2026-06-16·Daily digest

🔬 Research Pulse

Daily Digest

June 16, 2026

🤖 AI

🧠 LLMs

1. TokenPilot: Cache-Efficient Context Management for LLM Agents

Authors: Buqiang Xu, Zirui Xue, Dianmou Chen... Published: 2026-06-15 | Citations: 0 arXiv | PDF

Research Question: How can LLM agents in long-horizon sessions reduce context-driven inference costs without breaking prompt cache continuity, given that existing pruning/eviction methods cause prefix mismatches and cache invalidation?

Summary: TokenPilot is a dual-granularity context management framework for LLM agents that stabilizes prompt prefixes at ingestion and defers segment eviction until task relevance expires, preserving prompt-cache hits. It achieves 56–87% cost reductions on PinchBench and Claw-Eval while keeping task quality competitive, and ships inside LightMem2.

Key Results: On PinchBench and Claw-Eval benchmarks, TokenPilot reduced costs by 61% and 56% in isolated mode, and 61% and 87% in continuous mode, while maintaining competitive task performance versus prior context-management systems.

Key Findings:

Unconstrained context pruning/eviction destroys prefix stability, invalidating prompt caches and undermining apparent token savings
Separating concerns into ingestion-time compaction and lifecycle-aware eviction preserves cache continuity while still reducing context size
Conservative batch-turn eviction tied to residual utility outperforms aggressive eviction on cost-performance trade-offs in both isolated and continuous agent modes

Technical Novelty: Explicitly framing the cache-continuity vs text-sparsity trade-off and addressing it with dual-granularity controls: a global ingestion gate that stabilizes prompt prefixes (rather than mutating them post-hoc) plus a local conservative batch-turn eviction schedule that only offloads segments once task relevance expires.

What's New: Prior work optimized for token count alone; TokenPilot is the first to treat prompt-cache continuity as a first-class constraint and co-design ingestion compaction with lifecycle-aware eviction around it.

Extension Opportunities:

Apply Ingestion-Aware Compaction to multi-agent orchestration frameworks where shared prefixes across agents could amplify cache reuse benefits
Extend Lifecycle-Aware Eviction with learned utility predictors (small classifier or LM) instead of heuristic batch-turn schedules to improve eviction precision
Integrate TokenPilot with retrieval-augmented memory stores (e.g., vector DB-backed long-term memory) to test cache-aware retrieval injection points

Replicability: Integrated into LightMem2 (open source at https://github.com/zjunlp/LightMem2). Benchmarks PinchBench and Claw-Eval are referenced. Compute needs not specified, but as an inference-side context manager it should be reproducible on modest GPU/API budgets typical of agent benchmarking.

Research Gaps:

No reported evaluation across heterogeneous model providers with differing cache semantics (e.g., Anthropic vs OpenAI vs open-source vLLM)
Eviction schedule appears heuristic; lacks a learned or adaptive policy benchmarked against the conservative batch-turn rule

2. Geometric Action Model for Robot Policy Learning

Authors: Jisang Han, Seonghu Jeon, Jaewoo Jung... Published: 2026-06-15 | Citations: 0 arXiv | PDF

Research Question: How can robot manipulation policies explicitly leverage 3D geometric priors for contact-rich tasks, rather than relying on 2D image frames or 2D-derived latent spaces as in current VLAs and video world-action models?

Summary: GAM is a language-conditioned manipulation policy that repurposes a pretrained geometric foundation model as a unified backbone for perception, temporal world modeling, and action decoding by splitting it and inserting a causal future-latent predictor. This minimal architectural change injects 3D geometric priors into VLA-style policies, yielding gains in accuracy, robustness, speed, and model size over current foundation-model-scale baselines.

Key Results: The abstract claims GAM outperforms foundation-model-scale baselines across a 'broad suite' of simulation and real-robot manipulation benchmarks on four axes: accuracy, robustness, speed, and parameter efficiency (lighter). However, no specific numerical results, benchmark names (e.g., LIBERO, CALVIN, RLBench), or quantitative deltas are provided in the abstract itself.

Key Findings:

A single geometric foundation model can be split and reused as both observation encoder and feature decoder, with a causal predictor in the middle producing future latents that double as action substrate
Language-conditioned temporal world modeling can be added to a GFM with minimal architectural modification while preserving its geometric priors
GAM is reported as simultaneously more accurate, more robust, faster, and lighter than foundation-model-scale VLA/WAM baselines across simulation and real-robot benchmarks

Technical Novelty: Splitting a pretrained geometric foundation model at an intermediate layer and inserting a causal future-token predictor at the split — using shallow GFM layers as a perception encoder and the deeper layers as a shared decoder for both predicted future geometry and actions. Prior VLAs/WAMs bolt action heads onto 2D vision-language backbones; GAM repurposes a 3D-aware backbone as the substrate for perception, world modeling, and action all at once.

What's New: Most prior VLAs (e.g., RT-2, OpenVLA, π0) build on 2D vision-language backbones; video world-action models add temporal priors but still in 2D latent space. GAM is distinctive in (a) using a geometric foundation model as the substrate and (b) the split-and-insert design that makes the same backbone produce both future geometry and actions.

Extension Opportunities:

Swap the geometric foundation model backbone (e.g., from a depth/3D-token GFM like VGGT or DUSt3R-style models) to test which geometric priors transfer best to manipulation
Apply the split-backbone-with-causal-predictor pattern to bimanual or mobile manipulation where multi-view geometric reasoning matters even more
Extend the future latent token predictor to longer horizons or hierarchical planning, enabling sub-goal decomposition while reusing the same GFM substrate

Replicability: The abstract does not mention code, weights, or dataset release. Reproduction would likely require a pretrained geometric foundation model (e.g., DUSt3R/VGGT-class, 100M–1B params), multi-GPU finetuning (likely 8×A100/H100-scale), and access to standard simulation benchmarks plus a real robot setup for the physical evaluations.

Research Gaps:

No quantitative results or specific benchmark identifiers in the abstract, making the magnitude of improvements unclear
Choice of split layer, predictor design, and sensitivity to the underlying GFM are not characterized — generality across different geometric backbones is unproven from the abstract alone

3. Exploring Extrinsic and Intrinsic Properties for Effective Reasoning with Code Interpreter

Authors: Patomporn Payoungkhamdee, Napat Laosaengpha, Jenta Wonglertsakul... Published: 2026-06-15 | Citations: 0 arXiv | PDF

Research Question: What behavioral properties (extrinsic tokens and intrinsic cognitive behaviors) underlie effective Code Interpreter (CI) reasoning in LLMs, and how can they be leveraged at inference and training time?

Summary: The paper characterizes what makes Code Interpreter reasoning effective by analyzing extrinsic crucial tokens and intrinsic cognitive behaviors (verification, backtracking, backward chaining). It shows these properties correlate with stronger CI models and can be leveraged at inference (prompt injection) and training (SFT/RL augmentation) to improve reasoning, reduce overthinking, and boost token efficiency.

Key Results: Across multiple LLMs, stronger CI reasoning models showed higher prevalence of crucial tokens and cognitive behaviors (verification, backtracking, backward chaining). Appending code-specific crucial tokens at inference improved math, ordering, and optimization tasks. Augmenting SFT and RL with code-specific cognitive behaviors improved performance in 2 of 3 evaluated models, reduced overthinking on incorrect responses, and improved token efficiency.

Key Findings:

Stronger CI reasoning models exhibit more crucial tokens and the three cognitive behaviors
Inference-time appending of code-specific crucial tokens helps math, ordering, and optimization but offers limited gains elsewhere
Training-time augmentation with cognitive behaviors improves SFT and RL in 2 of 3 models, reducing overthinking on wrong answers and improving token efficiency

Technical Novelty: First systematic decomposition of CI reasoning into extrinsic (crucial tokens) vs intrinsic (cognitive behaviors) properties, transferring the natural-language-reasoning analysis framework to code-execution settings, and using these as both inference-time prompts and training-time augmentations.

What's New: Unlike prior work on natural-language reasoning behaviors or generic CI benchmarking, this paper is the first to systematically separate and measure extrinsic vs intrinsic code-reasoning properties and validate their causal usefulness in both inference and training regimes.

Extension Opportunities:

Build a CI-aware decoding wrapper that dynamically injects crucial tokens only on task types (math/ordering/optimization) where they help, avoiding regressions elsewhere
Construct a training dataset auto-labeled with the three cognitive behaviors (verification, backtracking, backward chaining) to fine-tune smaller open models for CI reasoning
Investigate the model-specific factors that prevented gains in the third model — likely a path to a router that picks behavior-augmented vs vanilla training per architecture

Replicability: Abstract does not mention released code/data. Reproduction would require multiple base LLMs plus SFT and RL pipelines with a Code Interpreter sandbox — moderate-to-high compute (likely multi-GPU for RL on reasoning benchmarks).

Research Gaps:

Why the third model failed to benefit from behavior augmentation remains only partially explained
Limited benefit of crucial-token injection outside math/ordering/optimization suggests an unmapped taxonomy of when these properties transfer

🦾 ROBOTICS

1. T-Rex: Tactile-Reactive Dexterous Manipulation

Authors: Dantong Niu, Zhuoyang Liu, Zekai Wang... Published: 2026-06-15 | Citations: 0 arXiv | PDF

Research Question: How can Vision-Language-Action (VLA) models for robotic manipulation effectively integrate high-frequency tactile signals for reactive dexterous manipulation, given that current VLAs either ignore tactile modality or use only static encoders due to data scarcity, architectural constraints, and encoder limitations?

Summary: T-Rex tackles the underuse of tactile signals in VLA models by combining a 100-hour primitive-focused tactile dataset, a variable-rate Mixture-of-Transformers architecture, and a temporal tactile VQ-VAE encoder. The system achieves >30% higher success rate over the strongest baseline across 12 force-sensitive and deformable-object manipulation tasks.

Key Results: Demonstrated tactile-reactive policies on 12 manipulation tasks requiring delicate force control and deformable object manipulation, achieving over 30% higher average success rate than the strongest baseline. Collected a 100-hour tactile-rich dataset via a data-efficient recipe prioritizing elementary motor primitives.

Key Findings:

A variable-rate MoT can preserve existing VLA capabilities while integrating naturally high-frequency tactile streams without forcing a uniform sample rate
Temporal (dynamic) tactile encoding via VQ-VAE outperforms static tactile encoders prevalent in prior work
Collecting demonstrations around elementary motor primitives is more data-efficient for tactile-reactive learning than task-centric collection, enabling a tractable 100-hour corpus

Technical Novelty: Introduces a variable-rate Mixture-of-Transformers (MoT) architecture that handles asynchronous modality rates (high-frequency tactile vs. lower-frequency vision/language), paired with a novel temporal tactile VQ-VAE encoder that captures dynamic touch signals rather than static cues. The data collection recipe centered on elementary motor primitives is also a methodological contribution distinct from task-centric demonstration collection.

What's New: Unlike prior VLAs that drop tactile input or use static-frame encoders, T-Rex jointly addresses dataset scale, architecture, and encoder design — introducing the first variable-rate MoT for asynchronous tactile-vision-language fusion and a temporal VQ-VAE tailored to touch dynamics.

Extension Opportunities:

Extend the variable-rate MoT architecture to incorporate additional high-frequency modalities (e.g., audio, proprioceptive force-torque) alongside tactile signals for richer multi-sensory manipulation
Apply the temporal tactile VQ-VAE encoder to bimanual or whole-body manipulation tasks, particularly contact-rich assembly or in-hand reorientation of irregular objects
Develop sim-to-real transfer pipelines using the 100-hour dataset to bootstrap tactile-reactive policies in simulation, reducing reliance on expensive real-world tactile data collection

Replicability: Abstract does not mention code/data release. Reproducing would require: (1) tactile sensor hardware on a dexterous manipulator, (2) ~100 hours of teleoperated data collection infrastructure, (3) significant GPU compute to train a VLA-scale MoT model with VQ-VAE — likely multi-GPU clusters for days to weeks.

Research Gaps:

No standardized benchmark or evaluation protocol for tactile-reactive dexterous manipulation across labs/sensors
Generalization across heterogeneous tactile sensor hardware (different resolutions, modalities) remains unaddressed by the proposed encoder

2. R2RDreamer: 3D-aware Data Augmentation for Spatially-generalized 2D Manipulation Policies

Authors: Xiuwei Xu, Haowen Sun, Angyuan Ma... Published: 2026-06-15 | Citations: 0 arXiv | PDF

Research Question: How can imitation-learned 2D RGB manipulation policies achieve spatial generalization across diverse object poses, robot configurations, and camera viewpoints from limited source demonstrations, without the sim-to-real gap of simulation augmentation or the 3D-policy bias of prior real-to-real methods?

Summary: R2RDreamer is a real-to-real demonstration augmentation framework that boosts spatial generalization of RGB-based manipulation policies by editing object pointclouds and end-effector trajectories in 3D, then completing the resulting masked, occlusion-aware control videos with a dense-control image-to-video model. This hybrid 3D-edit/2D-complete design avoids the heavy 3D scene parsing of prior real-to-real methods and produces temporally coherent RGB observations directly consumable by 2D diffusion and VLA policies.

Key Results: R2RDreamer demonstrates improved spatial generalization on spatially shifted manipulation tasks for both 2D diffusion-style policies and vision-language-action (VLA) policies starting from limited source demonstrations. Ablation analyses validate the individual contributions of (1) lightweight 3D editing of pointclouds/trajectories, (2) occlusion-aware projection into masked control videos, and (3) dense-control image-to-video completion. Specific numerical benchmarks, success rates, or dataset sizes are not disclosed in the abstract.

Key Findings:

Splitting augmentation into lightweight 3D action-observation editing and 2D video-space visual completion preserves geometric consistency while removing the need for full 3D scene/geometry completion.
Occlusion-aware projection of edited 3D scenes into masked control videos is important for the downstream image-to-video model to produce usable RGB observations.
Both 2D diffusion-style policies and vision-language-action policies benefit from R2RDreamer augmentations when generalizing from limited source demonstrations to spatially shifted tasks.

Technical Novelty: The hybrid split between geometry and appearance: prior real-to-real augmentation methods perform full 3D scene parsing and geometry completion to produce pointcloud-policy-friendly observations, while R2RDreamer keeps 3D work lightweight (editing incomplete pointclouds + EE trajectories in a shared frame) and pushes visual completion into 2D video space via occlusion-aware masked control videos fed to a dense-control image-to-video model — yielding temporally coherent RGB suitable for 2D/VLA policies.

What's New: Unlike prior real-to-real augmentation methods that target 3D pointcloud policies via heavy 3D parsing and geometry completion, R2RDreamer is explicitly designed for 2D RGB and VLA policies by offloading visual completion to a dense-control image-to-video model while keeping the 3D step minimal and geometry-consistent.

Extension Opportunities:

Extend the dense-control I2V completion module to multi-camera/multi-view synchronized generation so policies can be trained for setups with several RGB streams simultaneously.
Couple R2RDreamer with an active demonstration-selection loop that identifies under-covered spatial regions in policy behavior and synthesizes targeted augmentations to close those gaps.
Adapt the 3D-edit + 2D-complete pipeline to deformable or articulated objects (cloth, hinged objects), where current pointcloud editing assumptions of rigid transforms break down.

Replicability: The abstract does not mention a code or data release. Reproduction would require: a real-robot demonstration dataset, RGB-D capture, a dense-control image-to-video diffusion model (likely tens of GB VRAM, multi-GPU for training/finetuning), and downstream training of 2D diffusion policies and a VLA — overall a substantial multi-GPU compute footprint typical of recent video-diffusion + robot-learning pipelines.

Research Gaps:

Abstract does not quantify how the approach scales as the spatial distribution shift grows large (e.g., extreme viewpoint or robot-base changes).
It is unclear how the method handles dynamic scenes, multi-object interactions, or non-rigid objects, since 3D editing is described in terms of object pointclouds and EE trajectories.

💻 COMPUTE

1. 3D Ising criticality with Platonic lattice superconducting qubits

Authors: Liyang Sui, Hong-Hao Song, Sainan Huai... Published: 2026-06-15 | Citations: 0 arXiv | PDF

Research Question: How can 3D Ising critical exponents be directly extracted via quantum simulation, given the connectivity, system-size, and spectral-resolution barriers that have made direct quantum study of 3D criticality intractable compared to classical Monte Carlo and conformal bootstrap?

Summary: The authors perform a digital quantum simulation of the 3D Ising critical point on a 9-qubit superconducting processor built with a Platonic (cubic) lattice connectivity, using an extended VQE with a phase-based loss to prepare low-energy eigenstates of the transverse-field Ising model. Via state-operator correspondence and Fourier-transform spectroscopy, they extract Δ_ε = 1.5850 and ν = 0.7067, agreeing with exact diagonalization to ±0.001 and matching established theoretical values.

Key Results: Using a 9-qubit superconducting processor with Platonic (cubic) lattice connectivity and an 8-qubit register, the authors variationally prepared low-energy eigenstates of the transverse-field Ising model via an extended VQE with a phase-based loss. Fourier-transform spectroscopy of the four lowest eigenenergies agreed with exact diagonalization to within ±0.001, yielding scaling dimension Δ_ε = 1.5850 and critical exponent ν = 0.7067, consistent with conformal bootstrap values.

Key Findings:

Platonic-lattice qubit connectivity enables direct encoding of 3D Ising geometry on a superconducting chip with only 8 data qubits
Phase-based VQE loss + Fourier analysis resolves the four lowest eigenenergies to ±0.001 vs exact diagonalization
Extracted Δ_ε = 1.5850 and ν = 0.7067 are quantitatively consistent with conformal bootstrap and Monte Carlo benchmarks

Technical Novelty: The combination of (1) a multiply-connected Platonic (cubic) lattice qubit topology that natively realizes the 3D Ising connectivity on a flat chip, (2) the state-operator correspondence of CFT to map bulk critical scaling dimensions onto eigenenergies of a finite quantum register, and (3) an extended VQE with a phase-based loss function plus Fourier-transform spectral analysis to resolve multiple low-lying eigenstates simultaneously.

What's New: Prior quantum simulations of Ising criticality have been largely confined to 1D/2D due to connectivity limits; this work is among the first to realize a 3D critical point on superconducting hardware by engineering a Platonic-solid lattice topology and pairing it with a CFT-informed eigenstate-extraction protocol.

Extension Opportunities:

Scale the Platonic lattice geometry to larger qubit registers (e.g., 27 or 64 qubits) to extract higher-precision exponents and additional operator content (Δ_σ, Δ_T) beyond the four lowest eigenenergies
Apply the same state-operator-correspondence + phase-based VQE pipeline to other 3D CFTs (O(N) models, percolation, gauge theories) where bootstrap data exist for cross-validation
Port the multiply-connected Platonic encoding to neutral-atom or trapped-ion platforms with native long-range connectivity to test hardware-agnostic universality of the protocol

Replicability: The abstract does not mention released code or datasets. Reproducing the experiment requires a superconducting processor supporting Platonic/cubic-graph connectivity on 9 qubits, classical optimization for VQE, and Fourier post-processing — classical exact-diagonalization cross-checks for 8 qubits are trivial (<1 GB RAM). Hardware access is the primary barrier.

Research Gaps:

System size is still small (8 data qubits) — finite-size scaling corrections to extracted exponents are not addressed in the abstract
The approach relies on variational state preparation whose scalability and noise resilience at larger Platonic lattices remains unverified

2. Quantum Nonlocal Games on Graph Ensembles

Authors: Joshua Tucker, Chris Weeks, Peter Drmota... Published: 2026-06-15 | Citations: 0 arXiv | PDF

Research Question: Can quantum entanglement provide a provable advantage for spatial coordination tasks (like rendezvous) between mobile agents operating under realistic conditions of topographical uncertainty and across physically separated quantum devices, rather than only in idealized single-processor simulations?

Summary: The authors generalize quantum nonlocal coordination games to settings where agents have only probabilistic knowledge of the network topology (graph ensembles), and demonstrate a rendezvous quantum advantage experimentally using remote entanglement between two physically separated ion-trap systems. They also show — via superconducting hardware simulations — the surprising effect that increasing local information available to players amplifies the quantum advantage, contrary to classical intuition.

Key Results: The paper (1) develops a theoretical framework for quantum nonlocal games on graph ensembles that models topographical uncertainty, (2) experimentally demonstrates quantum advantage in rendezvous between two physically separated ion-trap systems linked via remote entanglement, (3) simulates a broader class of coordination problems on superconducting hardware, and (4) shows the counter-intuitive result that giving players more local information increases the quantum advantage — a behavior impossible classically.

Key Findings:

Quantum advantage in rendezvous survives realistic topographical uncertainty modeled via graph ensembles, not just fully-known graphs
First demonstration of nonlocal coordination advantage between two physically separated quantum devices linked by remote entanglement (ion traps), not co-located qubits
More local information given to players increases the quantum-over-classical gap — a monotonicity classical strategies cannot match
A broader family of coordination problems was simulated on superconducting hardware, suggesting generality beyond rendezvous

Technical Novelty: Two contributions: (a) formal theory of nonlocal games defined over ensembles of graphs rather than a single known graph, capturing agents' uncertainty about the underlying topology; (b) first cross-device (separated ion-trap) experimental demonstration of quantum coordination advantage using genuine remote entanglement, as opposed to qubits co-located in one processor.

What's New: Prior demonstrations assumed full topology knowledge and used a single quantum processor as a stand-in for separated agents. This work removes both idealizations: it formalizes uncertainty via graph ensembles and uses genuinely separated ion-trap nodes connected by remote entanglement, putting the experiment closer to a true distributed quantum protocol.

Extension Opportunities:

Scale the graph ensemble framework to larger, dynamic networks (e.g., time-varying graphs or stochastic mobility models) and benchmark quantum-vs-classical advantage as topology evolves
Build a portable quantum coordination demo using networked NV-centers or photonic entanglement distribution to test rendezvous in field conditions outside the lab
Extend the formalism beyond pairwise rendezvous to multi-agent coalition formation, swarm consensus, or distributed task allocation under uncertainty

Replicability: No explicit code/data link mentioned in the abstract. Reproduction would require either (a) access to a networked ion-trap system with remote entanglement distribution (a handful of labs worldwide, e.g., Oxford, Innsbruck, Delft), or (b) for the broader sweep, a superconducting backend (IBM/Google-class processors) — accessible via cloud. The graph-ensemble theory and classical bounds should be reproducible analytically/numerically on commodity hardware.

Research Gaps:

Scalability beyond two players and small graphs — distributed entanglement budgets, decoherence, and entanglement-distribution rates likely bound near-term practicality
No discussion (in the abstract) of robustness to noise, loss, or imperfect entanglement fidelity in the remote-entanglement link, which would govern real-world deployment

3. Embedded Arena: Iterative Optimization via Hardware Feedback

Authors: Zhihan Zhang, Alexander Le Metzger, Jiuyang Lyu... Published: 2026-06-15 | Citations: 0 arXiv | PDF

Research Question: Can an LLM agent autonomously co-optimize ML models and firmware for heterogeneous microcontrollers under hard physical constraints (memory, power, temperature) while preserving accuracy — a task currently requiring manual expert tuning?

Summary: The paper introduces a hardware-in-the-loop agent arena where LLM agents iteratively compile, flash, and measure on real microcontrollers to co-optimize ML models and firmware under hard memory/power/thermal constraints. With physical feedback, agents achieve 250x vision and 400x audio compression and surpass human experts within seven iterations, demonstrated on a battery-free elk-detection trap and a phonetic-transcription wearable.

Key Results: Frontier models (Claude Opus 4.7, Gemini 3.1 Pro) achieve 0% deployment success without hardware feedback. With hardware-in-the-loop: first successful deployment within 3 iterations, surpasses human experts within 7. Achieves 250x compression on vision models with <3.3% accuracy loss, 400x on audio with <6% FER loss. Real deployments: elk-detection camera trap at 96.7% accuracy and phonetic-transcription wearable at 8.44% FER, enabling battery-free solar-harvested operation on a commercial MCU.

Key Findings:

Without hardware feedback, frontier LLMs (Claude Opus 4.7, Gemini 3.1 Pro) have 0% deployment success — closed-loop physical measurement is the critical enabler
Agentic co-optimization yields 250x vision compression with <3.3% accuracy loss and 400x audio compression with <6% FER loss, enabling solar-powered battery-free operation
Real-world deployments validate the approach: 96.7% accuracy elk-detection camera trap and 8.44% FER child-speech phonetic transcription wearable

Technical Novelty: A hardware-in-the-loop agent arena that closes the optimization loop by physically compiling, flashing, and measuring on real MCUs each iteration — moving beyond simulator/proxy-based NAS or quantization search. Novel framing of joint model+firmware co-optimization as an LLM agentic task with grounded physical reward signals.

What's New: First framing of embedded model+firmware co-design as an LLM agent task grounded in real-hardware feedback rather than simulators or proxies, with empirical evidence that physical signals are necessary (not just helpful) for frontier models to succeed at MCU deployment.

Extension Opportunities:

Extend the arena to additional MCU families (RISC-V, ESP32, Nordic) and add NPU/accelerator targets to test cross-architecture generalization of the agent's optimization strategies
Add multi-objective Pareto-front exploration (energy vs accuracy vs latency) instead of single-trajectory iteration, letting the agent propose deployment portfolios rather than one model
Apply the closed-loop hardware feedback paradigm to other physical-constraint domains (FPGA synthesis, ASIC PPA tuning, robotics controller co-design) where simulation-only loops similarly fail

Replicability: Abstract does not mention code release. Reproduction requires: target MCUs, flashing/measurement instrumentation (power meters, thermal probes), solar harvesting setup for battery-free experiments, and API access to frontier models (Claude Opus 4.7, Gemini 3.1 Pro). Compute cost is modest on the LLM side but hardware lab setup is the gating factor.

Research Gaps:

Generalization across MCU architectures and toolchains beyond the studied targets is unproven
No analysis of cost/wall-clock budget per optimization run, or whether the iterative loop converges reliably across task types beyond vision and audio

⚡ ENERGY

1. Dynamically suppressed lattice rotations in SrTiO$_3$ as a basis for photo-induced ferroelectricity

Authors: Huaiyu Hugo Wang, Michael Fechner, Giovanni De Vecchi... Published: 2026-06-15 | Citations: 0 arXiv | PDF

Research Question: How do antiferrodistortive (AFD) rotations in the low-temperature tetragonal phase of SrTiO₃ couple with coherently driven Ti-O stretching vibrations to enable photo-induced ferroelectricity, and can a single unified theory describe the nonlinear lattice dynamics across both cubic and tetragonal phases?

Summary: The paper extends the understanding of photo-induced ferroelectricity in SrTiO₃ by measuring, via time-resolved elastic X-ray diffraction, a long-lived suppression of the static antiferrodistortive rotation angle in the low-temperature tetragonal phase after resonant mid-IR pumping of the Ti-O stretching mode. A unified first-principles nonlinear lattice-dynamics theory reproduces both this tetragonal-phase angle reduction and the previously observed cubic-phase AFD fluctuation suppression, providing a coherent microscopic basis for light-induced ferroelectricity across the phase diagram.

Key Results: Using time-resolved elastic X-ray diffraction in the low-temperature tetragonal phase, the authors observe a long-lived reduction of the equilibrium AFD rotation angle following mid-IR pumping of the Ti-O stretching mode. They demonstrate that a unified theory built from first-principles (DFT) calculations of nonlinear phonon-phonon coupling reproduces the dynamics in BOTH the high-T cubic phase (where AFD fluctuations are suppressed) and the low-T tetragonal phase (where the static AFD rotation angle is reduced), unifying prior diffuse-scattering observations with new elastic-scattering measurements.

Key Findings:

Resonant mid-IR driving of the Ti-O stretching vibration produces a long-lived reduction of the equilibrium AFD rotation angle in tetragonal SrTiO₃
The same nonlinear phonon-phonon coupling explains both fluctuation suppression (cubic phase) and static-angle reduction (tetragonal phase)
First-principles calculations quantitatively reproduce experimental dynamics across both phases without phenomenological tuning, supporting AFD-mode softening as the microscopic gateway to photo-ferroelectricity

Technical Novelty: Prior work had shown AFD fluctuation suppression correlates with photo-ferroelectricity in the CUBIC phase via diffuse scattering. This paper is the first to measure the complementary signature — a reduction of the STATIC AFD rotation angle — in the tetragonal phase via elastic X-ray diffraction, and crucially provides a single first-principles nonlinear-lattice-dynamics theory that quantitatively bridges both regimes (fluctuation suppression vs. equilibrium-angle reduction).

What's New: First time-resolved elastic XRD evidence of static AFD-angle suppression in the tetragonal phase, combined with a unified ab initio nonlinear-phonon theory that simultaneously explains cubic-phase fluctuation suppression and tetragonal-phase equilibrium suppression — closing a gap between two previously disconnected experimental regimes.

Extension Opportunities:

Apply the unified nonlinear-phonon framework to other quantum paraelectrics (KTaO₃, EuTiO₃) or hybrid improper ferroelectrics to predict photo-induced polar phases and validate via time-resolved XRD
Build a real-time TDDFT + nonlinear-phonon simulation pipeline that takes a pump pulse waveform as input and outputs predicted diffuse/elastic scattering signals — useful for designing pump protocols at XFEL beamlines
Couple this mechanism to thin-film heterostructure engineering: epitaxial strain tuning of the AFD/polar mode hierarchy to maximize the photo-induced polarization lifetime, with potential for ultrafast non-volatile memory prototypes

Replicability: No mention of code/data release in the abstract. Reproduction requires: (1) an XFEL or synchrotron with fs time-resolved elastic XRD capability and a mid-IR pump source resonant with the Ti-O stretching mode (~17 THz); (2) high-quality SrTiO₃ single crystals cooled below ~105 K into the tetragonal phase; (3) DFT/DFPT compute for the first-principles nonlinear phonon couplings (modest — tractable on a small HPC cluster, ~thousands of CPU-hours).

Research Gaps:

Direct measurement of the transient polar mode amplitude/symmetry in the tetragonal phase is not addressed — the link from AFD suppression to ferroelectric order parameter remains inferred rather than directly observed
Domain-level heterogeneity (tetragonal twins) and the role of disorder/defects in the photo-induced state are not characterized

2. Orbital-selective band evolution and out-of-plane correlation in the FeGe-family kagome antiferromagnet ScFe$_6$Ge$_6$

Authors: Jae Hyuck Lee, Ze Yan, Tongrui Li... Published: 2026-06-15 | Citations: 0 arXiv | PDF

Research Question: What conditions enable spin-correlated charge density wave (CDW) order in magnetic kagome metals, and what role do orbital selectivity and out-of-plane correlations play? Specifically, why does FeGe exhibit CDW order within an antiferromagnetic state while the structurally similar ScFe6Ge6 does not?

Summary: The paper investigates ScFe6Ge6, a FeGe-family kagome antiferromagnet lacking CDW order, to isolate the magnetic contribution to band reconstruction across a 195 K transition. Using ARPES and Raman spectroscopy, they identify orbital-selective evolution of a single kagome Dirac band coupled to an out-of-plane phonon, suggesting these ingredients are essential precursors for the unconventional CDW seen in FeGe.

Key Results: Using ARPES and Raman spectroscopy on ScFe6Ge6, the authors demonstrated: (1) a magnetic transition at T* = 195 K (in contrast to FeGe's CDW-AFM coexistence), (2) orbital-selective band evolution confined to a single kagome Dirac band across this transition, and (3) electron-phonon and magnetoelastic coupling to a specific out-of-plane phonon mode. The absence of CDW order in ScFe6Ge6 isolates the magnetic-only contribution, providing a controlled comparison to FeGe.

Key Findings:

ScFe6Ge6 hosts a magnetic transition at T* = 195 K without accompanying CDW order, contrasting FeGe
Band reconstruction across T* is orbital-selective, affecting only a single kagome Dirac band rather than the full electronic structure
Electron-phonon and magnetoelastic coupling specifically targets an out-of-plane phonon mode, highlighting 3D correlations in nominally 2D kagome physics

Technical Novelty: First combined ARPES + Raman investigation of ScFe6Ge6 as a 'CDW-free control' for FeGe, isolating orbital-selective magnetic effects on a single kagome Dirac band. The identification of magnetoelastic coupling to an out-of-plane phonon mode in a kagome antiferromagnet without CDW order is new, framing out-of-plane correlations as a key ingredient for spin-correlated CDW formation.

What's New: Unlike most kagome studies focused on in-plane physics, this work emphasizes out-of-plane correlations as essential to spin-correlated CDW order. Using a CDW-absent compound as a control to isolate orbital-selective magnetic effects is a novel comparative strategy in the FeGe-family kagome literature.

Extension Opportunities:

Perform analogous ARPES/Raman studies on other RFe6Ge6 family members (R = Y, Lu, rare earths) to map out a systematic phase diagram correlating R-site chemistry with CDW emergence
Apply uniaxial strain or pressure to ScFe6Ge6 to test whether enhancing out-of-plane coupling can induce CDW order, directly testing the paper's hypothesis
Develop DFT+DMFT calculations with orbital-resolved electron-phonon coupling to model the selective Dirac band renormalization and predict CDW instability conditions in related kagome magnets

Replicability: No code/data availability mentioned in the abstract. Reproduction requires high-quality ScFe6Ge6 single crystals, synchrotron ARPES beamline access (likely with polarization control for orbital selectivity), and a low-temperature Raman setup capable of sub-200 K measurements. Significant experimental infrastructure rather than computational resources.

Research Gaps:

Mechanism connecting orbital-selective Dirac band reconstruction to the lattice symmetry breaking specific to FeGe remains unresolved
Quantitative theoretical framework predicting which kagome antiferromagnets will develop CDW order based on orbital character and out-of-plane coupling strength is still absent

3. Counterdiabatic Raman Atom Optics for Compact High-Sensitivity Gravimetry

Authors: Asad Ali, Hamid Arian Zad, Saif Al-Kuwari... Published: 2026-06-15 | Citations: 0 arXiv | PDF

Research Question: How can large-momentum-transfer (LMT) atom interferometry scale to high momentum orders without accumulating pulse-transfer errors across long Raman pulse sequences, which currently limits compact quantum gravimeters?

Summary: The paper proposes embedding counterdiabatic (shortcut-to-adiabaticity) corrections directly into stimulated Raman pulse envelopes (STIRSAP) to enable high-fidelity large-momentum-transfer atom optics for compact gravimeters. Simulations show 1 μs pulses with F_π = 0.99902 fidelity and identify a shot-noise-optimal momentum order n ≈ 270, with practical limits set by wave-packet separation and noise rather than pulse speed.

Key Results: Numerical simulations of an effective Raman model show 1 μs STIRSAP pulses achieve single-pulse π-transfer fidelity F_π = 0.99902 with negligible pulse-time overhead at high momentum order. The analysis identifies an unconstrained shot-noise optimum near n ≈ 270 photon recoils, and shows extreme-LMT operation is bounded by wave-packet separation, vibration noise, Doppler detuning, and systematics — not by pulse duration.

Key Findings:

1 μs STIRSAP Raman pulses reach single-pulse fidelity F_π = 0.99902 with no significant time overhead vs. conventional pulses
Counterdiabatic correction can be folded into the Raman envelope itself, eliminating auxiliary microwave/RF control hardware
An unconstrained shot-noise optimum exists near n ≈ 270 photon recoils, balancing phase enhancement against compound contrast decay
At extreme LMT orders, the binding constraints are wave-packet separation, vibration noise, Doppler detuning, and systematics — not pulse duration

Technical Novelty: Encoding the counterdiabatic correction term directly into the amplitude/phase envelopes of the two Raman lasers — removing the need for auxiliary microwave or RF control fields that prior shortcut-to-adiabaticity schemes required — and applying it specifically to scalable Mach–Zehnder LMT sequences.

What's New: Prior shortcut-to-adiabaticity schemes for Raman transitions typically required extra microwave/RF control fields to realize the counterdiabatic Hamiltonian; this work shows the correction can be absorbed entirely into the two Raman laser envelopes, making it directly compatible with existing compact gravimeter hardware and scalable Mach–Zehnder LMT sequences.

Extension Opportunities:

Experimentally implement STIRSAP-shaped Raman envelopes on an existing Rb/Cs cold-atom gravimeter and benchmark π-pulse fidelity vs. standard square/Gaussian Raman pulses at n = 10–50
Combine STIRSAP envelopes with vibration-isolation feedback or hybrid classical-accelerometer co-sensing to push usable n beyond the wave-packet-separation limit identified in the paper
Extend the counterdiabatic encoding from Raman to single-photon Bragg or double-Bragg atom optics for clock-state interferometers, and to multi-species/dual-isotope geometries for equivalence-principle tests

Replicability: No code or data release is mentioned in the abstract. The work is purely theoretical/numerical based on an effective Raman two-photon model; reproduction requires only a standard scientific Python/Matlab stack to integrate the Schrödinger equation for a multilevel Raman system — minimal compute (laptop-scale).

Research Gaps:

No experimental validation — results are simulation-only on an effective two-photon Raman model, leaving open how well the predicted fidelities survive spontaneous emission, laser phase noise, and intensity inhomogeneity in real hardware
The analysis identifies wave-packet separation and vibration noise as the true LMT-limiting factors but does not propose concrete mitigation strategies for reaching the n ≈ 270 optimum

🏥 HEALTHCARE

1. Latent space mapping of interpretable structural coordinates from stochastic single-molecule signals

Authors: Matteo Cartiglia, Sandro Kuppel, Wouter Botermans Wannes Peeters... Published: 2026-06-15 | Citations: 0 arXiv | PDF

Research Question: How can we extract reliable molecular identity from solid-state nanopore signals when stochastic translocation dynamics (variable speed, conformation, device drift) warp the time-domain signal and break alignment-based decoding of DNA barcodes?

Summary: The paper introduces a contrastive encoder, trained solely on physics-simulated nanopore signals, that embeds solid-state nanopore traces of engineered DNA barcodes into an interpretable latent space aligned with structural barcode parameters and invariant to translocation kinetics and device. This enables single-pass molecule identification ~1000× faster than alignment methods and supports mixture quantification, rare-variant detection, consensus reconstruction, and real-time analysis.

Key Results: A contrastive encoder trained only on physics-simulated signals maps experimental nanopore traces of engineered DNA barcodes into a latent coordinate system that (a) is responsive to structural barcode parameters yet invariant to acquisition conditions and translocation conformation, enabling cross-device pooling; (b) identifies molecules in a single forward pass, reducing compute ~1000× vs alignment-based methods; (c) was validated on mixture quantification, rare-variant detection, consensus barcode reconstruction, and real-time acquisition. Specific accuracy/throughput numbers are not given in the abstract.

Key Findings:

A contrastive encoder trained only on physics-informed simulations generalizes to real solid-state nanopore data without experimental labels
The latent representation disentangles structural barcode parameters from acquisition/translocation nuisance variables, allowing data pooling across devices
Single forward-pass identification cuts compute by three orders of magnitude vs alignment-based decoding and supports real-time use

Technical Novelty: Reframes stochastic-sensor analysis from temporal alignment to learned structural-coordinate mapping: a contrastive encoder trained purely on physics-informed simulations produces an interpretable latent space whose axes correspond to barcode design parameters and that is invariant to translocation kinetics — eliminating the need for experimentally labeled training data or per-event dynamic time warping.

What's New: Prior nanopore analysis relies on time-domain alignment (DTW, HMMs) or supervised classifiers trained on labeled experimental events. This work is the first to combine physics-based simulation with contrastive learning to produce an interpretable, invariance-engineered latent coordinate system for stochastic sensor signals.

Extension Opportunities:

Apply the sim-to-real contrastive latent mapping to biological nanopores (e.g. MinION) for direct nucleotide/methylation decoding without basecalling alignment
Extend the physics-informed simulator + encoder to protein nanopores or other stochastic single-molecule sensors (FETs, plasmonic, tweezers) to test paradigm generality
Build a closed-loop adaptive sampling system that uses the real-time encoder embedding to trigger selective sequencing or molecule recapture decisions

Replicability: Abstract does not mention code, data, or model release. Reproduction would require a solid-state nanopore rig, engineered DNA barcode library, and the physics-informed translocation simulator; encoder training is small-scale (contrastive on simulated 1D signals) and feasible on a single modern GPU.

Research Gaps:

Quantitative accuracy/sensitivity numbers and comparison baselines are not surfaced in the abstract
Generalization beyond engineered DNA barcodes to native biomolecules and to non-DNA analytes remains untested

2. FusionRS: A Large-Scale RGB-Infrared Remote Sensing Dataset for Dual-Modal Vision-Language Foundation Models

Authors: Jiaju Han, Ben Zhang, Xuemeng Sun... Published: 2026-06-15 | Citations: 0 arXiv | PDF

Research Question: How can vision-language foundation models for remote sensing be extended beyond RGB-only training to leverage infrared modality, given the absence of a large-scale aligned RGB-infrared-text dataset?

Summary: FusionRS introduces the first large-scale RGB-infrared-text dataset for remote sensing vision-language modeling, constructed via RGB-to-IR translation paired with both conventional and IR-aware captions. Foundation models trained on it show improved RGB-IR alignment, IR-to-text retrieval, and dual-modal captioning, with ablations proving modality-specific captions are essential for infrared-language grounding.

Key Results: The paper introduces FusionRS, the first large-scale RGB-infrared-text dataset for remote sensing, built by translating public RGB images into IR-style counterparts with paired conventional and IR-aware captions. Experiments demonstrate that training CLIP-style and generative VLMs on FusionRS improves RGB-IR alignment, infrared-to-text retrieval, and dual-modal captioning compared to RGB-only and non-IR-aware baselines. Ablations confirm that IR-aware captions (vs generic captions) are critical for infrared-language alignment.

Key Findings:

Synthetic RGB-to-IR translation can produce useful aligned dual-modal training data at scale when real paired RGB-IR datasets are unavailable
IR-aware captions that explicitly describe thermal/infrared visual cues outperform generic scene captions for infrared-language alignment
CLIP-style contrastive training and generative VLM fine-tuning both benefit from the dual-modal dataset, improving retrieval and captioning over RGB-only baselines

Technical Novelty: First dataset to pair RGB remote sensing imagery with IR-style counterparts AND modality-specific (IR-aware) textual captions explicitly describing thermal/infrared visual properties — moving beyond generic scene captions reused across modalities.

What's New: Prior remote sensing VLMs focus almost exclusively on RGB; existing IR datasets lack aligned text. FusionRS is the first to combine (1) large-scale RGB-IR pairing via translation and (2) modality-aware textual supervision tailored to infrared properties.

Extension Opportunities:

Replace synthetic RGB-to-IR translation with paired real-world thermal/IR satellite imagery (e.g., Landsat thermal bands, Sentinel-3) to reduce domain gap between synthetic and true infrared signatures
Extend the dual-modal framework to tri-modal learning by incorporating SAR or hyperspectral channels alongside RGB-IR, enabling broader Earth observation foundation models
Apply the IR-aware captioning approach to downstream tasks like nighttime object detection, wildfire monitoring, or thermal anomaly detection where illumination-invariance is critical

Replicability: Abstract does not mention public code/data release. Reproduction would require: a large RGB remote sensing corpus, an RGB-to-IR translation model (likely GAN or diffusion-based), caption generation pipeline (likely LLM-assisted), and multi-GPU compute for CLIP-style contrastive training plus VLM fine-tuning — estimated 8–64 GPUs for foundation-model-scale training.

Research Gaps:

Reliance on synthetically translated IR rather than authentic thermal sensor data may not capture true radiometric properties
No evaluation reported on real-world IR remote sensing downstream tasks (e.g., thermal target detection, nighttime scene understanding)

🔬 MATERIALS

1. Octave bandwidth 3D-Printed Couplers for Low-Loss Thin-Film Lithium Tantalate Circuits

Authors: Erik Jung, Xinyu Ma, Jan Brandes... Published: 2026-06-15 | Citations: 0 arXiv | PDF

Research Question: How can we achieve efficient, broadband, low-loss fiber-to-chip coupling for the emerging lithium tantalate on insulator (LTOI) photonic platform, which is needed to unlock its potential for high-speed electro-optic modulation, quantum squeezing, and second harmonic generation applications?

Summary: The paper presents the first 3D-printed polymer out-of-plane couplers for lithium tantalate on insulator (LTOI) photonic circuits, achieving sub-1.3 dB coupling loss and octave-spanning bandwidth from 850–1740 nm. Combined with sub-0.1 dB/cm waveguide propagation losses and 1 W power handling, this work establishes LTOI as a viable platform for broadband electro-optic modulation and quantum applications including second-harmonic-based squeezing.

Key Results: Demonstrated first 3D direct laser written polymer out-of-plane couplers for LTOI waveguides achieving: 0.9 dB/interface coupling loss (strip waveguides), 1.25 dB/interface (rib waveguides), 3 dB optical bandwidth spanning >1 octave (850–1740 nm), stable operation at 1 W optical input power, and on-chip propagation losses <0.1 dB/cm.

Key Findings:

Ultra-low coupling losses of 0.9 dB (strip) and 1.25 dB (rib) per interface using 3D laser-written polymer couplers
Octave-spanning 3 dB optical bandwidth (850–1740 nm) covers both telecom and visible/SH wavelengths in a single coupler
Stable operation under 1 W input power with on-chip propagation loss below 0.1 dB/cm, enabling high-power and dense circuit operation

Technical Novelty: First demonstration of 3D direct laser written polymer coupling interfaces specifically tailored for LTOI (both strip and rib geometries), versus prior work that focused on lithium niobate or other platforms. The octave-spanning bandwidth (850–1740 nm) on an EO platform combined with sub-dB losses is unprecedented.

What's New: First polymer-based 3D direct laser written couplers on LTOI for both fully-etched strip and partially-etched rib waveguides. Prior LTOI work lacked efficient I/O interfaces; this enables practical chip-to-fiber coupling across an octave of bandwidth, which is rare even on more mature platforms like LNOI or SiN.

Extension Opportunities:

Integrate these couplers with active LTOI electro-optic modulators to build complete high-speed data transmission links and benchmark end-to-end insertion loss at >100 Gbps
Use the octave-spanning bandwidth to co-couple fundamental (1550 nm) and second-harmonic (775 nm) signals in a single device for on-chip quantum squeezed light generation
Scale to dense fiber-array packaging by fabricating 2D coupler arrays via parallel 3D laser writing, enabling massively parallel I/O for photonic computing chips

Replicability: No mention of open code/data. Reproduction requires specialized facilities: LTOI wafers, e-beam lithography for waveguide patterning, ion-beam or RIE etching, and a two-photon 3D direct laser writing system (e.g., Nanoscribe). Significant capital investment (~$1M+ equipment) and cleanroom access required.

Research Gaps:

No demonstration of active devices (modulators, frequency converters) integrated with these couplers — full system-level performance unverified
Long-term reliability, environmental stability (temperature, humidity), and packaging robustness of the 3D-printed polymer interfaces remain unstudied

🔥 GitHub Trending

1. StokerReminisce/MATLAB-R2026a-Full-Toolboxes

⭐ 13 stars | Unknown

MATLAB R2026a with Simulink — model-based design, signal processing, computer vision, robotics and parallel computing

activator data-analysis data-science deep-learning engineering machine-learning

2. StratumManateeSpirit/Topaz-Video-AI-4K-Upscaler

⭐ 13 stars | Unknown

Topaz AI 2026 — Photo AI and Video AI with noise removal, sharpening, face recovery, frame interpolation and 4K upscaling

4k-upscale ai-tools ai-upscaler deep-learning denoising face-restoration

3. Vectorurvary/Topaz-Photo-Video-AI-Setup

⭐ 13 stars | Unknown

One-command installer for Topaz AI — upscale images and video to 4K/8K, remove noise, sharpen and recover faces with AI

4k-upscale ai-tools ai-upscaler deep-learning denoising face-restoration

4. GripScholarBlade/MATLAB-R2026a-Desktop-Setup

⭐ 11 stars | Unknown

One-liner setup for MATLAB R2026a — technical computing with Simulink, Signal Processing, Image Processing and Deep Learning

activator data-analysis data-science deep-learning engineering machine-learning

5. KapaSique/kaggle-dominator

⭐ 3 stars | Shell

A Claude Code skill that turns the agent into a relentless Kaggle Grandmaster — battle constitution + technique arsenal.

ai-agent anthropic claude claude-code data-science kaggle

6. build-with-ml/zero-to-mle

⭐ 1 stars | Unknown

🧠 24-week roadmap: Software Engineer → Machine Learning Engineer at Amazon, Meta, Google, OpenAI, Anthropic. Build in public.

career deep-learning interview-prep llm machine-learning mle

7. YashrajThube/Yashraj-AI-Assistant-OS

⭐ 1 stars | Python

Enterprise Generative AI scheduling assistant that transforms natural language into actionable workflows using Gemini AI, AI Agents, Google Calendar APIs, and intelligent planning systems.

ai-agents artificial-intelligence automation fatsapi gemini-ai generative-ai

8. codinghub37/Ai-document-assistant-rag

⭐ 1 stars | Python

An AI-powered document question-answering system that uses RAG, FAISS, Sentence Transformers, and Gemini 2.5 Flash to answer questions from uploaded PDF documents.

artificial-intelligence document-qa faiss gemini-api machine-learning nlp

9. Itsakkurana26/Preddictive-Modeling-using-Machine-Learning

⭐ 1 stars | Python

🌸 Machine Learning project using the Iris Dataset to predict flower species. Implemented a Random Forest Classifier, trained and tested the model, evaluated performance using accuracy score and confus

classification data-science iris-dataset machine-learning predictive-modeling python

10. Geeta3521/adaptive-multistage-rag-system

⭐ 1 stars | Jupyter Notebook

Production-ready Adaptive Multi-Stage Retrieval-Augmented Generation (RAG) pipeline for long-document question answering.

adaptive-rag agentic-ai chromadb document-question-answering faiss generative-ai

11. Jarpula-Nirjala/HybridDualNetPath

⭐ 1 stars | Jupyter Notebook

A hybrid deep learning system for brain MRI disease classification, combining EfficientNetV2-S and Swin-Tiny Transformer through cross-fusion attention to detect Alzheimer's Disease, Parkinson's Disea

alzheimers-detection attention-mechanism brain-mri cnn computer-vision deep-learning

12. Dangmotm/Datathon

⭐ 1 stars | Python

Datathon 2026 VinTelligence - Sales forecasting and EDA for Vietnamese fashion e-commerce. LightGBM, SHAP, NeurIPS report.

datathon-forecasting-lightgbm-eda-shap-vietnam deep-learning machine-learning

13. Alonessam/pneumonia-xray-detection

⭐ 1 stars | Python

Pneumonia detection from chest X-Ray images using Deep Learning (ResNet-50 vs CNN) and Explainable AI (Grad-CAM), featuring a Streamlit web UI.

computer-vision deep-learning explainable-ai grad-cam medical-imaging pytorch

14. praveenjoe115/pixel-truth

⭐ 0 stars | HTML

AI Generated Image Detection System using Deep Learning and Computer Vision.

ai bca-project computer-vision deep-learning flask image-classification

15. Giganoti/solar-activity-deep-learning

⭐ 0 stars | Jupyter Notebook

Deep learning models for predicting extreme solar activity using satellite observations.

artificial-intelligence cnn coronal-mass-ejections deep-learning keras lstm

Generated by Research Pulse on 2026-06-16 06:07