2026-08-13·Daily digest

🔬 Research Pulse

Daily Digest

August 13, 2026

🤖 AI

🧠 LLMs

1. Diagram-MMU: A Multi-Modal Benchmark for Scientific Diagrams

Authors: Weihao Bo, Shan Zhang, Yanpeng Sun... Published: 2026-08-12 | Citations: 0 arXiv | PDF

Research Question: How well can current MLLMs parse, edit, and reason about scientific diagrams — particularly for diagram-to-TikZ code generation used in scientific writing workspaces like OpenAI Prism?

Summary: Diagram-MMU is a 3.7k-diagram, 18.3k-question benchmark evaluating MLLMs on scientific diagram parsing, editing, and QA across six domains, including agentic variants. Evaluation of 12 models shows diagram-to-code is substantially harder than QA, and that agentic scaffolding helps parsing/editing but hurts QA — except for Claude-4.6 Opus, which improves everywhere.

Key Results: Built Diagram-MMU benchmark with 3.7k curated diagrams and 18.3k human-validated questions across six domains. Evaluated 12 MLLMs across three tasks (diagram-to-code parsing, diagram-to-code editing, diagram QA) plus agentic variants. Demonstrated that diagram-to-code tasks are harder than diagram QA; in agentic settings, most models improve on parsing/editing but degrade on QA, while Claude-4.6 Opus improves consistently across all three tasks.

Key Findings:

Diagram-to-code (parsing and editing) is systematically harder for MLLMs than diagram question answering
Agentic settings help parsing/editing but degrade QA performance for most models
Claude-4.6 Opus is the only tested model to gain consistently across all three tasks under agentic settings

Technical Novelty: First benchmark to jointly evaluate diagram parsing, code editing, and QA on scientific diagrams with human-validated questions, and to compare agentic vs non-agentic pipelines across all three tasks — prior diagram benchmarks focus mainly on QA or single-format chart-to-code.

What's New: Combines three tightly related tasks (parse, edit, QA) on scientific diagrams with a human-validated QA set, and explicitly benchmarks agentic vs non-agentic pipelines — motivated by real workspace features (Prism's diagram-to-TikZ) rather than synthetic tasks.

Extension Opportunities:

Train a specialized diagram-to-TikZ model via supervised fine-tuning or RL on the 3.7k Diagram-MMU corpus, using rendered-output similarity as reward to close the parsing gap
Build a diagram-editing agent that decomposes edits into locate→modify→re-render loops, targeting the specific agentic-editing degradations reported for QA
Extend the benchmark beyond TikZ to Asymptote, Matplotlib, Mermaid, and PGFPlots to test cross-format code generalization

Replicability: Project page listed (vi-ocean.github.io/projects/diagram-mmu) but code/data release status not confirmed in the abstract. Reproduction requires API access to 12 MLLMs (including Claude-4.6 Opus, GPT-class models) plus a TikZ rendering pipeline — evaluation is inference-only, so compute is modest (API cost-bound rather than GPU-bound).

Research Gaps:

No method proposed to actually close the diagram-to-code gap the benchmark exposes
Unclear why agentic scaffolding degrades QA for most models but not Claude-4.6 Opus — the mechanism is not analyzed

2. One Frozen Simulator Is Not Enough: Simulator Collapse in Multi-Agent RL

Authors: Simon Yu, Nicholas Tomlin, Marwa Abdulhai... Published: 2026-08-12 | Citations: 0 arXiv | PDF

Research Question: Does training an LLM policy against a single frozen LLM simulator generalize to unseen simulators and real users in multi-agent RL for human-AI interaction?

Summary: The paper identifies 'simulator collapse' — a failure mode where RL policies trained against a single mode-collapsed LLM simulator overfit and fail to transfer to new simulators or real users. It proposes Verbalized Sampling (inference-time) and Co-Training against a simulator population (training-time), showing 9–14% held-out gains across three benchmarks plus a human study, and releases SCOPE as an open framework.

Key Results: Demonstrated 'simulator collapse': policies trained via single-simulator RL overfit to the simulator's dominant mode. On three multi-turn benchmarks (Persuasion for Good, τ²-bench, CooperBench), Verbalized Sampling improved held-out success by up to 9% over single-simulator RL, and Co-Training pushed gains to 14%. A human study confirmed similar gains on real users, and both methods preserved policy diversity that otherwise collapses.

Key Findings:

Single-simulator RL systematically fails to generalize due to mode collapse in the simulator LLM, not just the policy
Verbalized Sampling (inference-time) yields up to 9% held-out success gain; Co-Training (training-time population) yields up to 14%
Human study confirms transfer gains to real users, and both methods preserve policy diversity

Technical Novelty: Two complementary interventions: (1) Verbalized Sampling — an inference-time method where the simulator emits an explicit response distribution to sample from, mitigating mode collapse without retraining; (2) Co-Training — a training-time method that jointly optimizes the policy against a population of trainable simulators rather than one frozen LLM. Prior multi-agent RL for LLMs typically uses a single fixed simulator; this work formalizes why that fails and proposes population-based training as the fix.

What's New: First to formalize simulator collapse as distinct from policy collapse in multi-agent LLM RL, and to propose population-based Co-Training of trainable simulators as a remedy — reframing the generalization problem as one of environment diversity, not policy regularization.

Extension Opportunities:

Apply SCOPE's Population Co-Training framework to domain-specific agents (e.g., customer support, tutoring, negotiation) with diverse trainable simulator populations
Extend Verbalized Sampling to elicit distributions conditioned on demographic/persona axes to test fairness and coverage of long-tail user behaviors
Investigate curriculum strategies over simulator populations — e.g., adversarially evolving simulators alongside policies to prevent equilibrium mode collapse

Replicability: Authors release SCOPE, an open-source framework for Population Co-Training multi-agent RL. Compute would require RL fine-tuning of both a policy LLM and a population of simulator LLMs across three benchmarks — likely multi-GPU (8xA100-class) for meaningful reproduction; exact scale not stated in abstract.

Research Gaps:

Scaling Co-Training to larger simulator populations and evaluating diminishing returns / compute trade-offs
Understanding how simulator diversity relates to real human behavioral diversity — the human study is a first step but coverage of long-tail users remains open

3. AVA-Encoder: Towards Agent-Native Video Representation Learning

Authors: Chuyue Li, Jinpeng Yu, Haozhe Wang... Published: 2026-08-12 | Citations: 0 arXiv | PDF

Research Question: How can creative agents learn structured, editable representations from high-quality films that are faithful to content and directly usable for agentic reasoning and manipulation of cinematic-grade video?

Summary: AVA-Encoder is an agentic video auto-encoder that encodes films into a hierarchical knowledge graph with linked text/image/audio/video assets and reconstructs video from it, trained via a textual-gradient optimization loop that converts reconstruction feedback into natural-language policy updates. It outperforms the strongest external baseline by 20.7 pp and beats a hand-tuned policy while using 74.3% fewer prompt tokens.

Key Results: AVA-Encoder improves by 20.7 percentage points over the strongest external baseline on their agentic video reconstruction benchmark. In a controlled policy-only setting, the pseudo-trained shot-level Agentic Video Encoder policy outperforms a carefully human-tuned policy while using 74.3% fewer system-prompt tokens. They release a new dataset of high-quality film KG representations.

Key Findings:

A structured KG-based latent representation with typed edges is a viable agent-native alternative to opaque neural video embeddings
Textual-gradient optimization (reconstruction diff → NL update direction) can pseudo-train encoding policies without gradient-based training
Auto-tuned policies can beat human-engineered prompts on both quality (+20.7 pp) and efficiency (−74.3% prompt tokens)

Technical Novelty: Two novel pieces: (1) an agentic auto-encoder that uses a hierarchical knowledge graph with linked text/image/audio/video asset layers and typed edges as the latent representation, and (2) a textual-gradient optimization loop where reconstruction differences are converted to natural-language update directions, split into Data-Independent Encoding Policy Pseudo-Training (outer) and optional Data-Dependent KG Representation Refinement (inner, test-time).

What's New: Prior work treats video representation learning as producing dense neural embeddings; this paper reframes it as an agentic auto-encoding problem where the latent is a human/agent-readable knowledge graph, and training happens via natural-language 'gradients' rather than backprop.

Extension Opportunities:

Extend the KG schema and typed edges to capture longer narrative/temporal arcs (multi-scene story graphs) rather than shot-level structure, enabling full-film generation agents
Apply the textual-gradient outer-loop pseudo-training to other multimodal auto-encoding domains (e.g., 3D scenes, game levels, or interactive simulations) where structured intermediate representations aid agent editing
Build an interactive film-editing tool on top of the KG layer that lets users query/edit nodes and re-render assets, exposing the agent-native representation as a creative UI

Replicability: Authors state they release the complete framework, benchmark, and film KG dataset. Compute needs are not specified in the abstract but agentic pipelines with video generation typically require substantial LLM/VLM API budget plus GPU inference for asset reconstruction — likely non-trivial to fully reproduce.

Research Gaps:

No standard benchmark for evaluating agent-editable video representations at cinematic quality — the paper introduces one but broader adoption is unproven
Scaling from shot-level to full-film, multi-scene narrative representations remains open

🦾 ROBOTICS

1. HandEdit: A Unified Benchmark for Egocentric Human-to-Robot Dexterous Hand Image Editing

Authors: Zhenjie Yang, Xingyu Jiao, Guopeng Zhong... Published: 2026-08-12 | Citations: 0 arXiv | PDF

Research Question: How can we bridge the embodiment gap between abundant egocentric human hand videos and scarce robotic teleoperation data to enable scalable dexterous manipulation learning, given that general image-editing models lack the embodiment-specific priors needed to transform human hands into robotic hands within egocentric frames?

Summary: HandEdit introduces a unified large-scale (200M+ instances, 26 URDFs) embodiment-aware image-editing dataset and benchmark that transforms egocentric human hands/arms into diverse robotic dexterous embodiments. It provides two evaluation tracks and a multi-dimensional metric suite, benchmarking 11 baselines to expose the embodiment gap in current editing models and unlock scalable robotic learning from human videos.

Key Results: The authors constructed HandEdit, containing over 200M editing instances derived from 5 diverse source datasets, covering 26 distinct URDFs (13 hand-only and 13 hand-arm configurations). They benchmarked 11 representative image-editing baselines across two tracks (Hand-only and Hand-Arm) using a multi-dimensional metric suite combining generic similarity metrics, VLM-based judgment, and embodiment-aware metrics, demonstrating the inadequacy of existing general-purpose editing models for embodiment-aware translation.

Key Findings:

General-purpose image-editing models lack embodiment-specific priors needed for accurate human-to-robot hand translation
URDF conditioning is a viable mechanism for controlling embodiment identity in generative editing
A multi-dimensional metric suite (similarity + VLM judgment + embodiment-aware) is required — generic similarity metrics alone don't capture embodiment fidelity

Technical Novelty: First large-scale, URDF-conditioned image-editing dataset explicitly designed for human-to-robot embodiment translation in egocentric frames, coupled with an embodiment-aware evaluation protocol — prior editing benchmarks target generic scene/object edits without embodiment priors or kinematic conditioning.

What's New: Unlike prior image-editing benchmarks focused on generic object/scene edits, HandEdit is embodiment-aware and URDF-conditioned, targeting the specific bottleneck of dexterous manipulation data scarcity, and spans both hand-only and hand-arm configurations at a scale (200M instances) unmatched in this niche.

Extension Opportunities:

Fine-tune diffusion-based editing models (e.g., InstructPix2Pix, FLUX) on HandEdit with URDF conditioning to produce a specialized embodiment-aware editor, then measure downstream policy transfer improvements on real dexterous manipulation tasks
Build a video-level extension that enforces temporal consistency across edited frames (rather than per-frame editing) to enable direct co-training of visuomotor policies from human-to-robot translated trajectories
Add novel URDFs (e.g., soft/pneumatic hands, prosthetics) via a plug-in adapter that generates synthetic editing pairs from a URDF spec alone, extending coverage beyond the 26 configurations

Replicability: The abstract states HandEdit is a released benchmark, implying dataset and protocol availability, but code release is not explicitly confirmed. Reproducing baseline evaluations across 11 models on 200M instances would require substantial GPU compute (likely hundreds of A100/H100 hours); dataset generation from 5 source datasets requires their licenses.

Research Gaps:

No evidence yet that models trained on HandEdit-translated data yield improved downstream robot policy performance — the loop from editing quality to manipulation success is not closed
Coverage is limited to 26 URDFs and 5 source datasets; generalization to unseen embodiments and out-of-distribution egocentric viewpoints remains unmeasured

2. G0.5: One Autoregressive Stream for Robot Reasoning and Action

Authors: Yicheng Liu, Zibin Dong, Baijun Ye... Published: 2026-08-12 | Citations: 0 arXiv | PDF

Research Question: How can we build a Vision-Language-Action (VLA) model where reasoning and action are unified under a single autoregressive objective, rather than bolting a separately-trained flow-matching action expert onto a frozen VLM that acts merely as a context encoder?

Summary: G0.5 replaces the standard VLM-plus-flow-matching-action-expert VLA recipe with a single autoregressive transformer that decodes reasoning and action tokens under one objective, enabled by a cross-embodiment action tokenizer, an interleaved chain-of-thought stream, and a visual memory module. Because reasoning and action share weights, VLM capabilities transfer to physical behavior, and the model beats π_0.5, GR00T-N1.7, and the BEHAVIOR 2025 challenge winner across 7 real and simulated benchmarks.

Key Results: G0.5 surpasses SOTA across 7 regimes: real-world fine-tuning on R1lite/R1pro (76.7% vs 53.3% for π_0.5 and 24.4% for GR00T-N1.7); 2025 BEHAVIOR Challenge on 50 long-horizon household mobile manipulation tasks (31.4% vs 26.3% for π_0.5 and 26.1% for the challenge winner); DROID post-training with zero-shot transfer to unseen scenes (82.5%); LIBERO (98.9%); RoboTwin 2.0 (93.3%); SimplerEnv-Bridge (87.3%); plus a language-following Pick-and-Place benchmark.

Key Findings:

A single autoregressive stream for reasoning + action outperforms decoupled VLM + flow-matching-expert designs at foundation-model scale (76.7% vs 53.3% for π_0.5 on real-world R1lite/R1pro fine-tuning)
A learnable cross-embodiment action tokenizer makes it feasible to share one action vocabulary across heterogeneous robots, enabling unified pretraining
Because weights are shared, natural-language prompts directly steer action granularity, task horizon, and out-of-distribution scene handling with no additional training — a capability generally absent from action-expert architectures

Technical Novelty: Unified single-transformer autoregressive decoder emitting both reasoning and action tokens under one objective (vs. the dominant VLM + separate flow-matching action-expert recipe of π_0/π_0.5 and GR00T). Three enabling pieces: (1) a learnable cross-embodiment action tokenizer producing a shared discrete action vocabulary across heterogeneous robots, (2) a native chain-of-thought stream interleaving task decomposition, object grounding, and action hints with action tokens, and (3) a visual memory module injecting multi-second history through the vision encoder.

What's New: Prior VLAs (π_0, π_0.5, GR00T-N1.7) treat the VLM as a context encoder feeding a separately trained flow-matching action expert. G0.5 collapses this into one decoder with one objective, and shows that the resulting model inherits instruction-following, grounding, and generalization directly from the VLM — while still winning on low-level control benchmarks.

Extension Opportunities:

Extend the learnable cross-embodiment action tokenizer to dexterous hands or humanoid whole-body control by expanding the shared action vocabulary and retraining on humanoid teleop datasets
Exploit the single-stream design for test-time scaling: use CoT sampling / self-consistency over the reasoning tokens before committing to an action trajectory in long-horizon tasks
Distill G0.5 into a smaller on-robot model using the reasoning traces as supervision, so edge deployments retain instruction-following behavior without foundation-model latency

Replicability: Abstract does not mention code, weights, or dataset release. Reproduction would require foundation-model-scale compute (large multi-robot pretraining corpus plus VQA data) on the order of hundreds to thousands of GPU-days, plus access to R1lite/R1pro hardware, DROID, BEHAVIOR, LIBERO, RoboTwin 2.0, and SimplerEnv-Bridge for evaluation — practical only for well-resourced labs.

Research Gaps:

No ablation-level evidence in the abstract on how much each of the three components (cross-embodiment tokenizer, interleaved CoT, visual memory) individually contributes to the gains
Inference latency and control frequency of an autoregressive action decoder vs. a flow-matching expert are not reported, which is critical for real-robot deployment

3. A Frequency-Space Terahertz Transceiver Chip for Multi-Agent Communications and Spatial Awareness

Authors: Xiaoyue Xia, Zhicheng Lin, Hao Guo... Published: 2026-08-12 | Citations: 0 arXiv | PDF

Research Question: How can indoor embodied-intelligence systems achieve scalable multi-agent THz connectivity and mutual spatial awareness without the complexity of dense beamforming networks, element-level phase control, or external THz illumination required by conventional phased arrays and programmable metasurfaces?

Summary: The authors present a monolithic 65-nm CMOS THz transceiver (208-258 GHz, 1.5x4.9 mm) that integrates broadband front ends with a heterogeneous leaky-wave metasurface, using frequency-controlled beam steering to enable spatial-frequency division multiple access. It demonstrates simultaneous multi-agent communication, 2D localization, and sensing-enhanced links on a single chip, replacing dense phased-array beamforming networks with just four meta-atoms.

Key Results: Demonstrated a fully integrated 208-258 GHz THz transceiver in 65-nm CMOS occupying only 1.5mm x 4.9mm, achieving 75-degree frequency-controlled beam scanning using only 4 meta-atoms via a heterogeneous leaky-wave metasurface (HLM). Proved simultaneous multi-agent transmit/receive via spatial-frequency division multiple access (SFDMA), 2D localization, and sensing-enhanced communication on a single monolithic chip.

Key Findings:

Only 4 meta-atoms achieve 75-degree beam scanning across 208-258 GHz via dispersive leaky modes
Frequency-to-space mapping enables SFDMA, letting distinct frequencies serve distinct spatial users simultaneously without dedicated beamformers
Single chip delivers combined ISAC: simultaneous multi-agent comms, 2D localization, and sensing-enhanced communication

Technical Novelty: Monolithic co-integration of broadband CMOS THz front ends with a heterogeneous leaky-wave metasurface aperture on the same die, plus joint frequency-domain and spatial-domain mixing co-design that maps frequency directly to beam angle. Unlike phased arrays (needing per-element phase shifters) or programmable metasurfaces (needing element-level control and external illumination), only 4 meta-atoms deliver 75-degree scanning.

What's New: First fully monolithic CMOS THz transceiver with an integrated heterogeneous leaky-wave metasurface aperture that eliminates the need for element-level phase control, external THz illumination, or dense beamforming networks - achieving multibeam ISAC through frequency-space co-design instead of hardware complexity.

Extension Opportunities:

Scale from 4 meta-atoms to larger heterogeneous arrays to expand angular coverage beyond 75 degrees and increase user capacity per chip
Build a multi-chip testbed for embodied-intelligence robots/drones combining SFDMA links with real-time SLAM using the on-chip sensing data
Extend the HLM co-design methodology to sub-THz bands (140 GHz, 300 GHz+) for 6G ISAC standardization or to silicon photonics for optical analogues

Replicability: No code, dataset, or tape-out files mentioned in the abstract. Reproduction requires 65-nm CMOS fabrication access (multi-$100K), THz measurement infrastructure (VNAs, THz probe stations, anechoic chambers to 258 GHz), and RF/EM co-simulation tooling (HFSS, Cadence Virtuoso). Effectively out of reach for non-fab research groups.

Research Gaps:

Conventional THz phased arrays and programmable metasurfaces do not scale for multi-agent operation due to dense beamforming networks and element-level control
Existing metasurface approaches typically require external THz illumination, preventing true single-chip integration for embodied systems

💻 COMPUTE

1. An Event-Driven Cloud-Native Wearable Analytics Framework for Real-Time Clinical Workloads

Authors: Elias Grünewald, Daniil Cherepko, Linus Gustafsson... Published: 2026-08-11 | Citations: 0 arXiv | PDF

Research Question: How can continuous physiological data from heterogeneous consumer wearables be ingested, normalized to a medical standard, and analyzed at scale in real-time without vendor lock-in, while meeting healthcare regulatory requirements?

Summary: The paper introduces an open-source, event-driven, cloud-native platform that ingests high-frequency wearable vital signs, normalizes them to FHIR via stream processing, and serves them into a medallion lakehouse for real-time analytics and ML. Its key contributions are a dependency-aware FHIR minimization scheme and a vendor-neutral architecture that meets healthcare compliance while sustaining sub-8ms median ingestion latency at 50 RPS.

Key Results: The ingestion pipeline sustains 50 full ingestion requests per second with median response times under 8 ms, satisfying low-latency requirements for real-time patient monitoring. The system demonstrates lossless FHIR resource reconstruction via a dependency-aware minimization scheme and integrates role-based access control plus secure service-to-service communication for regulatory compliance.

Key Findings:

A microservice, event-driven architecture can meet real-time clinical latency budgets (<8 ms median) at 50 RPS on commodity infrastructure
Dependency-aware FHIR minimization can shrink storage without losing reconstructability, addressing FHIR's verbosity problem
A cross-platform mobile client leveraging native health frameworks is sufficient to abstract over fragmented wearable ecosystems without vendor SDK lock-in

Technical Novelty: A dependency-aware FHIR minimization scheme that reduces storage overhead while allowing lossless reconstruction of FHIR resources, combined with an event-driven stream-processing pipeline that maps heterogeneous wearable data to FHIR in a cloud-native microservice architecture — prior work typically bolts FHIR mapping onto batch ETL or requires vendor-specific SDKs.

What's New: Unlike prior wearable pipelines that are vendor-specific, batch-oriented, or skip standards compliance, this work unifies vendor-neutral acquisition, streaming FHIR normalization, storage-efficient FHIR representation, and a full ML lifecycle (medallion lakehouse + model serving) in a single compliance-ready open-source stack.

Extension Opportunities:

Add federated learning across institutions on the medallion lakehouse layer so ML models can train on multi-hospital wearable streams without moving PHI
Extend the FHIR minimization scheme to other emerging standards (e.g., OpenEHR, SMART on FHIR bulk data) and benchmark compression ratios across vital-sign types
Plug in on-device edge inference (via the cross-platform mobile app) to pre-filter anomalies before ingestion, reducing bandwidth and enabling offline clinical alerts

Replicability: The authors state the implementation is open-source. Reproduction requires a Kubernetes-class cluster, a stream processing engine (e.g., Kafka/Flink), a medallion lakehouse stack (e.g., Delta/Iceberg), and mobile devices with native health frameworks (HealthKit/Health Connect) for end-to-end evaluation. Load testing at ~50 RPS is modest and reproducible on a small cluster.

Research Gaps:

No end-to-end clinical validation study — performance is measured in RPS/latency, not in downstream diagnostic accuracy or clinician workflow impact
Scalability ceiling beyond 50 RPS and behavior under adversarial conditions (network partitions, device dropouts, burst traffic from thousands of patients) is not characterized

2. FQTree: Fine-grained Quantization and Hardware Generation of Boosted Decision Trees

Authors: Zhiqiang Que, Chang Sun, Haiyang Wang... Published: 2026-08-12 | Citations: 0 arXiv | PDF

Research Question: How can boosted decision trees (BDTs) be quantized and compiled into FPGA hardware in a way that minimizes LUT cost without sacrificing accuracy, avoiding the waste of uniform or hand-tuned fixed-point formats?

Summary: FQTree introduces a quantization-aware training method for boosted decision trees that uses a global step plus per-tree shift to produce compact non-negative integer leaves, and folds this into the boosting loop so later trees correct earlier quantization error. Paired with the QXGB compiler, it auto-generates FPGA implementations that cut LUT usage by 26–57% versus prior BDT hardware designs while preserving accuracy on JSC, MNIST, and NID.

Key Results: On JSC, MNIST, and NID benchmarks, FQTree/QXGB reduces FPGA LUT usage by 26–57% versus state-of-the-art FPGA-based BDT designs while matching or improving accuracy. The scheme uses a global quantization step plus per-tree shift, controlled clipping/pruning, and bias folding, with quantization applied during boosting so later trees compensate for earlier quantization error.

Key Findings:

Fine-grained leaf quantization with a global step + tree-wise shift yields substantially cheaper datapaths than uniform fixed-point BDT representations.
Applying quantization during (not after) boosting lets later trees compensate for accumulated quantization error, preserving accuracy at low bitwidths.
Automated compiler-based lowering (QXGB) achieves 26–57% LUT reduction over SOTA FPGA BDTs on JSC, MNIST, and NID with matched or improved accuracy.

Technical Novelty: A hardware-oriented leaf-value quantization scheme using a shared global step with tree-wise shifts to yield compact non-negative integer leaves, plus quantization-aware boosting where subsequent trees are trained against the already-quantized ensemble (rather than post-training quantization of a trained BDT), combined with bias folding to shrink the accumulator datapath.

What's New: Prior FPGA BDT work relied on uniform or hand-tuned fixed-point formats applied post-training; FQTree instead does quantization-aware boosting with a hardware-motivated leaf encoding (global step + shift, non-negative integers, bias folding) and ships an end-to-end compiler that turns the trained model directly into low-latency hardware.

Extension Opportunities:

Extend the fine-grained leaf-value quantization + tree-wise shift scheme to other tree ensembles (LightGBM, CatBoost, Random Forests) and gradient-boosted regressors, not just XGBoost classification.
Combine FQTree with feature-input quantization and mixed-precision split thresholds so the whole datapath (features, comparators, accumulator) is co-optimized, not only leaves.
Port QXGB's compiler flow to ASIC / RISC-V vector back-ends or integrate it into hls4ml-style pipelines for particle-physics triggers and other sub-microsecond inference domains.

Replicability: Code is released at https://github.com/ecs-bristol/FQTree. Reproduction requires standard XGBoost-style training compute (single GPU or CPU sufficient for JSC/MNIST/NID) plus an FPGA toolchain (e.g., Vivado/Vitis HLS) for hardware synthesis and LUT measurements.

Research Gaps:

Extends only leaf-value quantization; feature and threshold precision are not jointly optimized in the same fine-grained framework.
Evaluation is limited to three benchmarks (JSC, MNIST, NID) on FPGA — generalization to larger tabular workloads, ASIC targets, and regression tasks is unaddressed.

3. Design of monolithic microcavities for enhancing organic quantum emitters

Authors: Tim Hebenstreit, Jaime Gimeno Balaguer, Fridtjof Betz... Published: 2026-08-12 | Citations: 0 arXiv | PDF

Research Question: How can organic single-molecule quantum emitters be integrated into monolithic microcavities to suppress phonon sideband emission and enhance the zero-phonon line (ZPL), overcoming the clean-room fabrication incompatibility that has historically prevented such integration?

Summary: The paper proposes three monolithic microcavity designs that provide Purcell enhancement of the zero-phonon line for organic single-molecule quantum emitters, addressing a long-standing incompatibility between organic materials and clean-room monolithic integration. The designs are obtained through Bayesian optimization coupled with AAA rational approximation for efficient spectral simulation, and are claimed compatible with existing fabrication and single-molecule preparation workflows.

Key Results: The paper proposes three monolithic microcavity designs providing sufficient Purcell enhancement to preferentially amplify emission into the 0-0 zero-phonon line. Designs were optimized using Bayesian optimization combined with the adaptive Antoulas-Anderson (AAA) rational approximation algorithm for efficient spectral reconstruction from scattering simulations. Specific Purcell factors, Q-factors, and mode volumes are not quoted in the abstract, but the designs are claimed compatible with both standard clean-room processing and single-molecule preparation techniques.

Key Findings:

Three monolithic cavity geometries can deliver sufficient Purcell enhancement to preferentially funnel emission into the 0-0 ZPL, suppressing the phonon sideband contribution
Cavity enhancement simultaneously improves spectral purity, photon extraction efficiency, and shortens the excited-state lifetime — relaxing constraints for indistinguishable photon generation
Bayesian optimization combined with AAA rational approximation is an effective design pipeline for optimizing high-Q photonic resonances against scattering simulations

Technical Novelty: First proposed monolithic microcavity designs specifically tailored to organic quantum emitters — a materials class previously excluded from monolithic integration. Methodologically novel in combining Bayesian global optimization with AAA rational approximation to reconstruct spectra efficiently from scattering simulations, enabling tractable global design of high-Q resonances.

What's New: Prior work on organic emitters has been unable to achieve monolithic microcavity integration due to clean-room incompatibilities; this paper is the first to present designs that reconcile the two. The AAA+Bayesian design methodology for organic-compatible resonators is also new to this domain.

Extension Opportunities:

Experimental fabrication and characterization of the proposed cavity designs with a real organic emitter (e.g., DBT in anthracene) to validate predicted Purcell enhancements and ZPL fractions
Extend the Bayesian + AAA optimization framework to co-optimize for emitter position tolerance and fabrication disorder, producing designs robust to realistic clean-room yield
Integrate the cavity-enhanced source into a photonic circuit for on-chip HOM interference experiments to benchmark indistinguishability improvements

Replicability: No code or data availability is stated in the abstract. Reproduction would require electromagnetic scattering solvers (e.g., JCMsuite, COMSOL, or Meep), a Bayesian optimization library, and an AAA implementation — all runnable on a modest workstation or small cluster. No experimental fabrication is reported.

Research Gaps:

No experimental demonstration — the paper is purely a design/simulation study, leaving fabrication yield and real-emitter coupling unproven
Quantitative comparison against non-monolithic (e.g., open Fabry-Perot) cavity approaches for organic emitters is not evident from the abstract

⚡ ENERGY

1. Electronic structure, band offset, and interface electron population of the LaInO$_3$/BaSnO$_3$ system

Authors: G. Hoffmann, A. A. Riaz, C. Kalha... Published: 2026-08-12 | Citations: 0 arXiv | PDF

Research Question: How do the electronic structure, band alignment, and interface charge distribution of BaSnO3/LaInO3 heterostructures explain the formation of high-mobility 2DEGs, and how do these properties evolve with LaInO3 overlayer thickness?

Summary: The paper combines soft and hard X-ray photoelectron spectroscopy with hybrid DFT to characterize the electronic structure of BaSnO3, LaInO3, and their heterostructures, directly measuring band offsets and 2DEG populations at buried interfaces as a function of LaInO3 overlayer thickness. It provides a unified picture of chemical bonding and interface charge that underpins the >100 cm²/Vs room-temperature electron mobilities recently reported in this system.

Key Results: Combined SXPS/HAXPES measurements with hybrid DFT-derived DOS characterize core, semi-core, and valence states of BaSnO3, LaInO3, and their heterostructures. Directly probes band offset and 2DEG population at the interface via HAXPES across varying LaInO3 overlayer thicknesses. Context: recent thin-film advances achieve room-temperature electron mobilities >100 cm²/Vs at the interface.

Key Findings:

Core, semi-core, and valence-state analysis yields a consistent chemical-bonding and electronic-structure picture for both parent oxides and the heterostructures
HAXPES directly resolves the BaSnO3/LaInO3 band offset and the population of 2DEG states at the buried interface
Hybrid DFT DOS quantitatively matches photoemission spectra, validating the theoretical description of these perovskite oxides

Technical Novelty: Direct experimental probe of buried interface band offset and 2DEG population using HAXPES (rather than inferring from transport alone), cross-validated against hybrid DFT DOS across the parent oxides and heterostructures with varying overlayer thickness.

What's New: Prior work established the existence and high mobility of the BaSnO3/LaInO3 2DEG largely through transport; this study directly spectroscopically probes the buried interface electronic structure and band offset with HAXPES, tied to hybrid-DFT calculations across systematically varied overlayer thicknesses.

Extension Opportunities:

Systematic thickness-dependent HAXPES + transport co-measurements to map 2DEG carrier density vs LaInO3 layer count and correlate with mobility
Explore alternative polar overlayers (e.g., LaAlO3, NdInO3, GdScO3) on BaSnO3 to engineer larger band offsets and higher 2DEG densities
Apply the same SXPS/HAXPES+hybrid-DFT workflow to gated or field-effect device stacks to probe operando band bending under bias

Replicability: No explicit code/data availability mentioned in the abstract. Reproduction requires MBE/PLD thin-film growth capability, synchrotron access for HAXPES (e.g., PETRA III), lab-based SXPS, and hybrid-DFT compute (VASP/CRYSTAL-class, tens to hundreds of core-hours per structure).

Research Gaps:

Lack of direct spectroscopic characterization of buried oxide interfaces (as opposed to transport-only measurements)
Limited experimentally-validated first-principles description of band alignment in transparent-conducting-oxide-based 2DEG systems

2. Near-Unity Excitation and Radiative Efficiencies in Electroluminescence Without External Carrier Injection

Authors: Rui Li, Xinrui Li, Jiachen Xie... Published: 2026-08-12 | Citations: 0 arXiv | PDF

Research Question: How can non-injecting electroluminescence (NI-EL) overcome its inherent limitations of weak emission and high driving voltage caused by low excitation and radiative recombination efficiencies, to rival injection-based EL devices?

Summary: The paper demonstrates non-injecting electroluminescence (NI-EL) devices that achieve near-unity excitation efficiency and dramatically improved radiative efficiency through a combined strategy of remote charge reservoirs and graded-shell quantum dots. The resulting RGB devices match or exceed conventional injection-based LEDs in brightness (291,628 nits green), power efficiency (302.6 lm/W), and stability, with sub-100 ns pulsed response.

Key Results: Demonstrated NI-EL with near-unity excitation efficiency (>20-fold improvement over benchmark) using remote state-abundant charge reservoirs, plus 7-8-fold radiative efficiency gain via nonmonotonically graded-shell QDs. RGB device metrics: red turn-on at 3.7 Vrms; green luminance of 291,628 nits with power efficiency of 302.6 lm/W; first QD-based blue NI-EL surpassing LED counterpart stability. External EL efficiency up to 45.4%, sub-100 ns response time, minimal dielectric loss.

Key Findings:

Remote charge reservoir architecture boosts NI-EL excitation efficiency >20-fold to near unity
Nonmonotonically graded QD shells improve high-field radiative efficiency 7-8x, enabling 45.4% external EL efficiency
First QD-based blue NI-EL device demonstrates emitter stability superior to its injection-LED counterpart

Technical Novelty: Two-part innovation: (1) operational paradigm using remote, state-abundant charge reservoirs to decouple excitation from injection bottlenecks, and (2) nonmonotonically graded QD shells that maintain radiative efficiency at high fields — combining device architecture with materials engineering to break the excitation/radiative efficiency ceiling in NI-EL.

What's New: Reframes NI-EL from a weak-emission curiosity into a competitive display/pulsed-source technology by attacking both fundamental bottlenecks (excitation and radiative efficiency) simultaneously, and delivers the first QD-based blue NI-EL — a historically difficult color for QD LEDs generally.

Extension Opportunities:

Apply the remote charge reservoir paradigm to perovskite or organic emitters to test generality beyond QDs
Engineer sub-100 ns pulsed NI-EL sources for visible light communication (Li-Fi) or time-of-flight sensing applications
Develop flexible/transparent NI-EL displays exploiting the AC-driven architecture's simpler electrode requirements vs injection LEDs

Replicability: Abstract does not mention code/data availability. Reproduction requires QD synthesis with graded shell architecture, thin-film device fabrication (likely cleanroom-scale), AC electrical drive electronics, and optical characterization (luminance meter, transient EL spectroscopy). Materials chemistry expertise is the primary barrier, not compute.

Research Gaps:

Long-term operational lifetime and degradation mechanisms of NI-EL under continuous AC driving are not addressed
Scalability of graded-shell QD synthesis and device uniformity across large-area displays remain unexplored

🏥 HEALTHCARE

1. AI4AI at Test-Time: Strong-to-Weak Capability Transfer via Harnesses

Authors: Cheng Qian, Wenting Zhao, Liangwei Yang... Published: 2026-08-12 | Citations: 0 arXiv | PDF

Research Question: Can the capabilities of a strong model be transferred to a weaker model at inference/test time via scaffolding, without any parameter updates (as opposed to conventional training-time distillation)?

Summary: The paper introduces strong-to-weak scaffolding, where a stronger 'builder' model iteratively constructs inference-time harnesses that let a frozen weaker 'target' model solve tasks more reliably. On four Theory-of-Mind benchmarks, this test-time transfer nearly doubles target performance (0.49 → 0.91), driven primarily by offloading reasoning into deterministic code, routing, and format enforcement.

Key Results: On four Theory-of-Mind benchmarks, iterative harness construction using 5% of data as validation nearly doubled average target-model performance from 0.49 to 0.91. Gains scaled monotonically with builder-model reasoning effort, weaker target models received the largest gains, and platform effects were modest relative to builder capability.

Key Findings:

Test-time harness transfer lifts average target performance from 0.49 to 0.91 across four ToM benchmarks
Gains come from offloading brittle reasoning to deterministic code, benchmark-specific routing, and format enforcement — not from more reasoning or broader sampling by the target
Builder reasoning effort improves harness quality monotonically; weaker targets benefit most; platform effects are small relative to builder capability

Technical Novelty: Reframes model-to-model capability transfer as an inference-time harness-construction problem: a strong 'builder' iteratively refines scaffolding (code offloads, routing, format enforcement) for a frozen weaker 'target', rather than updating target parameters as in teacher-forcing or on-policy distillation.

What's New: Prior distillation work updates weights of the smaller model (teacher forcing, on-policy distillation). This work shows capability transfer can happen entirely at test time via scaffolding, positioning harness design as a complement — not a substitute — for parameter-based distillation.

Extension Opportunities:

Extend beyond Theory-of-Mind to domains where offloading to deterministic code is harder (open-ended generation, creative writing, multi-modal reasoning) to test generality
Study compositional/reusable harnesses: can harnesses built for one benchmark transfer or be combined across tasks, or are they always benchmark-specific routers?
Explore hybrid pipelines that combine test-time harness construction with lightweight fine-tuning — distilling the harness's structural patterns back into target weights

Replicability: Abstract does not mention code/data release. Reproduction would require API/inference access to multiple builder+target model pairs plus the four ToM benchmarks; compute is modest since no training is involved, but iterative harness refinement over multiple rounds implies non-trivial inference cost from the strong builder.

Research Gaps:

Evaluation is limited to Theory-of-Mind benchmarks; unclear whether gains generalize to tasks where code-offloading and strict format enforcement are not applicable
Harnesses appear benchmark-specific (routing, format rules) — the paper doesn't establish whether a single harness generalizes across tasks or how much of the gain is overfitting to the 5% validation slice

🔬 MATERIALS

1. Twist-Reconfigurable van der Waals Moiré Photonic Crystals

Authors: Hugo Quard, Jiyun Kim, Anastasiia Zalogina... Published: 2026-08-12 | Citations: 0 arXiv | PDF

Research Question: How can moiré photonic crystals be realized in the visible range with dynamic, post-fabrication twist reconfigurability, overcoming the challenges of nanoscale structuring and in-situ tunability that have prevented practical implementations?

Summary: The authors build a mechanically reconfigurable moiré photonic crystal from two patterned WS₂ slabs sandwiching an hBN spacer, and dynamically tune the twist angle between 3.8° and 8.4° while measuring the resulting 3D photonic dispersion. Reducing the twist angle densifies folded/hybridized resonances and drives a 30-fold enhancement of emission from embedded color centers, establishing a visible-range platform for in-situ dispersion engineering.

Key Results: Demonstrated a mechanically reconfigurable moiré photonic crystal built from two patterned WS₂ slabs separated by an hBN spacer, reconfigured across twist angles from 3.8° to 8.4°. Reconstructed the full 3D dispersion via momentum-resolved reflectivity spectroscopy, and observed a 30-fold enhancement of emission from embedded color centers as twist angle decreased and folded/hybridized resonances densified.

Key Findings:

A vdW moiré photonic crystal (WS₂/hBN/WS₂) can be mechanically reconfigured post-fabrication across a 3.8°–8.4° twist range
Momentum-resolved reflectivity fully reconstructs the 3D photonic dispersion and reveals folded, hybridized bands that evolve with twist angle
Smaller twist angles produce denser resonance manifolds and yield a 30-fold enhancement of color-center emission, demonstrating twist-controlled light-matter coupling

Technical Novelty: First demonstration of a mechanically twist-reconfigurable moiré photonic crystal operating in the visible range using patterned vdW slabs (WS₂/hBN/WS₂). Prior moiré photonics work relied on fixed, pre-fabricated twist angles or operated at longer wavelengths; this work uniquely combines nanoscale vdW patterning with in-situ mechanical twist tuning and demonstrates measurable light-matter coupling enhancement.

What's New: Combines two previously separate capabilities — visible-range moiré photonic band engineering and in-situ mechanical twist reconfiguration — in a single vdW heterostructure device, whereas prior moiré photonics work has been limited to fixed twist angles or non-visible wavelengths.

Extension Opportunities:

Integrate single-photon emitters (e.g., hBN defect centers or WSe₂ quantum dots) into the moiré cavity to build twist-tunable single-photon sources with programmable Purcell enhancement
Extend the platform to nonlinear photonics — leveraging WS₂'s χ⁽²⁾ response for twist-controlled second harmonic generation or parametric down-conversion
Couple the mechanical twist actuation to piezoelectric or MEMS drivers for real-time electrical control of photonic band structure, enabling reconfigurable optical filters or neuromorphic photonic devices

Replicability: Abstract does not mention code/data availability. Reproduction requires substantial nanofabrication capability: exfoliation and patterning of WS₂ monolayers/slabs, hBN spacer transfer, precise dry-transfer with rotational alignment stages, and a momentum-resolved (Fourier-plane) reflectivity spectroscopy setup. Compute demands are modest (photonic band simulations via MPB/COMSOL/FDTD), but wet-lab and cleanroom demands are high.

Research Gaps:

Absence of electrical or all-optical (rather than mechanical) tuning mechanisms limits switching speed and integration
No demonstration yet of quantum-optical functionality (single-photon emission statistics, strong coupling, cavity QED) despite the emission enhancement result

2. Mismatch between Raman shear modes and ferroelectric polarization in 3R-MoS$_{2}$

Authors: Johannes Schwandt-Krause, Jan-Niklas Heidkamp, Mohammed El Amine Miloudi... Published: 2026-08-12 | Citations: 0 arXiv | PDF

Research Question: How can sliding-ferroelectric stacking order and polarization state in parallel-stacked 2D van der Waals materials (specifically 3R-MoS$_2$) be reliably and non-destructively identified, given that individual characterization techniques appear insufficient?

Summary: The paper shows that neither KPFM nor low-frequency Raman spectroscopy alone can uniquely assign the stacking order and polarization state of trilayer 3R-MoS$_2$, since ABA and BAB configurations are KPFM-indistinguishable yet Raman-distinct. Notably, the standard bond-polarizability model fails to predict the observed shear-mode differences, indicating that interlayer Raman response involves physics beyond simple polarizability, and motivating a combined scanning-probe + spectroscopy approach.

Key Results: Using combined KPFM and low-frequency Raman spectroscopy on exfoliated trilayer 3R-MoS$_2$ on hBN, the authors demonstrated that ABA and BAB stackings (both with zero net polarization) are indistinguishable via KPFM but show drastically different low-frequency shear-mode Raman activity. The result was reproduced across multiple flakes and corroborated by low-temperature photoluminescence. The standard bond-polarizability model fails to account for the observed shear-mode difference between ABA and BAB.

Key Findings:

ABA and BAB trilayer 3R-MoS$_2$ stackings are indistinguishable via KPFM (both have zero net polarization) but show drastically different low-frequency shear modes
Low-temperature photoluminescence corroborates the shear-mode-based distinction, and the effect is reproducible across multiple flakes
The bond-polarizability model fails to explain the ABA vs BAB shear-mode activity difference, revealing physics beyond simple polarizability in the interlayer Raman response

Technical Novelty: First demonstration that KPFM alone cannot distinguish zero-net-polarization ABA vs BAB stackings in trilayer 3R-MoS$_2$, and that low-frequency shear-mode Raman exposes a stacking-dependent response not captured by the standard bond-polarizability model — motivating a combined spectroscopic + scanning-probe protocol.

What's New: Prior work assumed spectroscopic or scanning-probe techniques individually could assign sliding-ferroelectric stacking order; this work explicitly demonstrates their insufficiency and identifies a theoretical gap in the standard bond-polarizability description of interlayer Raman modes.

Extension Opportunities:

Develop a beyond-bond-polarizability theoretical framework (e.g., incorporating charge-transfer or electron-phonon coupling) to quantitatively predict shear-mode activity differences between ABA/BAB stackings
Extend the combined KPFM + low-frequency Raman protocol to other sliding ferroelectrics (WSe$_2$, hBN, MoSe$_2$) and multi-layer stacks (4L, 5L) to build a general stacking-assignment toolkit
Build an automated ML classifier that fuses KPFM maps and low-frequency Raman spectra to assign domain stacking order in real time for device fabrication workflows

Replicability: No mention of released code or data in the abstract. Reproduction requires exfoliated trilayer 3R-MoS$_2$ on hBN, a KPFM-capable AFM, low-frequency (sub-50 cm$^{-1}$) Raman spectrometer with volume Bragg gratings, and a low-temperature PL setup — significant experimental infrastructure but no unusual compute.

Research Gaps:

Absence of a theoretical framework beyond bond polarizability that captures stacking-dependent shear-mode activity in parallel-stacked vdW ferroelectrics
Lack of a standardized, non-destructive protocol for unambiguous stacking-order assignment in sliding ferroelectric device fabrication

🔥 GitHub Trending

1. vuphongle/oss-pr-reviewer

⭐ 74 stars | TypeScript

AI-powered CLI for reviewing GitHub pull requests, detecting potential bugs, security risks, regressions, and missing tests, with structured Markdown reports for open-source maintainers.

ai cli code-review developer-tools github llm

2. thiientv/godmode

⭐ 63 stars | Python

Production-grade Agent Skills for AI coding agents—composable workflows for planning, TDD, debugging, review, UI/UX, releases, incidents, and evals.

agent-evaluation agent-skills ai-agents ai-coding claude-code code-review

3. MaincodeHQ/from-tin-to-tokens

⭐ 16 stars | Unknown

A free textbook on building an LLM inference stack from scratch in Rust.

book gpu inference llm machine-learning rust

4. DukeYu65ss/presage

⭐ 4 stars | Python

The Everything Prediction Engine — five AI analyst personas debate, 10,000 Monte Carlo worldlines unfold, and superforecasting calibration converges any question into one actionable probability.

fastapi forecasting llm monte-carlo multi-agent prediction

5. uulong950/qingming-qwen3-tts

⭐ 3 stars | C++

Device-native Qwen3-TTS inference for AMD RX 7900 XTX and NVIDIA RTX 4090 with explicit HIP/CUDA kernels, streaming and Resident execution.

amd cuda gpu hip inference low-latency

6. aylmerwawa/fraud-detection-lightgbm

⭐ 2 stars | Jupyter Notebook

Fraud detection on 410k transactions with LightGBM — feature engineering, 5-fold CV, test AUC 0.9726

feature-engineering fraud-detection imbalanced-data lightgbm machine-learning python

7. siddemmohankrishna/AI-Real-Time-GYM-Coach

⭐ 2 stars | Python

Real-time AI gym coach using computer vision and pose detection to track exercises, count reps, analyze form, and provide AI voice feedback.

ai-coach artificial-intelligence computer-vision exercise-tracking fitness llm

8. falcons-eyes/pokemon_battle_prediction

⭐ 2 stars | Jupyter Notebook

🔬 Poké Lab — 포켓몬 승부 예측 데스크톱 앱(PyQt6) + 데이터 수집·EDA·머신러닝·통계 검증 파이프라인 (GroupKFold 93.6%)

data-analysis data-visualization eda jupyter-notebook korean machine-learning

9. trixmix821/deepseek-delegate-mcp

⭐ 2 stars | JavaScript

Local MCP server that lets Codex (GPT Sol) delegate boilerplate and bulk code generation to DeepSeek V4 — hybrid planner + cheap executor.

ai-agents codex deepseek llm mcp open-source

10. reh8n/a2acompress

⭐ 2 stars | Python

Lossless wire format for agent-to-agent handoffs. 36.6% fewer real cl100k_base tokens on held-out ToolBench, every byte of decode overhead counted, byte-exact round trip.

agents ai-agents compression context-window llm multi-agent

11. aliguerreroj/credit-risk-mlops

⭐ 1 stars | Python

Pipeline MLOps end-to-end en AWS para scoring de riesgo crediticio: Glue, SageMaker Pipelines, doble despliegue (Endpoint + FastAPI/ECS), monitoreo de drift, automatización con Lambda/EventBridge, Ter

api-gateway aws aws-ec2 aws-lambda aws-s3 aws-sagemaker

12. Kim-FengLei/fenglei-memory

⭐ 1 stars | Unknown

Learn, Don't Store — a long-term memory methodology for AI agents. Zero-dependency skill that turns agent collaboration logs into reusable experience, judgment standards, and skill packs.

agent-memory ai-agents ai-skills artificial-intelligence claude-skills llm

13. hibafatima-ahsan/handwritten-digit-recognition

⭐ 1 stars | Python

hand written digit recognition prediction project

anaconda learning-by-doing machine-learning python streamlit-dashboard

14. ayushmangupta1990/hybrid-sku-recommender

⭐ 1 stars | Python

Explainable hybrid alternative-SKU recommendation with business constraints

machine-learning python ranking recommender-system

15. Prajwaln08/NeurIPS-2024-BELKA

⭐ 1 stars | Unknown

🧬 Predicting small molecule–protein binding with ML — my Kaggle submission to the NeurIPS 2024 BELKA competition (EDA, LightGBM, XGBoost, Random Forest, Optuna, ensembling)

cheminformatics data-science drug-discovery ensemble-learning healthcare kaggle-competition

Generated by Research Pulse on 2026-08-13 06:07