AI papers — 2026-05-06

Problem: Single-agent autonomous research systems are prone to plausible-but-unsupported outputs—claims that outrun their evidence, silently misreported results, or hallucinations that go unchecked because the generator and validator share the same model family.

Method: ARIS is an open-source research harness organized into three layers: an execution layer with 65+ Markdown-defined reusable skills, MCP model bridges, and a persistent per-project research wiki; an orchestration layer covering five end-to-end workflows (idea discovery, experiment bridge, auto-review loop, paper writing, and rebuttal); and an assurance layer that runs integrity verification, result-to-claim mapping, and claim auditing against raw evidence, plus a five-pass scientific-editing pipeline and PDF visual inspection. The default configuration pairs a cross-family executor/reviewer duo (e.g., Claude Code as executor, GPT-5.4 as reviewer) to break same-model blind spots. A prototype self-improvement loop proposes harness edits that are gated by reviewer approval before adoption.

Key results:

• System footprint: 65+ skills, 5 workflows, 6 MCP model bridges, tested on 3 executor platforms (Claude Code, Codex CLI, Cursor)
• 4 effort presets (lite/balanced/max/beast) with no skill-level dependencies
• Qualitative deployment reports show the adversarial review loop iterates paper scores from ~4/10 to 7.5/10 across 2–3 revision rounds

Why it matters / caveats: ARIS is the first openly described research harness that treats experimental integrity as a first-class architectural concern rather than a post-hoc check, and its cross-family adversarial design directly addresses correlated errors in same-model self-refinement. The paper itself notes that human-in-the-loop significantly improves output quality, and that autonomous end-to-end research without human oversight remains unreliable for high-stakes findings.

Problem: In multimodal LLM post-training, supervised fine-tuning (SFT) introduces distributional drift—degrading both visual grounding and reasoning—that compounds during subsequent reinforcement learning with verifiable rewards (RLVR), limiting final model performance.

Method: PRISM is a three-stage post-training pipeline (SFT → distribution alignment → RLVR) that inserts an explicit pre-alignment stage between SFT and RL. The alignment stage formulates distribution repair as a minimax adversarial game between the policy and a Mixture-of-Experts (MoE) discriminator containing dedicated perception and reasoning experts, trained with a Bradley-Terry loss on policy rollouts versus a supervision pool—requiring no teacher logits. The SFT stage uses a curated 113K corpus distilled from Gemini 3 Flash (targeting zero-pass-rate problems with dense visual grounding) combined with 1.26M public demonstrations, totaling ~1.37M samples. The resulting aligned policy then serves as initialization for GRPO, DAPO, or GSPO.

Key results:

• PRISM+GRPO outperforms SFT→GRPO by +4.4 average accuracy points on Qwen3-VL-4B across multimodal benchmarks
• PRISM+GRPO outperforms SFT→GRPO by +6.0 average accuracy points on Qwen3-VL-8B
• Gains are consistent across all three tested RL algorithms (GRPO, DAPO, GSPO)

Why it matters / caveats: PRISM demonstrates that the SFT-to-RL gap is a tractable bottleneck addressable with a logit-free, black-box-compatible alignment stage, which matters for production pipelines where teacher models are proprietary. Current experiments are limited to Qwen3-VL at 4B/8B scale; whether gains persist at larger scales or with different base model families remains untested.

Problem: Developing frontier search agents has required resource-intensive pipelines (CPT + SFT + RL) exclusive to well-funded industrial teams, creating a barrier for academic and open-source researchers.

Method: OpenSeeker-v2 is a 30B-parameter search agent trained with a single SFT run on 10.6K high-difficulty trajectories, building on the ReAct paradigm. Three data synthesis modifications drive quality: (1) scaling the knowledge-graph expansion budget K to produce larger evidence subgraphs that force multi-hop reasoning; (2) expanding the tool set A to teach more versatile interaction strategies; and (3) strict low-step filtering that removes any trajectory solvable in too few tool-call steps, enforcing a minimum difficulty floor. The base model is fine-tuned once using standard SFT with no RL or continual pre-training.

Key results:

• 46.0% on BrowseComp (vs. 43.4% for Tongyi DeepResearch with full CPT+SFT+RL)
• 58.1% on BrowseComp-ZH (vs. 46.7% for Tongyi DeepResearch)
• 34.6% on Humanity's Last Exam (vs. 32.9% for Tongyi DeepResearch)
• 78.0% on xbench (vs. 75.0% for Tongyi DeepResearch)

Why it matters / caveats: OpenSeeker-v2 is the first academic, SFT-only model to reach state-of-the-art among 30B-scale ReAct agents, directly challenging the assumption that RL and CPT are necessary for frontier search performance. The model weights are fully open-sourced. The evaluation scope is limited to 4 benchmarks at a fixed 30B/ReAct scale, and gains may not transfer to smaller models or paradigms requiring agentic RL generalization.

Problem: Existing segmentation MLLMs are fragmented: image-only models (e.g., LISA) lack visual prompting and video support, while video models (e.g., VISA, VideoLISA) lack a unified architecture for both modalities and both textual and visual prompts.

Method: X2SAM is a unified segmentation MLLM that couples an LLM backbone with a Mask Decoder and a newly introduced Mask Memory module, which stores guided vision features to maintain temporally consistent mask generation across video frames. The model processes both textual (T-Prompts) and visual (V-Prompts) instructions in a single interface and supports 14 segmentation task types spanning generic, referring, reasoning, grounded conversation generation (GCG), interactive, visual grounded (VGD), object-centric, and out-of-domain segmentation for both images and videos. Training uses a unified joint strategy co-training on heterogeneous image and video datasets. The paper also introduces the Video Visual Grounded (V-VGD) segmentation benchmark for evaluating visual-prompt-driven video instance tracking.

Key results:

• X2SAM supports 7 image segmentation tasks and 7 video segmentation tasks in a single model, more than any prior method
• Achieves strong video segmentation performance with consistent gains on video benchmarks while remaining competitive on image segmentation benchmarks
• Specific numeric benchmark scores not stated in the available text excerpt

Why it matters / caveats: X2SAM provides the most comprehensive unified segmentation interface published to date, enabling conversational visual prompting for videos—a capability absent in all prior image-only or video-only MLLMs. Detailed per-benchmark quantitative comparisons against LISA, VISA, Sa2VA, and X-SAM are present in the paper but not fully captured in the available text; the V-VGD benchmark is newly introduced with no established baselines.

Problem: Agentic orchestration harnesses achieve strong performance through complex system designs, but the core mechanism driving that performance—and whether it can be distilled into a portable, model-internal capability—is not understood.

Method: HeavySkill frames "heavy thinking" as a two-stage inference pipeline: (1) a parallel reasoning phase that generates K independent reasoning trajectories from the same LLM (πθ) without cross-trajectory communication, and (2) a sequential deliberation phase where a second LLM (πϕ, optionally the same model) receives a serialized memory cache of pruned and shuffled trajectories and synthesizes a final answer. This pipeline is encoded as a portable plain-text "readable skill" document—compatible with Claude Code, CodeX, and Hermes harnesses—that instructs the orchestrator to spawn parallel subagents and then deliberate. Reinforcement learning is used to further scale both the width (K) and depth of the deliberation capability.

Key results:

• Heavy-Mean@K (HM@K) consistently outperforms Mean@K, Vote@K, and standard Best-of-N across all tested models on AIME25, BeyondAIME, HMMT25-Feb, and GPQA-Diamond
• Performance hierarchy confirmed: Heavy-Pass@K ≥ Heavy-Mean@K ≥ Vote@K ≥ Mean@K
• Frontier models (Kimi K2 Thinking, GPT-5-Thinking) approach Pass@K upper bounds under heavy thinking on several benchmarks
• HeavySkill outperforms majority voting most significantly on hardest benchmarks (BeyondAIME, HMMT, GPQA-Diamond); parity with voting only on near-saturated benchmarks (AIME25 scores >90)

Why it matters / caveats: HeavySkill offers a unified explanation for why orchestration harnesses work—the key ingredient is parallel exploration plus deliberative synthesis, not complex system scaffolding—and packages this as a transferable skill requiring no model retraining. The main caveat is that the compute cost scales with K parallel generations, and the benefit diminishes on benchmarks where the base model is already near ceiling.

Problem: Video VAEs optimized purely for reconstruction quality do not necessarily produce latent spaces that are well-suited for downstream generative modeling (diffusability), and improving reconstruction does not reliably translate into better video generation.

Method: Predictive Video VAE (PV-VAE) augments standard video VAE training with a predictive reconstruction objective: at each training step, a random number of future frame groups are dropped before encoding, and the decoder is trained to reconstruct the full video (including dropped frames) from the partial-observation latents padded with uninformative prior vectors. A motion-aware auxiliary loss—requiring reconstruction of temporal frame differences—prevents the model from exploiting static-region copy shortcuts. PV-VAE uses 3D causal convolutions with 16× spatial and 4× temporal downsampling (latent channel dimension 64), pretrained on images for 300K steps, then trained on video for 50K steps, followed by a 50K-step decoder fine-tuning stage with frame-dropping disabled to close the training–inference gap. Generation is evaluated using the Latte Transformer-based latent diffusion model.

Key results:

• 34.42 FVD improvement over Wan2.2 VAE on UCF101 class-conditional generation (17-frame, 256×256)
• 52% faster convergence on UCF101 compared to Wan2.2 VAE
• 22.31 FVD improvement over SSVAE on UCF101
• 63.93 FVD improvement over Hunyuan-VAE on UCF101 with 2.68× training speedup and 62% memory reduction
• 38%/19% faster inference vs. Hunyuan-VAE/Wan2.2 VAE; 64%/15% memory reduction at inference

Why it matters / caveats: PV-VAE demonstrates that a predictive learning objective integrated into VAE training—rather than separate world-model pretraining—is sufficient to substantially improve both generation quality and training efficiency, with latent probing tasks (optical flow, point tracking, next-frame prediction) confirming richer temporal structure in the learned representations. The approach is currently validated only with the Latte diffusion architecture at 256×256 resolution; generalization to larger-scale diffusion transformers and higher resolutions is not yet demonstrated.

Problem: Existing patent examination benchmarks treat the process as static classification or extraction, failing to model the iterative, multi-turn dialogue between examiners and applicants (Office Actions and rebuttals).

Method: PatRe is a benchmark of 480 real USPTO patent examination cases covering two generative tasks: Office Action (OA) generation and applicant rebuttal generation. OA generation is evaluated under three settings — zero-shot direct prompting (OA-DP), oracle reference provision (OA-RO), and BM25 retrieval-simulated (OA-RS). Evaluation uses a hierarchical metric suite combining deterministic metrics (Decision Accuracy, Statute Precision, ROUGE-L) with LLM-as-a-Judge scoring via Gemini-3.1-Flash-Lite across five dimensions (soundness, clarity, constructiveness, completeness, language style). A range of proprietary and open-source LLMs are benchmarked under all settings.

Key results:

• Dataset spans 480 patents across all 8 IPC sections (A–H), with up to 15 examination rounds per patent
• Proprietary models consistently outperform open-source models on both OA and rebuttal tasks
• A task asymmetry is observed: models perform better on examiner-side OA generation (proactive) than applicant-side rebuttal generation (reactive)
• Specific numeric accuracy scores not stated in the extracted text

Why it matters / caveats: PatRe is the first benchmark to model the full patent examination lifecycle as a dynamic, multi-turn generative process, enabling more realistic LLM evaluation for legal-technical reasoning. The benchmark is limited to USPTO cases, and evaluation relies partly on an LLM judge, which may introduce its own biases.

Problem: Standard Gaussian Splatting primitives (3DGS, 2DGS) assign a single view-dependent color and a scalar opacity per primitive, requiring a large number of Gaussians to represent scenes with complex textures, leading to redundant and non-compact representations.

Method: SVGS (Spatially Varying Gaussian Splatting) augments 2D Gaussian surfels (2DGS) by making color and opacity functions of the 2D intersection point on each surfel. Three spatially varying functions are implemented and compared: bilinear interpolation across four quadrant-assigned learnable colors/opacities; movable kernels (four learnable exponential kernels whose centers are optimized); and a tiny three-layer MLP per surfel that maps local coordinates (u, v) to color and opacity. The movable kernels design uses 1.40× the parameters of baseline 2DGS per Gaussian.

Key results:

• SVGS with movable kernels (Ours-MK) achieves PSNR of 34.10 on Synthetic Blender (no Gaussian limit) vs. 33.65 for 2DGS baseline
• At 50K Gaussian limit on Blender, Ours-MK reaches 33.09 PSNR vs. 32.64 for 2DGS
• On DTU (no limit), Ours-MK reaches PSNR 37.76 / SSIM 0.957 / LPIPS 0.117 vs. 2DGS baseline 36.35 / 0.947 / 0.165
• All three spatially varying functions outperform the 2DGS baseline on every dataset tested

Why it matters / caveats: SVGS demonstrates that per-primitive spatial variation meaningfully improves texture reconstruction compactness without changing the explicit Gaussian representation paradigm. The MLP variant carries significantly more parameters (1.88× 2DGS), and the current implementation is tied to 2DGS surfels rather than 3DGS ellipsoids.

Problem: Existing agent benchmarks evaluate tasks on pre-specified or synthesized files with limited real-world inter-file dependencies, leaving agents' ability to reason across large, heterogeneous workplace file collections untested.

Method: Workspace-Bench constructs realistic multi-workspace environments with 5 worker role profiles (Operations Manager, Logistics Manager, Product Manager, Researcher, Backend Manager), 74 file types, 20,476 files (up to 20 GB), and 388 tasks, each annotated with a file dependency graph and evaluated over 7,399 rubrics. Four agent harnesses — ClaudeCode, DeepAgent, Hermes, and OpenClaw — are evaluated with seven backbone LLMs including Claude Opus 4.7, Gemini 2.0 Flash, DeepSeek, and Kimi. A smaller Workspace-Bench-Lite (100 tasks) is also released, reducing evaluation cost by ~70%.

Key results:

• Best agent configuration (Opus 4.7 + ClaudeCode/DeepAgent/Hermes) reaches 68.7% rubric pass rate
• Average pass rate across all agent-LLM combinations is only 47.4%
• Human + tools baseline achieves 80.7%, a gap of ~12 points over the best AI agent
• Top three configurations all use Opus 4.7; performance varies substantially across both harnesses and LLMs at lower tiers

Why it matters / caveats: Workspace-Bench is the first benchmark to evaluate agents on multi-workspace structures with explicit file lineage relations and over 70 file modalities, exposing a large gap between current AI agents and human workers on real workplace tasks. The benchmark is derived from ByteDance's internal Lark platform, which may limit generalizability to other enterprise environments.

Problem: LLM performance on medical diagnosis has only been validated on curated clinical vignettes, not on naturalistic, layperson-reported symptom conversations, where information is incomplete and non-expert.

Method: SymptomAI is a set of conversational AI agents built on Gemini 2.0 Flash, deployed via the Fitbit app in a 9-month randomized study (June 2025 – April 2026) with N=13,917 participants across five prompting-strategy arms: (1) base user-guided; (2) fixed canonical history-taking questions; (3) flexible canonical questions; (4) dynamic agent with live differential diagnosis (DDx) updates; (5) dynamic agent with final DDx only. Accuracy was assessed by three board-certified Family Medicine physicians who blindly ranked DDx from SymptomAI vs. two independent clinicians on a subset of 517 cases. An LLM-based auto-rater was then used to scale evaluation to the full N=1,228 self-diagnosed cohort.

Key results:

• Arms 2–5 (structured interview strategies) achieved 27.34% higher DDx top-5 accuracy than the base user-guided arm (p < 0.001)
• SymptomAI top-5 DDx accuracy: 73% vs. 60% for independent clinicians (Median OR = 2.47, p < 0.001)
• Clinical raters ranked SymptomAI DDx as best in >50% of cases (OR = 2.20, p < 0.001)
• Dynamic arms (4 & 5, combined 71.4%) performed comparably to canonical-question arms (2 & 3, combined 75.6%, p = 0.155)
• Wearable physiological signals (Fitbit) identified OR > 7 for influenza association

Why it matters / caveats: This is one of the largest real-world deployments of a conversational diagnostic AI agent using objective clinical accuracy measures, demonstrating that structured AI-led interviews can outperform clinicians given the same transcript context. Limitations include self-reported ground-truth diagnoses and a study population drawn from Fitbit users, who may not represent the general population.

Problem: As LLM agents form coordinated multi-agent teams, existing RL frameworks designed for single-agent trajectories lack a shared abstraction for the reward design, credit assignment, and orchestration learning challenges specific to agent coordination.

Method: This is a survey and taxonomy paper. It introduces the orchestration trace — a temporal interaction graph recording sub-agent spawning, delegation, communication, tool use, return, aggregation, and stopping — as the central unit of analysis. Using this abstraction, the paper organizes the literature across three axes: eight families of reward design (R1–R8), eight credit/signal-bearing units from token to team, and five orchestration sub-decisions (O1: when to spawn, O2: whom to delegate, O3: how to communicate, O4: how to aggregate, O5: when to stop). Industrial systems analyzed include Kimi K2.5/K2.6 (PARL training, up to 300 sub-agents, 4,000 coordinated steps), OpenAI Codex, and Anthropic Claude Code.

Key results:

• Kimi K2.5 scales to 100 sub-agents / 1,500 coordinated steps; K2.6 expands to 300 sub-agents / 4,000 steps
• Explicit counterfactual message-level credit assignment is the sparsest area in the literature
• No RL training method for the stopping decision (O5) found in the curated 84-paper pool as of May 2026
• 15 open research problems enumerated across algorithms, rewards, systems, safety, and evaluation

Why it matters / caveats: The orchestration trace framing provides a unifying vocabulary for a fragmented literature and exposes specific under-addressed gaps (stopping decisions, message-level credit). Industrial claims (Kimi, Codex, Claude Code) are drawn from public deployment reports rather than independently reproduced training traces, limiting evidence level.

Problem: Multi-modal point cloud completion methods that project sparse 3D points onto image planes suffer from Cross-Modal Entropy Collapse: the projected support is near-zero in measure, severing gradient flow from visual features to geometry and causing models to degenerate into unimodal template retrievers.

Method: SplAttN replaces hard projection with Differentiable Gaussian Soft Splatting to produce a continuous, dense image-plane density field. The GS-Bridge module uses geometric tokens (from EdgeConv + Graph Transformer hybrid tokenizer) as queries over visual features encoded by TinyViT, with soft splatting weights combining a Gaussian spatial kernel and a differentiable depth (inverse-z) prior. A Global-Local Decoder then hierarchically upsamples a coarse point skeleton P0 → P1 → P2 using Structure Self-Attention and Cross-Attention over local EdgeConv features. Training uses a Weighted Arc-Chamfer Distance (arccosh of CD) to compress outlier sensitivity. Counterfactual evaluation via a Semantic Consistency Score on KITTI tests whether models genuinely depend on visual cues.

Key results:

• PCN benchmark: CD-Avg 6.36 × 10⁻³, F1 0.854 — best among all compared methods (vs. prior SOTA GeoFormer: CD 6.42, F1 0.853)
• State-of-the-art on ShapeNet-55 and ShapeNet-34 (specific numbers not stated in extracted text)
• Counterfactual evaluation shows baselines (SVDFormer, GeoFormer) are insensitive to visual removal, while SplAttN maintains strong visual dependency on KITTI

Why it matters / caveats: SplAttN provides both a theoretical diagnosis (Cross-Modal Entropy Collapse via measure-theoretic analysis) and a practical fix (soft splatting), backed by a novel counterfactual test for genuine cross-modal dependency. Improvements over prior SOTA on PCN are incremental in CD terms; the method adds complexity through per-point Gaussian rendering and a TinyViT visual encoder.

Problem: Evaluating interactive world models is hindered by the lack of a large-scale, unified benchmark covering diverse scenes, perspectives, and action modalities, while existing benchmarks address only general video generation and lack interaction-specific or memory tasks.

Method: iWorld-Bench processes 12 high-quality source datasets (KITTI-360, Waymo, nuScenes, RealEstate10K, DL3DV-10K, etc.) into a standardized dataset of 330K video clips spanning 4 agent perspectives (UGV, UAV, human, robotics), 9 outdoor weather conditions, and 5 indoor lighting conditions. An Action Generation Framework defines an 81-action dictionary with modality-agnostic encoding to unify text, one-hot, and camera-parameter control inputs across heterogeneous world models. Six task types with varying difficulty levels—including memory tasks—yield 4,900 evaluation tasks assessed along 9 metrics covering visual quality, action following, and memory.

Key results:

• 14 interactive world models evaluated, including 5 text-controlled, 2 one-hot encoding-based, and 7 camera-parameter-controlled models
• Benchmark contains 4,900 evaluation tasks selected from 2,100 curated video clips
• iWorld-Bench is the only benchmark among 8 compared that supports multiple inputs, interactive tasks, camera control, memory ability, multi-perspective, and all-weather evaluation simultaneously

Why it matters / caveats: iWorld-Bench provides the first standardized framework for fairly comparing interactive world models across heterogeneous action modalities, which is a prerequisite for training agents in simulated environments for robotics and autonomous driving. The benchmark currently relies on synthesized camera trajectories, and sim-to-real transfer of learned interaction behaviors is not evaluated.

Problem: Indic ASR systems fail dramatically on entity-dense audio (phone numbers, currency, addresses, brand names, code-mixed speech), with open-source SOTA achieving only 2.7% Entity-Hit-Rate on Telugu—far below what production IVR and fintech deployments require.

Method: The TTS-STT Flywheel generates ~22,000 entity-dense Indic-English code-mix utterances using a multi-system TTS pipeline (Praxy R6, Chatterbox, IndicF5, ElevenLabs v3, Cartesia sonic-3) seeded with entity dictionaries and Anthropic Haiku-4.5 for text generation; the resulting Entity-Dense Synthetic Audio (EDSA) corpus is used to LoRA fine-tune vasista22/whisper-telugu-large-v2 (and Hindi/Tamil counterparts built on Whisper-large-v3). Evaluation uses Entity-Hit-Rate (EHR), a semantic normalization metric handling currency parsing, brand aliasing, spelled-digit matching, and NFKC pincode normalization. An EDSA-isolation ablation (LoRA on FLEURS-Te only) confirms virtually all gains come from the EDSA corpus rather than the LoRA recipe itself.

Key results:

• β-Te EHR: 0.473 vs. open SOTA 0.027 (17×) and commercial Deepgram Nova-3 0.16 (3×)
• β-Hi EHR: 0.337 (7× vs. open SOTA); β-Ta EHR: 0.543 (22× vs. open SOTA and vs. Deepgram)
• Native-human-recorded sanity check (n=20 Telugu): EHR 0.516, confirming transfer to real speech
• Read-prose regression bounded to +6.6 pp WER on FLEURS-Te
• Total synthesis cost: <$50

Why it matters / caveats: The flywheel demonstrates a cost-effective path to closing entity-dense ASR gaps for low-resource languages without requiring native-recorded training data. All three models fall below pre-registered targets (0.75/0.65 EHR), the recipe is contraindicated on Hindi/Tamil for the Script Collapse fix, and commercial systems outperform on Hindi where they already have substantial entity coverage.

Problem: Conversational Aspect-based Sentiment Quadruple Analysis (DiaASQ) suffers from structural noise and positional encoding failures: standard GCNs propagate information across semantically unrelated conversation threads, and standard RoPE causes Distance Dilution where logically adjacent utterances separated by many tokens lose positional correlation.

Method: TCDA combines two components over a pre-trained language model backbone. Thread-Constrained Directed Acyclic Graph (TC-DAG) replaces general GCNs by enforcing strict intra-thread boundaries—only propagating information within each semantic thread while maintaining global connectivity via a shared root node—and incorporates temporal speaker order. Discourse-Aware Rotary Position Embedding (D-RoPE) replaces standard RoPE with a dual-stream projection that maps tokens and utterances to independent subspaces and applies topology-adaptive coordinate transformation using tree-like distances derived from the thread structure, coupling local syntactic cues with utterance-level progression before attention computation.

Key results:

• State-of-the-art performance on two DiaASQ benchmark datasets
• Specific F1 figures not captured in the extracted text

Why it matters / caveats: TCDA addresses a structural gap in conversational sentiment analysis by making thread-level topology a first-class citizen in both graph propagation and positional encoding. The evaluation is limited to two datasets, and the approach's scalability to very long multi-party conversations with many threads is not discussed.

Problem: Existing UAV SAR research relies on decoupled classical perception and path planning without semantic reasoning, and current embodied aerial benchmarks depend on fine-grained step-by-step language instructions rather than high-level goal-oriented tasks, leaving no unified benchmark for evaluating autonomous MLLM-driven agents in realistic wilderness rescue scenarios.

Method: ESARBench formalizes the Embodied Search and Rescue (ESAR) task, requiring agents to autonomously explore, discover multi-modal clues, and reason about victim locations across four sequential mission phases. Four large-scale environments (Alpine, Desert, Snowy Peak, Coast) are constructed in Unreal Engine 5 using real-world GIS satellite imagery and DEM data, with AirSim providing flight dynamics simulation. An "Event-Snapshot-Task" generation framework produces 600 tasks derived from real SAR cases with dynamic weather (13 conditions), time-of-day variation, and stochastic clue placement; performance is measured via SR (success rate), TSR (target search rate), CDS (clue discovery score), and RS (rescue score).

Key results:

• 600 tasks across 4 environments and 4 difficulty levels
• Map scales ranging from 2×2 km (Alpine) to 5×5 km (Desert)
• Direct transfer of ground-based MLLM agents (ObjectNav) to aerial ESAR is insufficient; aerial-adapted agents outperform but all baselines reveal bottlenecks in spatial memory and safety-efficiency trade-offs
• Specific SR numbers not stated in the extracted text

Why it matters / caveats: ESARBench is the first photorealistic, GIS-grounded benchmark for embodied UAV search and rescue, enabling evaluation of holistic agent capabilities rather than isolated sub-tasks. The benchmark is simulation-only (UE5/AirSim), and sim-to-real transfer of trained policies to physical drones has not been validated.

Problem: Existing medical AI environments cover only fragments of clinical reasoning—lacking broad domain coverage, authentic tool ecosystems, or RL-compatible training—and multi-turn agentic RL training collapses into verbose single-turn monologues due to sparse terminal rewards misaligned with sequential clinical trajectories.

Method: Healthcare AI GYM is a Gymnasium-compatible environment with 10 clinical domains, 3,600+ tasks, 135 domain-specific clinical tools (lab ordering, severity scoring, etc.), and an 828K-passage knowledge base, evaluated by a 5D safety-aware reward function. The proposed Turn-level Truncated On-Policy Distillation (TT-OPD) trains Qwen3.5-9B using GRPO with a gradient-free EMA teacher (decay α=0.995, updated every 5 steps with periodic hard-copy fallback every 30 steps) that receives outcome-privileged correctness hints; the teacher's KL divergence against the student provides dense turn-level regularization. A cosine length-controlled reward (scaled by remaining budget toward Lmax) suppresses response length explosion without truncating necessary reasoning.

Key results:

• TT-OPD achieves best performance on 10 of 18 benchmarks with average +3.9 pp over non-RL baseline
• MedQA (USMLE): 87.1% (+16.4 pp over base); MedMCQA: 66.2%; MMLU-Med: 65.5%
• Vanilla GRPO achieves +9.4 pp training accuracy but suffers instability; TT-OPD sustains 7.0–7.4 tool-use turns with controlled response lengths of 5.7–9.3K tokens

Why it matters / caveats: TT-OPD identifies and addresses the "agentic collapse" failure mode in multi-turn RL—a broadly relevant finding for any sequential decision-making agent trained with sparse rewards. A key limitation is the agentic-to-textual transfer gap: RL-trained agents do not improve on static text-based QA benchmarks due to format-reward dilution, and all experiments use a single backbone (Qwen3.5-9B) without SFT warmup.

Problem: Neural lossless compressors either require externally stored pre-trained weights (hundreds of MB, GPU inference) making archives non-self-contained, or rely on prohibitively slow per-token online training with Transformers, preventing practical fully-online neural compression.

Method: StateSMix trains a 2-layer Mamba-style State Space Model (DM=32, NL=2, ~120K active parameters) entirely from scratch online during compression, updated via Adam after every 32-token chunk with no GPU or pre-trained weights. Sparse n-gram tables (bigram through 32-gram, 2^24=16M slots each) are blended with the SSM output via a softmax-invariant logit-bias: n-gram evidence adds directly to SSM logits, touching only non-zero-count tokens and scaling bias magnitude by the SSM's predictive entropy (entropy-adaptive mixing). A compact vocabulary remapping reduces the effective vocabulary from 49,152 to 18K–44K tokens, and linear-probing collision resolution with probe depth 8 is used for n-gram hash tables. Range arithmetic coding encodes the final probability estimates.

Key results:

• enwik8 (10 MB): 2.161 bpb, beating xz -9e by 0.7%
• enwik8 (100 MB): 2.130 bpb
• SSM alone (no n-gram): beats xz on enwik83M (840 KB vs. 852 KB); 46.6% size reduction over frequency-count baseline
• N-gram only (no SSM): 16.1% size reduction
• Throughput: ~2,000 tok/s on 4 CPU cores with AVX2 SIMD / OpenMP; no GPU required

Why it matters / caveats: StateSMix demonstrates that a tiny online-trained Mamba SSM can match or beat a best-effort LZMA compressor on English text without any pre-trained weights, establishing a new practical point in the online neural compression tradeoff space. The system's advantage over xz disappears around 30 MB files, where LZMA's long-distance repetition exploitation (beyond the fixed-horizon n-gram tables) takes over, and results are reported only on English text (enwik8).