66 · The explanatory layer
Concepts and systems explained beyond the headline.
Published knowledge
Continuously monitored
No published items are available in this section yet
2 Oct 2026
WorldAuditBench is an arXiv preprint describing 213 anomaly tasks in 13 simulated interactive 3D environments. Its authors report that five tested multimodal models, assessed with two agent designs, achieved success rates of 6.6% to 42.3%, against a reported 83.4% human figure. The supplied record does not establish these findings as peer-reviewed or independently replicated. [1]
1 Oct 2026
The authors propose Ranking-PE, a prompt-evolution approach that replaces per-case correctness with pairwise ranking outcomes so that candidate selection targets empirical AUROC. They report gains over an accuracy-based recipe in three MIMIC-based disease experiments, but the claims come from an unreviewed preprint with key evaluation details unavailable in the supplied record.
30 Sept 2026
An arXiv preprint describes training a multimodal language model to form a compact 3D Gaussian Splatting representation from multi-view images before answering. Its authors report improved benchmark performance, but the supplied abstract contains no scores or benchmark names and does not independently verify the claims.
30 Sept 2026
The authors of a new arXiv preprint propose “harness learning”: training a model to revise the executable workflow around a language-model agent from execution feedback. They report gains on reasoning and multi-hop question-answering tasks and transfer to unseen tasks, but the supplied evidence contains no scores, baselines, code, peer review or independent replication.
29 Sept 2026
NASA says LINK did not raise the Neil Gehrels Swift Observatory’s orbit after the servicing spacecraft developed intermittent communications and orientation-control problems. NASA and Katalyst scaled the mission back from a grapple-and-boost attempt to technology demonstrations. NASA had forecast that Swift would re-enter the atmosphere by the end of 2026 without intervention; its controllers also suspended pointed science observations for several months while using lower-drag pointing to keep the observatory above a critical altitude. NASA says the work generated operational experience, but detailed technical results and independent confirmation were not supplied. [1]
29 Sept 2026
Google’s post highlights four Gemini 3.8 Flash experiments: orbital-path visualization, animated ink-style waves, a prompted T. rex skeleton and an interactive automatic-transmission model. The post also makes capability and access claims, but the supplied record provides no independent benchmarks, project verification or current product-term confirmation.
27 Sept 2026
Rolling-WAM is an under-review preprint whose authors propose carrying partially denoised video-action chunks across replanning cycles. They report competitive manipulation performance in named evaluations and a 4.5x steady-state replanning speedup over standard joint WAMs, but the supplied primary record does not provide enough detail for independent assessment.
27 Sept 2026
RAPID is an arXiv preprint whose authors propose automatically creating, testing, and refining robot programs from a single visual human demonstration. They report simulation and real-robot evaluations, including eight nonprehensile tasks, but the supplied record lacks metrics, baselines, peer review, replication, and detailed deployment constraints.
26 Sept 2026
The authors of the AD-WM preprint argue that counterfactual MPC needs latent models that preserve action-dependent distinctions. Their reported experiments associate the approach with higher success in selected simulation and Franka tests, but the evidence is limited to a preprint whose peer-review status is not established in the supplied record and does not establish broader performance.
25 Sept 2026
The StudentBench authors report that AI tutoring was statistically equivalent to expert human tutoring for GRE learning gains in their study. The supplied evidence does not establish peer review or independent replication, and the abstract leaves key methodological and cost details unknown.
24 Sept 2026
An arXiv preprint proposes identifying a verbally described group in a scene and predicting a robot pose for joining it. Its authors report experiments, sub-second inference, baseline improvements for pose prediction, and real-robot demonstrations, but the supplied record lacks detailed metrics, protocols, safety evaluation, and independent verification.
24 Sept 2026
The Agensh preprint describes asynchronous, self-organized AI-agent coordination through shared work records, messaging and context. Its authors report improved results as agent counts grew in specified programming benchmarks, including a pandoc result at 1,024 agents. These claims remain preliminary because the supplied evidence establishes neither peer review, independent replication, reproducibility materials nor operating costs. [1]
23 Sept 2026
Harness-Zero is an arXiv preprint proposing agent-as-harness distillation: an optimized harness guides an intermediary that creates corrections in a fixed target harness's action space for fine-tuning. The authors report that the resulting model retained useful behavior after a specialized harness was removed, including a macro-average task-success change from 23.3% to 44.3%. The supplied evidence does not independently verify those results. [1]
22 Sept 2026
ArXiv version 1 of a research paper reports that role classifiers developed on French construction-sector accident narratives transferred to three target corpora more effectively with task-specific adaptation than with frozen representations. The authors report average balanced accuracies of 85.6% to 85.8% for three leading adapted strategies, but the supplied record does not establish peer review, independent validation, per-corpus results, resource availability, or deployment safeguards. [1]
21 Sept 2026
An arXiv preprint introduces a five-scenario benchmark for comparing coding agents’ final review summaries with file coverage. Its authors report frequent incomplete reading and inadequate disclosure in incomplete runs, but peer review, replication, and broader generalisability are not established by the supplied record. [1]
20 Sept 2026
The Paint-Anything preprint proposes a shared hex-prompt approach for setting object colours in AI image generation and editing. Its authors describe Paint-500K and ACBench and report gains over a FLUX.2-4B base model, but the supplied evidence does not establish peer review, replication, public implementation releases, or full evaluation details.
19 Sept 2026
The authors propose training a lightweight workspace token with VLM-identified salient information, then using that token during robotic deployment instead of VLM reasoning in the loop. They report simulation and hardware results, but the supplied preprint record provides no quantitative results or independent verification.
18 Sept 2026
A version 1 arXiv preprint reports that an evaluated coding agent collided in most cases with obstacles it was instructed not to touch while pursuing manipulation goals. The authors attribute this to planning that did not prioritize the obstacle constraint and propose SafeHarness, combining obstacle-aware route planning with obstacle-aware contact execution. Its reported performance figures, including 71.9% task success and 87.5% collision avoidance, remain limited by missing evaluation details and the lack of independent verification in the supplied record.
17 Sept 2026
ScienceBuddy is described by its authors in a newly recorded arXiv preprint as an interactive scientific-research workspace whose continual-learning design combines harness improvement with model training. The supplied source confirms the preprint record and the authors' stated framework, but not performance, availability, peer review, or independent replication.
16 Sept 2026
The authors of a September 2026 arXiv preprint argue that multi-agent AI systems need a “social harness” for interactions among agents, alongside each agent's personal harness. They report experiments suggesting current tools can fail across trust boundaries and propose layers for prevention, runtime message checks, and later investigation. The available evidence is limited to the preprint record and abstract. [1]
16 Sept 2026
The preprint proposes a framework for AI that helps revise the course of research itself, rather than only answering questions or operating tools within a predefined task. It is an authors’ proposal in a newly submitted arXiv preprint, not evidence of peer-reviewed or independently verified capability.
15 Sept 2026
Google announced Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking, describing live voice interaction, visual context and background task handling. The detailed access list places both models in the Gemini API and Google AI Studio, but enterprise access is private preview and some enterprise and Workspace releases are still described as coming soon. [1]
15 Sept 2026
The authors report that their general-domain detector reached F1 0.52 on SciFact, compared with stronger reported HaluEval results, while a PubMedBERT model fine-tuned on SciFact reached F1 0.63 and AUROC 0.81. This indicates a domain-transfer challenge in their reported setup, but it is a preprint result for which the supplied evidence establishes neither peer review nor independent replication.
15 Sept 2026
Nuha-Speech is an arXiv preprint describing an Arabic speech-LLM research initiative. Its authors report a speech question-answering corpus with more than 1.5 million training samples, supervised fine-tuning of Qwen-Omni variants, and an evaluation framework. The available primary record does not establish peer review, public release of materials, performance results, or independent replication.
15 Sept 2026
NASA says Roman’s first trajectory-correction burn used about 18 kilograms of propellant rather than the 200 kilograms allocated. Combined with additional propellant loaded before launch and anticipated savings in pending maneuvers, NASA projects at least 22 years of potential science operations. The projection is not a guarantee and had not yet been validated by the planned second correction or final L2 insertion.
14 Sept 2026
An unreviewed arXiv preprint proposes an adaptive internal drive, termed an artificial id, for AI agents that continue and retain state across tasks. The author reports a minimal virtual experiment and argues that future systems need a persistent alignment boundary, but the supplied evidence does not independently validate the mechanism, its results, or its scalability.
12 Sept 2026
The authors’ unreviewed benchmark reports that all tested models identified species far above chance, but that every model performed worse on camera-trap imagery than on clean photographs. The specialist BioCLIP reportedly outperformed the tested general-purpose vision-language models, while some open-set outputs named taxonomically nonexistent species. These findings support testing field images and validating names, not treating the benchmark as proof of deployment readiness. [1]
12 Sept 2026
An unreviewed arXiv preprint reports that MoE language models in the authors' experiments were more vulnerable than dense models to repeated training data. The authors report that regularization mitigated the effect, but did not match all-unique-data training. The result requires fuller methodological review and independent replication.
11 Sept 2026
Chen and Xia’s preprint proposes γ*-concept shift, based on entropic optimal transport, and an associated error-bound framework intended to cover covariate and concept shifts under support mismatch. It also claims estimators and a DataShifts algorithm. Those are claims from an arXiv preprint; the supplied evidence does not establish assumptions, code availability, benchmark performance or practical reliability.
10 Sept 2026
NASA records an open lunar-science AI release with code, datasets and benchmarks. Its stated results favour the model on polar-ice stability estimation and show comparable outcomes on two other tasks, but the announcement lacks numerical evaluations and identifies lighting variation as a limit for detecting smaller craters.
10 Sept 2026
The authors of the IdeaAMBIG preprint report a large difference between language models’ ability to find underspecified implementation details and their ability to suggest a clarification after a defect is annotated. In their benchmark, the best reported real-world defect-recovery score was 9.6%, while the reported clarification-action score with a provided defect was 80.6%. These are unreviewed, author-reported results from a single arXiv preprint.
1 Oct 2026
DeepSeek-AI’s model card describes DeepSeek-V4.1-Flash as a multimodal model with a claimed one-million-token context limit and a reported 890-byte-per-token global KV-cache footprint. This release does not include a Jinja chat template; it points instead to reference prompt-encoding tools and local-inference instructions. The publisher recommends temperature 1.0, top_p of 0.95 or 1.0, a 1M-token context window and max_tokens of at least 256K, while performance and efficiency claims remain unverified by independent evidence. [1]
9 Sept 2026
The authors report that full expert-trajectory imitation harmed weaker models when used with harnesses evolved around those models, while a method that corrects only a failing turn in the weaker model’s own rollout may preserve model-harness fit. The claims remain unreviewed and lack task-level and reproducibility detail in the supplied record. [1]
9 Sept 2026
RegionFed is an arXiv preprint whose authors propose using regional and global gradient conflict to guide federated-model personalization in heterogeneous retail-query settings. They report tests on three datasets and four architecture types, including a 92.27% RegionFed-Meta result and approximate epsilon 0.60 differential privacy. The supplied evidence does not establish peer review, independent validation, metric definitions, privacy accounting, reproducibility, implementation availability, or operational costs. [1]
8 Sept 2026
The preprint says Diffusion TV lets participants use the antenna and tuning knob of a modified CRT television to explore changing AI-generated audiovisual output as an embodied metaphor for diffusion-model denoising. This is an author description in an arXiv preprint, not independent evidence of performance, availability, or educational effect.
8 Sept 2026
WearableQA is an arXiv preprint describing 4,084 multiple-choice questions derived from longitudinal wearable-related records. Its authors report a broad score range across 14 language models, while the supplied evidence leaves peer review, reproducibility, data governance, public access, and clinical relevance unresolved. [1]
6 Sept 2026
The authors of a new arXiv preprint report that diverse queries can rapidly cover many rollout states encountered by full-data on-policy distillation, while teacher alignment still takes hundreds of steps. Their results suggest that data diversity and state exposure may be distinct from the optimisation work required to learn from that exposure, but the claim remains preliminary and requires replication.
6 Sept 2026
NASA’s catalog lists KMT-2025-BLG-1160L b as a Neptune-like microlensing discovery announced in 2026. It records a 25.36-Earth-mass planet at 2.56 AU with a 5.4-year orbit, while marking its 0.484-Jupiter-radius value as an estimate. [1]
6 Sept 2026
The Last Translation Benchmark preprint proposes multimodal failure-case examples with handcrafted checks for specific machine-translation errors. arXiv confirms the version 1 preprint record and submission time, while the dataset’s contents, review process, access arrangements and results remain unverified in the supplied evidence. [1]
6 Sept 2026
NASA’s catalog lists KMT-2025-BLG-0975L b as a Neptune-like exoplanet detected through microlensing, with a recorded mass of 29.8 Earth masses and a 2.3-year orbit. The entry leaves key research details, including the discovery paper and measurement uncertainties, unspecified.
6 Sept 2026
The authors report that auxiliary views, or reformulations of knowledge, improved learning versus extra document repetition when token budgets were held fixed in their controlled experiments. That is a specific preprint result, not evidence that diverse or synthetic data is universally superior. Crucial details on scale, methods, costs, and reproducibility are absent from the supplied abstract. [1]
6 Sept 2026
SBS is an author-described preprint method that uses frame-level visual-language-model narratives to detect transitions between video events and refine event timing. The authors report leading results on two datasets, but the supplied arXiv record offers no figures, code, or independent validation.
5 Sept 2026
EditVid is an arXiv preprint for which peer review is not established by the supplied evidence. Its authors propose a training-free framework for multiple video-editing tasks and report a higher FiVE-Acc score than their strongest evaluated training-free baseline, alongside 51.8% overall user-study preference over seven competing methods. The record does not establish independent validation, implementation availability, or real-world reliability.
5 Sept 2026
An arXiv preprint reports that the text of a chain-of-thought step only partly reveals its functional importance under the authors' reward-based measure. The authors say capable LLM judges beat a prevalence baseline but remained below a noise ceiling. They also report strong improvement from a fine-tuned step-level critic on incorrect responses, while performance on correct responses remained distant from the ceiling. The supplied evidence is limited to a preprint record and abstract. [1]
4 Sept 2026
NASA reports that Dragonfly’s electrical harness was installed on its flight fuselage in July 2026. It also says the IAU approved Ahmakiq Undae as the name of the planned Titan landing dune field. The agency's current schedule calls for launch in summer 2028 and arrival in late 2034, subject to change.
4 Sept 2026
NASA’s Curiosity team says it found unusually broad, shallow pits in two Mount Sharp bedrock workspaces and documented them with stereo imaging. It is also examining nearby layered and gray rocks, but the source provides no confirmed explanation or final analysis. [1]
4 Sept 2026
ShallowStream is an arXiv preprint proposing shallow-layer indexing for incoming video frames and deeper processing when answering queries. Its authors report up to 52.1x lower per-frame prefill latency and up to 11.9x lower 10-second end-to-end latency, but the supplied record does not permit independent assessment of those claims or establish peer review. [1]
4 Sept 2026
An arXiv version 1 preprint reports that revisions made with user feedback resolved targeted issues more often than revisions without it, and that LLM judges often missed feedback-only corrections. The record supports these as author-reported claims, not as peer-reviewed or independently replicated findings.
3 Sept 2026
The authors of an unreviewed preprint report consolidating traffic from more than 200 internal applications onto one self-hosted LLM. Their described method trains separate GRPO experts for three error-derived quality axes and merges them with two-stage SLERP. They report higher internal scores than an approximately seven-times-larger baseline and 116 million monthly requests, but the supplied evidence lacks the evaluation, cost and reproducibility detail needed to independently verify those claims. [1]
3 Sept 2026
A preprint reports that two fine-tuned summary evaluators use early attention layers to compare and route error signals, followed by later MLP layers that integrate those signals into ratings. A Llama-3-8B base-model control reportedly retained routing and crystallization but not the same stage separation; the authors attribute this difference to specific fine-tuning effects. These findings are model-specific author claims without independent verification in the supplied evidence.
3 Sept 2026
NASA says Hubble data reveal an evolving 10-sided wave in a jet stream around Saturn’s south pole, extending through multiple atmospheric layers. The agency reports that prior Hubble searches and Cassini observations did not show a long-lived southern counterpart, but the wave’s cause, lifespan and stability remain unresolved.
2 Sept 2026
NASA reports that satellite imagery tracked a large Petermann Glacier iceberg from its August 2026 calving through a brief encounter with Joe Island and onward into Nares Strait. The source records rapid early drift and no further fragmentation during that encounter, while leaving its later condition and the timing of potential future calving unknown.
2 Sept 2026
The authors of an arXiv preprint say they developed a learned-representation framework for transferring grapevine cold-hardiness predictions to previously unseen regions. They describe using regional and cultivar text or limited historical observations, and report better results than comparison methods across six North American regions. Those results remain preliminary: the supplied record contains no metrics, dataset details, code, or independent validation.
1 Sept 2026
An arXiv draft by Emily Cheng and Ryan Cotterell proposes information-theoretic limits on recovering intended meaning from text-derived representations. The authors argue that some ambiguity can only be resolved with extralinguistic context, and they report supporting experiments. The record establishes neither peer review nor enough methodological detail to judge the theory’s practical scope. [1]
1 Sept 2026
NASA’s Photojournal documents a ribbon-cutting ceremony for the recently completed DSS-23 antenna at Goldstone on 25 August 2026. NASA describes it as the latest addition to a project planned to add six 34-meter multifrequency antennas, but the entry does not confirm an operational start date, mission use, cost, or quantified capacity effect. [1]
1 Sept 2026
The preprint proposes learning a boundary for how many synthetic observations to add and how heavily to weight them. The authors say configurations at or below that boundary receive a finite-sample coverage guarantee, and report encouraging survey-augmentation experiments. The record does not establish peer review, underlying assumptions, or detailed experimental evidence. [1]
1 Sept 2026
The Aero Hand Open preprint says it releases a tendon-driven hand design alongside simulation, actuation-mapping, reinforcement-learning, and deployment resources. Its reported no-fine-tuning deployment workflow has not been independently verified in the supplied evidence.
1 Sept 2026
The authors of an arXiv preprint report that simulated speech-recognition errors can make some voice-controlled embodied-AI safety failures more likely in their benchmark evaluations. They also report inconsistent benefits from automatic correction. Detailed methods and independent verification are not supplied.
2 Sept 2026
DeepSeek’s model card documents an experimental image-and-text model repository with prompt-encoding tools, minimal PyTorch inference and example vLLM and SGLang serving paths. Its capability and benchmark statements are vendor-reported, while access conditions, full resource requirements and operational limits remain unestablished in the supplied evidence.
31 Aug 2026
MAELLE is an author-described reaction-prediction method based on discrete flow matching over graph-structured electron occupations. The preprint reports interpretable electron-rearrangement paths, competitive USPTO-480K performance, out-of-distribution robustness and side-product prediction. The available arXiv record does not provide the quantitative, methodological or independent evidence needed to verify those claims.
31 Aug 2026
MCR-Bench is an arXiv preprint whose authors propose a defect state-aware benchmark for multi-round code review. They describe 2,269 tasks across five languages with defect metadata and cross-round state labels, and report that tested mainstream LLMs struggle more as review rounds increase. The supplied record does not provide the methods, model scores, access details or independent verification needed to assess those claims fully.
30 Aug 2026
The authors of the SWE-Prime preprint propose filtering software-agent training data at trajectory and segment levels while retaining full sequence context. They report that a selected 10% trajectory subset beat full-dataset training on two SWE-Bench evaluations, but the supplied evidence is limited to a preprint abstract for which peer review is not established and does not support conclusions about reproducibility or broader performance.
30 Aug 2026
WikiSkill is an arXiv preprint whose authors propose consolidating AI-agent execution experience into a persistent wiki that informs later skill updates. They report gains over comparison methods and no-skill baselines, stronger benefits for larger models, cases where smaller models with skills surpass larger unskilled models, and cross-model transfer in which other-model skills can outperform self-evolved ones. The supplied abstract lacks the evaluation detail and independent validation needed to confirm those claims.
29 Aug 2026
An arXiv preprint proposes without-replacement sampling and inverse-probability reweighting as an alternative to threshold-only pruning for estimating probabilities in transduced language models. Its authors report improved compute-variance or error results in text and DNA tests, plus major runtime savings in one DNA-to-amino-acid case, but the supplied evidence does not establish peer review or replication.
28 Aug 2026
MyoMechanix is an unreviewed preprint proposing a multimodal benchmark, a structured fitness knowledge graph and a compositional analysis system for examining weight-loaded movement. Its scale and reported performance are author claims that remain unverified in the supplied evidence.
28 Aug 2026
The authors propose a Visual Dependence-Aware framework for continual, unlabeled post-training of multimodal language models. Its two mechanisms are intended to reduce cross-modal forgetting and support new-task adaptation, but the supplied arXiv record does not provide enough experimental detail to verify performance, costs, or reproducibility. [1]
