AI papers — 2026-08-24

Problem: As LLM-based tasks grow more complex, a single agent's execution loop cannot adequately schedule parallel/interdependent subtasks, integrate heterogeneous specialized expertise and verification, or maintain persistent, recoverable runtime state, limiting what the authors call "Individual Intelligence."

Method: This is a survey that introduces "Graph Engineering," a paradigm using explicit, dynamic graph structures to represent tasks, agents, and system states in order to achieve "System Intelligence." It organizes the field along three axes: Task Organization (goal decomposition and workflow optimization, e.g., HuggingGPT, ReWOO, LLMCompiler's dataflow DAG compilation, Plan-over-Graph, TDAG, Flow), Agent Coordination (agent capability modeling, team topology, and communication, e.g., MetaGPT, AgentVerse, G-Designer, DyLAN, MasRouter), and Runtime State Management (state recording, fault localization, and failure recovery, e.g., LangGraph, SagaLLM, AgentGit, Shepherd). It also traces the broader evolutionary taxonomy from Model Intelligence (Pre-training/Post-training, Prompt/Context Engineering) through Individual Intelligence (Harness Engineering and Loop Engineering) to System Intelligence (Graph Engineering) and a proposed future stage, Ontology Engineering.

Key results:

• Not stated (this is a survey/taxonomy paper reviewing methods, benchmarks, and open-source libraries rather than reporting original experimental results)

Why it matters / caveats: The paper proposes a structural framework and vocabulary (Graph Engineering, System Intelligence) for organizing the rapidly growing multi-agent/LLM-agent literature, and provides an accompanying resource list (github.com/DEEP-JLU/Awesome-Graph-Engineering); as a survey it synthesizes and categorizes prior work rather than introducing or benchmarking a new model or system itself.

Problem: Finding the optimal learning rate for pretraining large-scale Mixture-of-Experts (MoE) models via exhaustive hyperparameter sweeps becomes computationally prohibitive at extreme model sizes and trillion-token training budgets.

Method: The authors propose a two-step hyperparameter transfer framework: (1) a Maximal Update Parameterization (µP) adaptation for MoE architectures using Multi-head Latent Attention (MLA) and the Muon optimizer, which classifies parameters as vector-like or matrix-like and applies learning-rate scaling only to matrix-like hidden weights (including router and expert FC1 weights) so optimal learning rates transfer across width- and sparsity-scaled models; and (2) a token-scale extrapolation method that trains small proxy models with a Warmup-Stable-Decay scheduler, applies Exponential Moving Average (EMA, α=0.6) to checkpoints every ~2B tokens, fits a second-order polynomial between log-learning-rate and validation loss to estimate the optimal LR at each token budget, then fits a log-log linear regression across token scales to extrapolate to a 10T-token horizon. The method is validated by pretraining a foundation MoE model (155B total, 17B active parameters) on Kakao/Upstage's internal Megatron-LM fork using NVIDIA H200 GPUs.

Key results:

• µP enables consistent learning-rate transfer across MoE models scaled 2×, 4×, and 8× in width (up to 30.7B total / 3.6B active params), whereas Standard Parameterization (SP) fails to transfer.
• The log-log linear regression of optimal LR vs. token budget achieves R²=0.95, extrapolating an optimal learning rate of 3.85×10⁻⁴ for 10T-token training from proxy runs at 255B–502B tokens.
• The full-scale 155B/17B-parameter target training required ~98× the total compute of the proxy runs (vs. an additional 240.3 ZFLOPs for a conventional 2D model-scale sweep on top of the 64.8 ZFLOPs proxy cost).
• The resulting model, trained on 10T tokens, achieves training-loss stability with no spikes and lies on the Pareto frontier of estimated training compute vs. MMLU-Pro accuracy, outperforming dots.llm1 and GLM-4.5-Air at comparable or lower compute.

Why it matters / caveats: The approach replaces costly 2D (model-scale × token-scale) hyperparameter sweeps with a lightweight 1D token-dimension search plus zero-shot width transfer, making learning-rate tuning tractable for 100B+ parameter MoE models. The authors note they cannot exhaustively verify optimality of the predicted LR due to infeasible full-scale sweeps, and acknowledge that sparsity and width scaling are coupled rather than independently disentangled in their large-scale recipe, leaving per-expert LR adaptation and isolated sparsity-axis transfer as future work.

Problem: Existing instruction-based video editors assume an in-place, fixed-time-span setting that fails for open-ended streams (e.g., restyling a live game or applying a camera move to an ongoing shot), where edits must instead extend to future frames as they arrive.

Method: InfinityEdit is a lightweight edit adapter attached to a frozen Helios-Distilled (14B autoregressive video diffusion transformer) backbone, inserting a trainable adapter block after each transformer layer with three attention modules — history cross-attention (anchors the denoising chunk to input frames), temporal causal self-attention (propagates cues only from earlier to later frames), and edit cross-attention (injects the edit instruction). Training uses a flow-matching objective with history corruption for exposure-bias mitigation, a mixture-Gaussian noise-level sampler, and a two-phase curriculum (uniform coverage then detail refinement); training data comes from a custom pipeline synthesizing (source, edit, target) triplets from UltraVideo sources, Gemini-3-Flash-generated edit instructions, Qwen-Image-Edit-2511 frame edits, and Wan2.2-I2V-A14B target generation. At inference, an "ignite-then-continue" strategy activates the adapter only on the chunk where an edit arrives, then lets the frozen backbone continue with an anchored sliding-window history and a reset anchor frame.

Key results:

• On a 200-sample, 3-round OOD sequential-editing benchmark, InfinityEdit achieves the best Camera Motion score (0.7654 vs. 0.5432 for the next-best baseline, a gain of about +0.22) and best Temporal Flickering (0.9660) among compared methods.
• VLM-as-Judge (Gemini-3.5-Flash) scores: InfinityEdit ranks first on all four dimensions — Edit Faithfulness 3.828, Visual Quality 3.765, Preservation 3.815, Coherence 3.840 — versus the next-best baseline (SANA-Streaming) at 3.303/3.093/3.060/3.030.
• Edit faithfulness stays stable across 3 editing rounds (3.860/3.805/3.820), with a standard deviation of only 0.023 across rounds, versus 0.268–0.715 for baselines; Aesthetic Quality stays nearly flat (Δ≈0) across rounds while baselines drop 0.02–0.04.
• Trained on 32 NVIDIA H20 GPUs at 384×640 resolution and 16 fps; demonstrated stable long-video editing over sequences exceeding 1000 frames without quality collapse.

Why it matters / caveats: The paper defines a new task (infinite video editing) with a dedicated benchmark and shows a frozen-backbone-plus-lightweight-adapter design can inject edits while preserving a pretrained streaming generator's long-video stability, avoiding costly full fine-tuning. Caveats acknowledged by the authors: it currently supports only natural-language instructions (no image/video reference conditioning), and transitions at edit boundaries can still be abrupt despite overall coherence.

Problem: Evaluating omni-modal LLMs as real-time interactive video assistants is difficult because a model's response changes the user's subsequent actions, creating diverging interaction paths that static offline video datasets cannot capture.

Method: The authors build OmniAssistBench by reverse-engineering existing Internet videos (mainly YouTube, plus curated clips from action-recognition, instructional, and emotion-analysis datasets): experts deduce fixed "prior knowledge" (a single canonical procedural path) from each source video, design multi-turn open-ended QA pairs with ground-truth sentences and key-point phrases, embed user prompts as TTS audio (or picture-in-picture video clips for gesture/OCR prompts) at clip boundaries, and segment videos into multi-turn clips to simulate streaming interaction. The benchmark uses a two-tier task taxonomy (Basic Interactive Understanding: Social/Temporal/Referential Perception, Non-audio Prompt Following; Advanced Interactive Understanding: Context-aware Response, Proactive Response, Process Tracking) totaling 7 major tasks/16 sub-tasks, plus 3 filmed Real World Cases (Meeting Simulation, Blind Assistance, Handicraft Process Tracking). Scoring uses an LLM judge (GPT-5, with GLM-5/DeepSeek-v3.2 for consistency checks) on a rubric-based 0–5 (normalized to 0–100) scale that penalizes redundancy and hallucination.

Key results:

• Gemini-3-Pro (best proprietary) scores 66.4/100 overall; Qwen3-Omni-Instruct (30B-A3B, best open-source) scores 51.2/100.
• Dataset statistics: 300 videos, 685 QA pairs, average video length 182.25s, built with over 1000 expert person-hours.
• On Real World Cases, closed-source models average 51.2 vs. 34.8 for open-source models; VITA-1.5 (7B) scores lowest overall at 24.6.
• Ablations: modality removal generally hurts scores (audio removal worse than vision removal); judge-model agreement is high (Pearson correlations 0.75–0.91 across GPT-5/GLM-5/DeepSeek-v3.2); resizing video to 360px min-edge for Qwen3-Omni-Instruct gains ~50s more context but yields no significant overall score change.

Why it matters / caveats: Results show current Omni-LLMs understand instructions but frequently give incorrect/incomplete answers, with four identified bottlenecks: poor gesture/visual-prompt following, limited long-term memory (context exhausted within minutes, e.g., MiniCPM-o-2.6 retains only ~380s), failure to delay responses until the right moment, and loss of cross-turn goal context — indicating substantial room for improvement before such models are reliable real-world assistants.

Problem: Parallel reasoning improves large reasoning model accuracy by exploring multiple solution branches, but existing control signals (final-answer consensus, token-level confidence, isolated intermediate probes) are delayed, weakly tied to reasoning progress, or too noisy for dynamic branch-level computation control.

Method: ParaTempo is a training-free, asynchronous parallel reasoning framework driven by "temporal confidence" — a branch-local signal computed by periodically probing each active branch (every τ=500 tokens) for a tentative answer-distribution, aggregating recent probe distributions over a sliding window (size W), and taking the exponentiated negative entropy of the aggregated distribution. Using this signal, a controller performs a warmup calibration phase (quantile-based pruning threshold), then asynchronously prunes low-confidence branches, retires branches whose dominant answer stays concentrated over X consecutive probes, forks new branches from high-confidence donors into freed compute slots, and triggers global early termination via a confidence-weighted vote. It is evaluated on Qwen3.5-35B-A3B and GPT-OSS-20B using vLLM, against baselines Zero-shot, Self-Consistency (SC), ESC, SAC, DeepConf, and Parallel-Probe on AIME 2026, HMMT November 2025, HMMT February 2026, and GPQA Diamond.

Key results:

• Versus standard self-consistency (SC), ParaTempo reduces average latency by 21.8–32.2% and total generated tokens by 18.1–30.3% while maintaining competitive accuracy (e.g., on Qwen3.5-35B-A3B: 71.1% avg accuracy vs. SC's 72.2%, within 1.1 points).
• Versus Parallel-Probe (strongest parallel controller baseline), ParaTempo improves accuracy by 3.9 points while reducing latency by 10.6% on Qwen3.5-35B-A3B, and achieves a 3.8-point accuracy improvement with comparable sequential cost on GPT-OSS-20B.
• Ablation on HMMT25 (Qwen3.5-35B-A3B): removing early retirement drops accuracy by 6.6 points (73.3%→66.7%); disabling forking drops accuracy by 3.3 points (73.3%→70.0%); removing pruning increases latency (205.7s→233.0s) with only marginal accuracy change.
• Preliminary study: instantaneous answer confidence has Spearman |ρ|=0.41 and AUC=0.71 for predicting future answer stability, versus |ρ|≤0.13 and AUC≤0.58 for token entropy/perplexity, motivating temporal confidence's design.

Why it matters / caveats: ParaTempo removes the cross-branch synchronization barrier that limits prior parallel-reasoning controllers, letting compute be reallocated asynchronously toward converging branches, which yields a better latency-accuracy Pareto frontier. Results are reported on two specific open models (Qwen3.5-35B-A3B, GPT-OSS-20B) and four math/science benchmarks on a single A100 GPU with vLLM; generalization to other model families or task types is not evaluated in the extracted text.

Problem: The generalization behavior of on-policy distillation (OPD) — where a student LLM is trained via reverse-KL supervision from a teacher on the student's own sampled trajectories — is poorly understood because prior work evaluates OPD only within a single domain on benchmarks close to the training data, leaving unclear how it behaves under distribution shifts, cross-domain transfer, and multi-teacher settings.

Method: The authors run a controlled study varying one generalization factor at a time, using PG-style OPD (reverse-KL estimated via sampled-token k1 approximation) across teachers including Qwen3-32B, Light-R1-14B, Polaris-7B/4B, OpenMath-Nemotron-1.5B/7B, JustRL-1.5B, Nemotron-1.5B, DeepScaleR-1.5B, and VibeThinker-1.5B, and students including DeepSeek-R1-Distill-Qwen-14B/7B/1.5B, Qwen3-8B-SFT, and Qwen3-4B. They distinguish same-origin (teacher/student share a base model) from cross-origin pairs, and study three research questions: in-domain distribution shift (training difficulty, English→Chinese, short→long reasoning horizon), cross-domain transfer (math, code, science, instruction-following), and multi-teacher OPD (MOPD), training on datasets like BigMath, DeepCoder-Preview, TextbookReasoning/SCP-116K, and Nemotron-Post-Training-IF, evaluated on benchmarks such as AMC2023, MATH-500, AIME2025/2026, BeyondAIME, OlymMATH, LiveCodeBench, GPQA-Diamond, and IF-Eval.

Key results:

• Training-problem difficulty barely matters: BigMath easy, hard, and random subsets converge to nearly identical final accuracy; even GSM8K/hardest DeepMath-103K slices recover over 80% of the OPD gain of BigMath-random.
• Dynamically discarding problems the student already solves gives a small but consistent gain, e.g., +0.6pp average (41.4%→42.0%) for Polaris-7B→DS-distill-1.5B and +0.4pp (52.4%→52.8%) for Light-R1-14B→DS-distill-7B across six math benchmarks.
• Same-origin teachers generalize far better than cross-origin ones: e.g., same-origin Nemotron-1.5B reaches 40.5% GPQA-Diamond (a 40.5% relative gain) vs. cross-origin JustRL-1.5B only 14.9%; on IF-Eval, JustRL-1.5B (same-origin) achieves a 64.8% gain vs. Nemotron-1.5B's 42.0% decline in the reversed setting.
• In MOPD, changing teacher mixture ratios (e.g., JustRL/Nemotron math-teacher share from 25/2 to 2/25) produces a "seesaw": student scores on GPQA-Diamond, LiveCodeBench, and IF-Eval shift toward whichever teacher gets a larger prompt share, regardless of assigned domain (math accuracy ranges from 25.1% to 27.1% average across mixture ratios).

Why it matters / caveats: The findings show OPD transfers a teacher's general reasoning behavior rather than problem-specific answers, and this broad transfer is a "double-edged sword": it enables strong generalization for same-origin teacher-student pairs but undermines the assumption in multi-teacher OPD that routing prompts to domain experts isolates their contributions, instead causing capability seesaws that complicate multi-teacher training design.

Problem: Existing multimodal image re-rankers either collapse multi-faceted query relevance into an opaque embedding or use free-form chain-of-thought reasoning that omits implicit constraints, hallucinates evidence, and covers semantic dimensions inconsistently across queries.

Method: EviRank recasts multimodal image re-ranking as semantic constraint satisfaction by parsing any query (text-only, image-only, or composed) into a modality-agnostic Evidence Frame of typed statements — required, forbidden, and ignorable — across six semantic slots (entities, attributes, actions, relations, scene, key details). A training-free MLLM teacher (Gemini-3-flash or Gemini-3-pro) mines this evidence in a two-step process (query normalization then evidence extraction), and re-ranking combines a deterministic rubric score (embedding-based matching with a forbidden-constraint penalty and cross-modal consistency term) with an evidence-grounded listwise refinement over the top-M candidates. The resulting structured supervision bundle (evidence, slot-wise satisfaction rates, calibrated listwise scores, confidence, hard pairs) is distilled via a combined score-distillation, hard-pair margin, and slot-wise BCE loss into a lightweight student, Qwen3-VL-2B-Thinking, trained on 20k queries across five datasets.

Key results:

• On Flickr30k with BLIP-2, EviRank-pro reaches 95.61% R@1, surpassing CoTMR by 6.32 points; on COCO with CLIP-ViT-L/14 it reaches 69.53% R@1 (+9.6 over CoTRR).
• On composed retrieval (I→I), EviRank-pro achieves 91.46% R@1 on Stanford Online Products and 86.89% on CUB-200, exceeding LoCoRE-base by 7.66 and 8.59 points respectively; on FashionIQ, R@10 improves up to 8.27 points on Shirt over ImageScope.
• The distilled student (EviRank) preserves over 90% of the teacher's capability while running at ~800ms/query (vs. ~382ms for CLIP coarse retrieval alone) with ~10 R@1 points gain; structured evidence extraction is stable across 10 repeated runs, prompt perturbations, and different teacher models (Kendall's τ≥0.89, Top-1 agreement≥91%, R@10 std ≤1.3).

Why it matters / caveats: The paper argues structured, typed evidence (especially forbidden constraints, often overlooked in retrieval design) is more robust and interpretable than free-form CoT, and rubric scoring and listwise comparison are shown to be complementary rather than redundant. Limitations noted by the authors: evaluation is limited to English public benchmarks and still images (no multilingual, evolving-distribution, video/3D/audio-visual settings), and no large-scale human-facing production user studies were conducted.

Problem: Hybrid-thinking multimodal large language models (MLLMs) that switch between deliberative thinking and fast non-thinking modes can produce inconsistent user-visible response quality across modes even when correctness-focused post-training keeps task accuracy high.

Method: The authors introduce PatternEval, a 2,415-prompt diagnostic benchmark (nine constituent categories across visual perception/grounding, OCR/structured-image understanding, and multimodal knowledge reasoning) that scores responses for four failure types — chain-of-thought leakage, response repetition, logical contradiction, and performative reasoning — using Qwen3-Max as the correctness verifier and Seed-2.0-Pro (with image access) as the calibrated pattern judge. They then train PatternRM, a Qwen3.5-27B-initialized reward model predicting the four binary pattern labels from 57,578 SFT instances, and use it to build PatternRL, which adds a category-weighted auxiliary penalty (CoT/Rep weight 0.05, Con/PR weight 0.02) on top of a GRPO correctness reward, applied to Qwen3-VL-4B-Instruct and Qwen3-VL-8B-Instruct on a 44,200-prompt multimodal RL mixture.

Key results:

• Across 25 evaluated model configurations (e.g., GPT-5.4/5.5, Seed-2.0/2.1-Pro, Claude-Opus-4.6/4.8, Kimi-K2.5/2.6, Qwen3.5/3.6/3.7 series), non-thinking mode Trigger rate exceeds thinking mode in every pair, with gaps up to 48.64 percentage points (Kimi-K2.6) and exceeding 20 points in 17 of 25 pairs.
• PatternRL reduces non-thinking Trigger rate by 13.08 points on Qwen3-VL-4B and by 14.35 points on Qwen3-VL-8B relative to correctness-only BaseRL, while aggregate accuracy changes by less than 1 percentage point.
• Meta-judge calibration on 2,500 human-annotated responses: GPT-5.5 achieves the highest overall pattern-judge F1 (77.2%/77.0%), while CoT-leakage F1 is high (90.2–95.3%) versus much lower contradiction (56.1–62.5%) and performative-reasoning F1 (41.8–64.5%).
• Response length correlates with Trigger rate (Pearson r=0.64 non-thinking, r=0.84 thinking); in the longest length sextile, non-thinking Trigger reaches ~86% for incorrect and ~56% for correct responses, versus ~52%/~22% under thinking.

Why it matters / caveats: The paper shows correctness metrics alone mask substantial user-facing quality gaps between hybrid-thinking modes, and that lightweight RL-stage pattern penalties can reduce but not eliminate these failures — PatternRL introduces an accuracy trade-off that is notably larger for the smaller 4B model than the 8B model on general reasoning benchmarks, suggesting deeper fixes may require intervening earlier than the RL stage (e.g., SFT/mid-training).

Problem: Semantic vision encoders (e.g., SigLIP, CLIP) produce tokens optimized for semantic abstraction that discard fine-grained pixel detail, forcing unified multimodal models to rely on separate, fragmented representation spaces (a semantic ViT plus a VAE latent space) for understanding versus generation/editing.

Method: The authors introduce Patch Reparameterization: the pretrained patch embedding of a frozen semantic ViT is kept for understanding, while a new reconstruction-aware patch embedding is added and fed through the same frozen Transformer blocks to recover pixel detail; the two token streams are combined via an explicit channel-wise Token Fusion (with the reconstruction stream projected/compressed) rather than an entangled MLP merge, and trained with a balanced flow-matching objective (λr=0.75) plus an RAE-style ViT decoder. This produces a family of unified tokenizers — PR-SigLIP2, PR-DINOv2, PR-Qwen-ViT — which are then scaled into UniSpace, an 8B Qwen3-based Mixture-of-Transformer-Experts model (following BAGEL's MoT design) that uses PR-Qwen-ViT as its sole frozen visual interface (no separate VAE) for understanding, text-to-image generation, and instruction-based editing, trained on a progressive multi-stage curriculum (256→512→1024 resolution) plus SFT.

Key results:

• PR-DINOv2 reconstruction: rFID 0.14, PSNR 30.84, SSIM 0.90 — vs. matched-backbone RAE (DINOv2-B) rFID 0.57 (75.4% relative reduction) and RAEv2 (DINOv3-L) rFID 0.29/PSNR 22.57.
• Multimodal understanding preserved: PR-SigLIP2 average 64.37 vs. SigLIP2-B baseline 63.39; PR-Qwen-ViT 68.94 vs. Qwen-ViT baseline 68.29 (LLaVA-v1.5 protocol).
• UniSpace (8B) on ImgEdit: overall score 4.28, beating SenseNova-U1 (3.90) and BAGEL (3.20); on GEdit: avg 7.395; on GenEval: 0.84 overall; on DPG-Bench: 86.49 overall; on OneIG-Bench: bilingual average 0.547, edging out Emu3.5 (0.546) and SenseNova-U1 (0.542).
• Diagnostic ablation: replacing SigLIP2's patch embedding with a random projection (all Transformer blocks frozen) raised last-layer PSNR by 3.70 dB (20.96→24.66), showing the bottleneck is the patch embedding, not the frozen backbone.

Why it matters / caveats: The work shows a single frozen, reparameterized semantic ViT can serve as the sole visual interface for understanding, generation, and editing at scale, eliminating the separate VAE pathway used by prior MoT systems like BAGEL. The authors note UniSpace's system-level understanding still lags dedicated VLMs, and because the encoder is frozen/shared, it cannot be further updated on understanding data without risking generative capability, limiting further understanding gains.

Problem: Agent training environments are typically built task-centrically (constructed manually or synthesized around predefined benchmark tasks), which limits how well they scale to reflect the diverse, evolving, multi-service workflows found in real business settings.

Method: AgentMercury introduces PLANET, an explicit environment-authoring role that generates a persistent, executable world (entities, services, tools, state, transition logic, and executable cross-service invariants) from a high-level business scenario, from which multiple tasks and trajectories are subsequently instantiated via a separate TASK generator; grading is deterministic (GRADE) using task rubrics plus hidden world-level invariants checked via SQL-based verifiers. The team synthesized 4,783 executable environments across 14 industries and 50 countries, and trained Qwen3.5-4B and Qwen3.5-35B-A3B policies on 43,300 task instances using GRPO (with Dr. GRPO coefficients) and single-rollout asynchronous optimization (SAO). Separately, they fine-tuned Qwen3.5-35B-A3B on 29,823 construction-trace samples to teach the world-authoring process itself.

Key results:

• Qwen3.5-4B + GRPO on EnterpriseOps-GYM improves average score from 12.3 to 15.7 (+27.6%); Qwen3.5-35B-A3B improves from 24.8 to 28.1 with GRPO and 28.3 with SAO.
• Out-of-domain transfer despite no benchmark-specific training data: Qwen3.5-4B improves AIME26 45.9→56.0, HMMT 28.5→35.4, LiveCodeBench 36.6→44.0, SciCode 22.6→25.7, GPQA-Diamond 76.5→77.5, and τ3-Airline 48.8→58.7.
• Fine-tuning Qwen3.5-35B-A3B on AgentMercury construction traces raises executable-world authoring oracle-pass rate from 3.3% to 83.3% on 30 held-out business briefs (p=1.2×10⁻¹⁰), matching strongest off-the-shelf API models (mean 78.3–80.7%).
• At 35B scale, BFCL (tool use) improves markedly with SAO from 31.1 to 42.1, and τ3-Telecom variance drops sharply (49.1±49.8 → 65.5±23.8), indicating more consistent policy behavior, not just higher average reward.

Why it matters / caveats: The results suggest environment diversity/realism — not benchmark-targeted task count — drives transferable agentic learning signals, and that environment construction itself can become a learnable model capability rather than a fixed engineering pipeline. Caveats: recipe-conditioning does not reliably help and can even hurt the fine-tuned authoring model (83.3%→10.0%); CSM performance regresses after training (9.2→5.6) for the 4B GRPO model; and the system does not yet close the loop with a learned world model to target environment generation at policy capability gaps.

Problem: Existing LLM/VLM-based user simulators fail to reproduce faithful multi-turn e-commerce browsing sessions because they mishandle long-range browsing history (either discarding it or naively concatenating it) and rely on step-wise imitation objectives that cannot detect or correct unrealistic session-level patterns like over-exploration or excessive passivity.

Method: The paper presents RecVerse, a GUI-grounded simulation agent (backbone: Qwen3.5-2B, with a Qwen3.5-4B scaling variant, trained via Megatron-LM) that perceives e-commerce pages through screenshots. It uses a cognitive-inspired hierarchical memory — Working Memory (FIFO short-term visuospatial buffer), Episodic Memory (textual in-session behavioral records), and Preference Memory (distilled high-level user intent) — where memory writes are treated as part of the action space ("memory-as-action") and learned jointly. Training proceeds via imitation-learning warm-start followed by trajectory-level GRPO reinforcement learning with a composite reward combining a macro-level reward (aligns aggregate action-type distributions with real users), a micro-level reward (intent-weighted hierarchical category-matching), and a format reward. The authors also release USB (User Simulation Benchmark), a 5,274-trajectory interactive GUI e-commerce dataset with pixel-level screenshots and an online multi-turn RL environment.

Key results:

• RecVerse-GUI with RL improves item-level F1 from 4.27 (STA baseline) to 7.19, HR from 5.92 to 10.45, and HCO (Hierarchical Category Overlap) from 23.11 to 32.64.
• Moving from text to GUI modality raises HCO from 19.03 to 30.00 under IL and from 24.38 to 32.64 under RL.
• In pairwise human evaluation (Fleiss' κ=0.834), real users were preferred over RecVerse in 74% of comparisons (vs. 98% for the STA baseline), and RecVerse was preferred over STA in 92% of head-to-head comparisons.
• USB dataset statistics: 5,274 trajectories, 8 action types, 69,842 actions, average trajectory length 13.24 steps, 90,095 items across 41/517/2,256 L1/L2/L3 categories, 5,222 users.

Why it matters / caveats: RecVerse narrows but does not close the gap to real human browsing behavior (annotators still preferred real users 74% of the time), and the authors note that fully human-like simulation would require deeper personalization and richer individual preference modeling; the work targets offline evaluation and RL-based recommender training as its primary use cases.

Problem: Small language models are typically designed like large GPU-serving models and only adapted for CPU afterward, whereas the batch-size-one, memory-bandwidth-bound, growing-KV-cache regime of single-user CPU decoding calls for an architecture chosen for that target from the start.

Method: Daedalus-150M (160.49M parameters) uses 18 blocks at d_model=768 with 6 full-attention blocks (grouped-query attention, 4 KV heads/12 query heads, RoPE) interleaved among 12 depthwise short-convolution blocks (kernel length 3, state fixed at 2 timesteps regardless of context length), plus tied embeddings and a 2048-wide feed-forward inner dimension. It was trained from scratch on 59.9B tokens (124,476 steps) from a ten-source English corpus (FineWeb-Edu, DCLM-baseline, Stack-Edu, FinePDFs-Edu, etc.), using Muon for 2D weight matrices and AdamW for embeddings/norms/biases, a WSD schedule with linear decay to zero, and 4-bit Q4_0 post-training quantization for deployment. A parameter-matched, pre-registered ablation compared the hybrid against a 24-layer all-attention dense twin (161.25M params) trained identically on 5B tokens to isolate the architecture's effect.

Key results:

• Scores 47.31 on a five-task benchmark (HellaSwag, ARC-Easy, PIQA, OpenBookQA, WinoGrande) against a pre-registered bar of 42.20, beating GPT-2 124M (42.2), OPT-125M (42.1), GPT-neo-125M (41.9), Pythia-160M (41.0), and MobileLLM-125M's published 46.3 (despite MobileLLM using 1T tokens vs. 59.9B).
• In the parameter-matched ablation, the hybrid beats the dense twin's validation bpb by 0.81% (0.910398 vs 0.917774), ties on five-task mean (44.68 vs 44.82), produces a 6.3% smaller 4-bit file (95.56 MiB vs 101.62 MiB), and decodes 1.76x faster at 2048-token context (739.3 vs 420.3 tok/s) — versus only a 1.17x speedup predicted by a first-order bandwidth cost model.
• Against an external 135M-parameter peer on the same harness, Daedalus decodes 2.08x faster at 2048 tokens (648.6 vs 312.4 tok/s) while carrying 19% more parameters; ~47.9% of convolution channels are inert (dead), representing ~13.6M/8.5% wasted parameters.

Why it matters / caveats: The result shows that shifting the attention-to-recurrence layer ratio (based on a CPU memory-traffic argument) yields decode speedups that grow with context length and cannot be explained by bandwidth savings alone, at no downstream quality cost — but the paper reports unresolved issues: quantization-aware training failed (non-finite loss) leaving a ~6% perplexity penalty instead of ~2.5%, the training mixture drifted beyond its pre-set 10% deviation limit, the vocabulary (49,152 tokens) is oversized for a 150M model, and all results come from a single seed on English-only data.

Problem: Human-centric research has advanced deeply within individual tasks (pose estimation, action recognition, motion/video generation, human-object interaction) but remains fragmented across tasks, modalities, and communities, with no existing survey systematically connecting the full spectrum of human-centric intelligence or examining its methodological evolution in the foundation-model era.

Method: The survey introduces a "full-spectrum human context taxonomy" with six interconnected levels organized under three perspectives: observable subjects (visual appearance, spatial geometry), dynamic actors (kinematic dynamics, interaction modeling), and situated agents (world simulation, embodied agency). It then reviews methodological foundations across three axes — five human-centric data/signal families (visual imaging, spatial-structural, sensorimotor, wireless/ranging, linguistic/acoustic), four computational architecture paradigms (single-step mapping, sequential factorization, iterative generative, and hybrid architectures), and optimization strategies split into seven training mechanisms and five inference-time mechanisms. Using this framework, it systematically reviews representative methods (e.g., Sapiens/Sapiens2, SapiensID, CosmicMan, SMPLer-X/SMPLest-X, SAM 3D Body, MotionLLM, HIS-GPT) level by level, tabulating each with modality, data scale, model size, training/inference strategy, and architecture type.

Key results:

• Not stated (this is a survey/taxonomy paper organizing and cataloguing prior work rather than reporting original experimental results; cited numbers like Sapiens2's 1B-image training set or CosmicMan's 6M images are dataset/model-scale statistics from surveyed papers, not the survey's own findings)

Why it matters / caveats: The survey argues that human-centric AI has not kept pace with foundation-model progress and offers a unifying taxonomy, methodological foundation, and evaluation reference intended to guide future work toward scalable, trustworthy, physically grounded, and deployable human-centric intelligence; it explicitly excludes human-computer interaction, human-oriented language modeling, and policy studies of human-centered AI.

Problem: Globally applied LLM safety tuning (e.g., SFT, standard LoRA) improves refusal of harmful prompts but also perturbs behavior on benign prompts, causing utility degradation and over-refusal (the "alignment tax").

Method: CLEAR (Continuous LatEnt Adapter Routing) freezes the backbone LLM and inserts a safety-specific LoRA adapter (on q_proj and v_proj) whose strength is continuously scaled by a learned scalar gate g(x)∈[0,1] computed from a lightweight MLP over aggregated hidden states from three intermediate/later transformer layers, via h' = (W + g·ΔW)h. The gate is trained on WildJailbreak prompts (vanilla/adversarial benign and harmful subtypes) with a subtype-weighted binary cross-entropy loss plus a hard pairwise margin loss to separate harmful/benign gate scores, while the LoRA adapter is trained only on unsafe prompt-completion pairs; gate and adapter parameters are jointly optimized with AdamW while the backbone stays frozen.

Key results:

• On Llama-3-8B-Instruct, CLEAR reduces HarmBench ASR from 32.25% (base) to 0.50%, versus 2.00% (SFT) and 0.00% (LoRA), while retaining 73.46% GSM8K accuracy vs. 66.34% (SFT) and 66.72% (LoRA) — up to ~7.1 points higher.
• On Gemma-2-2B-it, CLEAR achieves 0.00% HarmBench ASR and 96.00% unsafe refusal, improving GSM8K to 41.62% and TruthfulQA MC2 to 53.19% (vs. ~38%/~47-48% for SFT/LoRA).
• Against safety-aligned baselines, CLEAR (Llama-3 family) cuts HarmBench ASR to 0.50% vs. 15.50-24.00% for Llama-3-8B-RR, SACPO, P-SACPO, and Beaver-7B; the CLEAR gate uses only 664,129 parameters vs. 278.8M (PromptGuard) and 8.03B (Llama Guard 3 8B).
• Under latent adaptive attacks on Gemma, CLEAR achieves ASRs of 0.0%/7.1%/18.4% across three injection depths vs. 54.0%/48.0%/31.0% (LoRA) and 47.0%/35.0%/22.0% (Full SFT).

Why it matters / caveats: CLEAR offers a parameter-efficient way to decouple safety intervention from general capability, improving robustness to latent perturbation attacks as well as standard jailbreaks. The authors note it is not uniformly best on every metric, depends critically on gate reliability (benign prompts can still trigger unnecessary intervention and low-scored unsafe prompts can bypass the adapter), and was evaluated mainly on relatively small open-weight models in single-turn text settings.

Problem: Vision-language models are too memory- and compute-heavy for mobile deployment, and aggressive low-bit quantization of VLMs is difficult in practice, especially without access to the original training data.

Method: The authors quantize Llama 3.2 11B Vision Instruct using a quantization-aware training (QAT) pipeline with knowledge distillation, where the full-precision bfloat16 model acts as teacher and a straight-through estimator propagates gradients through the quantization operator; training data is synthetically generated by prompting the teacher model itself on ImageNet images with randomly sampled instruction-style prompts. They introduce S3D8 (signed 3D 8-bit), a 2.7-bit-per-parameter format that packs three signed weights per byte via a shared 5-bit centroid index (fit with Lloyd-Max/k-means++) plus 3 sign bits, decoding to INT8 using Arm SIMD table-lookup (TBL) instructions with only 5 logical instructions for index construction. They compare against direct casting, GPTQ, and scalar baseline formats under the same QAT procedure, with a custom C++ inference implementation for Arm CPU/Android deployment.

Key results:

• S3D8 compresses Llama 3.2 11B Vision Instruct from 21,340 MB (bfloat16, 16 bpp) to 3,569 MB (2.68 bpp), with average VQA task score dropping from 0.744 to 0.661 (VQAv2 0.702, ChartQA 0.648, DocVQA 0.740, AI2D 0.554), outperforming size-matched INT (0.347), student-t (0.565), and lloyd-max (0.436) formats.
• At ~2.7 bpp, S3D8 with GPTQ alone reaches 0.340 average task performance vs. 0.018 for rate-matched INT with GPTQ; adding QAT raises S3D8 to 0.661.
• End-to-end on a Pixel 8a, the 3.73 GB S3D8 model generates at a median 3.8 tokens/s (12.5 GB/s bandwidth); on Graviton4, S3D8 achieves 36.8 tokens/s (120 GB/s) vs. 26.4 tokens/s (258 GB/s) for INT8.
• Matrix-multiply benchmarks show S3D8 speedups for memory-bound small-batch decoding (e.g., 1031 GMAC/s vs. 461 for INT8 on Graviton4 at m=1) but slight slowdowns for compute-bound larger-batch prefill/vision encoding.

Why it matters / caveats: The work shows sub-3-bit VLM compression with limited quality loss is achievable on-device without training-data access, via a data-free synthetic QAT pipeline and a hardware-co-designed format for Arm CPUs. The authors note experiments are limited to a single model and CPU execution, S3D8's decoder is Arm-specific and may need new kernels on other architectures, and evaluation uses only a small set of VQA-style benchmarks.

Problem: Open-ended culinary reasoning benchmarks typically rely on model judges, small human preference panels, or brittle exact-match keys, which entangle the evaluator with the systems being tested and discard useful differences between plausible answers.

Method: FlavourBench uses Epicure, a versioned culinary runtime that represents 1,790 ingredients in a 300-dimensional space with deterministic operations for substitution, pairing, dietary feasibility, and regional composition, to pre-score all 56 three-ingredient portfolios (from 8 candidates) for each task before any model runs. A task compiler builds 534 shared tasks across three families (Substitution, Pairing, Constraints; 178 tasks each) drawn from two independently compiled 89-task-per-family panels, and 27 frontier endpoints (e.g., GPT-5.6 Sol/Terra/Luna, Claude Opus/Sonnet/Fable 5, Gemini 3.1 Pro/3.6 Flash, Grok 4.6, Llama 4 Maverick, Qwen3.8 Max/A95B, GLM 5.2/5.3, DeepSeek V4 Pro/Flash, Mistral Large 3, Cohere Command A/R+, etc.) are evaluated via fixed provider routes. The FlavourBench Score is the equal-family mean of frozen task scores, with uncertainty from 50,000 anchor-cluster bootstrap replicates and 100,000 sign-flip draws with Holm correction across all 351 pairwise model contrasts.

Key results:

• Grok 4.6 ranks first with a FlavourBench Score of 65.1 (95% CI [61.0, 69.2]), followed closely by Gemini 3.1 Pro (65.0) and GPT-5.6 Sol Pro (64.2); Command R+ ranks last at 47.9.
• Only 101 of 351 pairwise model comparisons remain statistically significant after Holm correction, meaning most adjacent-rank differences are not resolved.
• The two independently compiled task panels correlate at Pearson r = 0.89 (Spearman ρ = 0.80) at the model level.
• The full evaluation covers 14,418 model–task response cells (27 models × 534 tasks), with each model completing exactly 89 valid tasks per panel/family.

Why it matters / caveats: The benchmark demonstrates an executable, judge-free evaluation paradigm (analogous to SWE-bench or BFCL) applied to culinary decision-making, yielding dense partial-credit scoring and rigorous multiplicity-controlled statistics rather than a flat leaderboard. The authors caution the score measures agreement with one specific Epicure release rather than universal human taste, evaluates constrained ingredient selection rather than full recipe generation or sensory/safety reasoning, and that results are tied to specific model routes and collection dates.

Problem: Existing reviews of hadith computational science document literature growth but fail to critically assess which transformer- and LLM-era advances are methodologically robust versus benchmark-bound, and largely omit Islamic scholar and domain-expert perspectives on authenticity, authority, and responsible use.

Method: The authors conduct a critical narrative review combining three elements: critique of six prior review papers, paper-level appraisal of representative original studies coded against six explicit dimensions (corpus realism, setting, transfer, reproducibility, expert input, scholarly use), and synthesis of Islamic scholar/domain-expert commentary. Searches were run iteratively across Google Scholar, Scopus, ACL Anthology, SpringerLink, ScienceDirect, arXiv, and manual domain repositories, yielding 42 shortlisted items refined to 32 records, organized using a three-level pipeline framework (Level 1: isnad-matn segmentation; Level 2: narrator/source verification and QA; Level 3: knowledge graphs and corpus-wide enrichment).

Key results:

• Review database: 42 items at initial shortlist, refined to 32 records after scope refinement.
• Cited resource-scale examples: Sanadset spans 650K narrations across 926 books; Asgari-Bidhendi et al. (2025) process 1.2M narrations with multi-task expert scoring using six domain experts.
• No pooled quantitative benchmark scores are reported — the authors explicitly state they "do not pool headline scores into a single quantitative ranking" because task definitions, corpora, and metrics vary too much across the literature.

Why it matters / caveats: The review argues hadith computation should be treated as an "evidence infrastructure problem" requiring knowledge integration, provenance tracking, and expert supervision rather than judged on isolated model performance; it proposes a five-part research agenda. As a narrative review, it is self-described as vulnerable to search/selection bias and underrepresentation of gray literature and poorly indexed Arabic-only venues, and reports no new experiments or datasets of its own.