture beats JiT-B/8 on DPG-Bench (62.21 vs 61.15) at 40.6% lower forward cost, though JiT-B/8 keeps the GenEval lead (29.78 vs 26.99), and ablations show that moving the global carrier to any other slot worsens FID while full-global attention still wins (5.84 vs 6.59) at 74.8% more compute.
• The headline gains are measured against an implementation-defined compute-matched control rather than a published JiT size, the paper concedes the cause of the larger 512 margin is not isolated, extra spectral slots show diminishing returns (effective slot utilization falls from 0.803 at K2 to 0.495 at K8), and evaluation is limited to base-scale models on ImageNet-1K and 256px text-to-image.

arXiv:2610.40647 PDF