2026-06-04·Daily digest

🔬 Research Pulse

Daily Digest

June 04, 2026

🤖 AI

🧠 LLMs

1. Towards Efficient and Evidence-grounded Mobility Prediction with LLM-Driven Agent

Authors: Linyao Chen, Qinlao Zhao, Zechen Li... Published: 2026-06-03 | Citations: 0 arXiv | PDF

Research Question: How can LLM-based mobility prediction move beyond static prompts and single-pass inference to adaptively gather evidence when signals are weak or conflicting, without requiring task-specific training?

Summary: AgentMob is a training-free LLM agent for next-location prediction that splits cases into a fast historical-regularity path and a slow iterative tool-use path invoking trajectory, behavior, stay-move, and geographic evidence. It achieves SOTA among training-free LLM methods across three datasets and demonstrates that adaptive evidence gathering — not just tool access — drives an 18-point Acc@1 gain on ambiguous cases.

Key Results: AgentMob with GPT-5.4 achieves 71.42% Acc@1 on BW, 33.14% on YJMob100K, and 33.50% on Shanghai ISP — strongest among training-free LLM methods. On BW non-fast-path (ambiguous) cases, the LLM controller boosts Acc@1 from 30.65% to 48.62% versus a same-tool statistical baseline, isolating the adaptive-reasoning gain at ~18 absolute points.

Key Findings:

Adaptive routing matters: the LLM controller's evidence-gathering logic adds ~18 Acc@1 points over a same-tool statistical baseline on ambiguous BW cases (30.65% → 48.62%).
Training-free agentic prompting can reach SOTA on individual mobility prediction: 71.42% / 33.14% / 33.50% Acc@1 on BW / YJMob100K / Shanghai ISP.
Most prediction value sits in the long tail — routine cases are easy via fast path, so the LLM's main payoff is concentrated in ambiguous, evidence-conflicting predictions.

Technical Novelty: A training-free dual-path agent that routes routine cases via historical-regularity fast path and triggers iterative tool use (recent trajectory, historical behavior, stay-move likelihood, geographical evidence) only for ambiguous cases — replacing static single-pass LLM prompting with adaptive evidence-controlled decision making.

What's New: Unlike prior LLM mobility work that uses fixed prompts and single-pass inference, AgentMob is the first to frame next-location prediction as an evidence-controlled agent loop with dynamic fast/slow routing and explicit tool selection over heterogeneous mobility signals — and it isolates the reasoning gain via a same-tool statistical ablation.

Extension Opportunities:

Add real-time contextual tools (weather, events, transit disruptions) to the agent's tool set to test whether external signals further reduce ambiguity on non-fast-path cases.
Distill the agent's iterative trajectories into a small fine-tuned model to retain interpretability while cutting inference cost and LLM API spend at city scale.
Extend the fast/slow routing paradigm to multi-step (next-k locations) or group/household mobility prediction, where evidence conflicts compound across agents.

Replicability: Code is open-sourced at github.com/Unknown-zoo/AgentMob. Reproduction requires API access to a frontier LLM (GPT-5.4) plus the three public-ish mobility datasets (BW, YJMob100K, Shanghai ISP); compute is modest (no training) but LLM inference cost scales with non-fast-path fraction.

Research Gaps:

No comparison against supervised SOTA sequence models on the same splits, so the accuracy ceiling vs. trained baselines is unclear.
Cost, latency, and tool-call budget analysis of the agent loop is not surfaced in the abstract — practicality at city-scale simulation remains an open question.

2. Audio Interaction Model

Authors: Zhifei Xie, Zihang Liu, Ze An... Published: 2026-06-03 | Citations: 0 arXiv | PDF

Research Question: How can we unify offline Large Audio Language Models (LALMs) and single-task streaming audio models into a single online LALM that listens, decides, and responds in real time across diverse audio interaction tasks?

Summary: The paper introduces the Audio Interaction Model — an online unified LALM that runs a continuous perceive-decide-respond loop to handle both offline audio tasks and real-time streaming interaction. It is realized through Audio-Interaction (the model), SoundFlow (the end-to-end training/inference framework), StreamAudio-2M (2.6M streaming examples), and Proactive-Sound-Bench, demonstrating proactive audio intervention capabilities absent in prior offline LALMs.

Key Results: Built Audio-Interaction model evaluated across 8 benchmarks, preserving competitive performance on mainstream audio tasks while enabling real-time ASR, streaming audio instruction following, and proactive help. Constructed StreamAudio-2M (2.6M-item streaming corpus spanning 7 fundamental abilities and 28 sub-tasks) and Proactive-Sound-Bench for evaluating proactive audio intervention.

Key Findings:

A single unified streaming LALM can preserve competitive offline-task performance while adding online instruction following and proactive responses
Comprehension-aware training plus asynchronous inference yields stable low-latency real-time audio interaction
Proactive intervention (deciding when to respond from stream semantics) is a measurable new capability — quantified via the new Proactive-Sound-Bench

Technical Novelty: Formalization of the 'Audio Interaction Model' regime with an always-on perceive-decide-respond loop, plus the SoundFlow framework combining streaming-native data construction, comprehension-aware training, and asynchronous low-latency inference — going beyond single-task streaming ASR or voice chat models by letting the model decide when to respond based on stream semantics.

What's New: Prior streaming audio models specialize in one task (ASR or voice chat); prior LALMs are offline. This work is the first to formalize and deliver an always-on unified streaming LALM that decides when to speak, with a dedicated benchmark for proactive audio intervention.

Extension Opportunities:

Extend the perceive-decide-respond loop to multimodal streams (vision + audio) for embodied agents that react to environmental sound and visual cues simultaneously
Adapt the proactive intervention mechanism for domain-specific assistants (medical monitoring, accessibility tools for hearing-impaired users, industrial anomaly detection)
Distill the streaming model into smaller on-device variants and benchmark latency/quality tradeoffs for edge deployment in wearables or smart speakers

Replicability: Abstract does not explicitly mention open-sourcing code/weights. Reproduction would require substantial compute given 2.6M streaming training items and likely a multi-billion-parameter audio LLM backbone; data construction pipeline appears to be a major engineering effort in itself.

Research Gaps:

No standard evaluation existed for proactive audio intervention before this work — broader community adoption of Proactive-Sound-Bench remains to be seen
Trade-offs between proactivity (false interjections) and reticence (missed opportunities) are not fully characterized in the abstract

3. Knowledge Index of Noah's Ark

Authors: Sheng Jin, Minghao Liu, Yunze Xiao... Published: 2026-06-03 | Citations: 0 arXiv | PDF

Research Question: How can knowledge benchmarks for LLMs overcome three structural flaws: scaling-driven designs lacking disciplinary representativeness, flat-payment annotation enabling lazy consensus, and unaudited ranking instability under bounded test budgets?

Summary: KINA is an 899-item, 261-discipline LLM knowledge benchmark that formalizes representativeness as a submodular coverage problem (with a (1-1/e) approximation guarantee) and introduces a bonus-on-bar tournament annotation mechanism proven to dominate flat-payment in review quality. Evaluation of 42 models reveals a tiered leaderboard topped by Gemini-3.1-Pro-Preview at 53.17%, with substantial headroom and tool augmentation adding up to 5.17 points.

Key Results: Introduces KINA, an 899-item benchmark spanning 261 fine-grained disciplines. Proves two formal results: (1) Proposition 1 — disciplinary representativeness via coverage-style proxy yields a (1-1/e) greedy approximation; (2) Theorem 1 — bonus-on-bar tournament weakly FOSD-dominates flat payment in released-review quality with incentive-compatibility threshold B > ΔC/Δp_min. Evaluated 42 models from 13 labs: Gemini-3.1-Pro-Preview leads at 53.17%, Claude-Opus-4.6 at 49.92%, GPT-5.4 at 48.55%. Tiered structure: frontier tier >48%, strong tier 38–45%, weak tier modestly above 10% chance baseline. Tool augmentation adds up to 5.17 points across five tool-use evaluations.

Key Findings:

Frontier models cap at 53.17% (Gemini-3.1-Pro-Preview), indicating substantial headroom and no saturation on disciplinary knowledge
The leaderboard is tiered (frontier >48%, strong 38–45%, weak near 10% baseline) rather than a smooth gradient, suggesting capability discontinuities
Tool augmentation gains are heterogeneous across models (up to +5.17 points), indicating tool-use ability is not uniformly distributed even among strong models

Technical Novelty: Combines a submodular coverage formulation of disciplinary representativeness (with greedy approximation guarantees) with a game-theoretic annotation payment scheme (bonus-on-bar tournament with FOSD dominance), plus explicit bootstrap-based ranking-stability reporting — three methodological levers that prior knowledge benchmarks (MMLU, GPQA, etc.) treat informally.

What's New: Unlike scale-first benchmarks, KINA operationalizes representativeness with a formal coverage proxy and approximation guarantee, replaces flat-rate annotation with an incentive-compatible tournament with a provable dominance result, and audits ranking stability via bootstrap — moving benchmark construction from heuristic to mechanism-design territory.

Extension Opportunities:

Apply the bonus-on-bar tournament incentive mechanism to other annotation pipelines (RLHF preference data, red-teaming) and empirically measure quality lift vs flat payment
Extend the coverage-style representativeness proxy beyond academic disciplines to industry domains (legal verticals, medical specialties, engineering codes) using expert-elicited anchors
Build an adaptive evaluation harness that uses the bootstrap ranking-stability statistics to dynamically allocate test budget to disambiguate adjacent ranks

Replicability: Abstract does not mention code/data release. Reproduction would require the 899 items, the 261-discipline taxonomy with expert anchors, and inference budget for 42 models (mostly API-based, modest compute for evaluation; tool-use evaluations require agent scaffolding).

Research Gaps:

The (1-1/e) guarantee applies to the proxy, not population representativeness — the gap between proxy and true coverage of human knowledge is unbounded
Incentive-compatibility threshold B > ΔC/Δp_min requires estimating annotator cost and probability parameters that may be unobservable in practice

🦾 ROBOTICS

1. GRAIL: Generating Humanoid Loco-Manipulation from 3D Assets and Video Priors

Authors: Tianyi Xie, Haotian Zhang, Jinhyung Park... Published: 2026-06-03 | Citations: 0 arXiv | PDF

Research Question: How can we scale humanoid loco-manipulation training data without the bottlenecks of teleoperation, motion capture, physical setups, and instrumented actors, while still producing robot-compatible demonstrations across diverse objects, whole-body motions, and scene geometries?

Summary: GRAIL is a fully-virtual data generation pipeline that synthesizes humanoid loco-manipulation demonstrations by combining 3D assets, simulator-ready scenes, and video foundation model priors, sidestepping teleoperation and motion capture entirely. Starting from fully-specified 3D configurations rather than in-the-wild videos enables accurate metric 4D human-object interaction recovery, which is retargeted to a humanoid robot. Policies trained purely on 20K+ GRAIL sequences transfer to a real Unitree G1 with 84% pick-up and 90% stair-climbing success.

Key Results: GRAIL generated over 20,000 synthetic sequences spanning pick-up, object manipulation, sitting, and terrain traversal. Using only GRAIL-generated data (no real-world demos), trained egocentric visual policies deployed on a Unitree G1 humanoid achieved 84% real-world success on diverse object pick-up and 90% success on stair-climbing, demonstrating sim-to-real transfer from purely synthetic data.

Key Findings:

Privileged 3D scene specification (known geometry/camera/scale) before video generation substantially reduces depth ambiguity and morphology mismatch during 4D HOI reconstruction compared to in-the-wild video reconstruction
Purely synthetic data from a fully-virtual pipeline is sufficient for sim-to-real transfer of egocentric humanoid policies, with no real teleoperation needed
Decoupling manipulation (object-aware latent adaptor) from locomotion (scene-aware tracker) as complementary task-general trackers enables a single pipeline to cover both pick-up and terrain traversal

Technical Novelty: Inverts the usual 'reconstruct from in-the-wild video' pipeline: instead of recovering unknown 3D structure from videos, GRAIL starts from fully specified 3D configurations (known geometry, camera, metric scale, robot-proportioned character) and uses VFMs to generate video conditioned on that scaffold. This 'privileged' setup eliminates depth ambiguity and human-robot morphology mismatch during 4D HOI reconstruction. The pairing of an object-aware latent adaptor for manipulation with a scene-aware tracker for traversal is also new.

What's New: Unlike prior humanoid data pipelines that rely on teleoperation, mocap, or reconstruction of unconstrained internet video, GRAIL inverts the generation/reconstruction loop—generating video from known 3D scaffolds—so reconstruction becomes well-posed. It is among the first to demonstrate full sim-to-real humanoid loco-manipulation from VFM-synthesized data alone.

Extension Opportunities:

Extend GRAIL to bimanual dexterous manipulation tasks (tool use, articulated object manipulation, in-hand reorientation) where current VFM priors are weakest
Add a closed-loop refinement step where deployed policy failures feed back as targeted 3D scene generations, creating an active data-generation curriculum
Replace single-character video generation with multi-agent scenes to synthesize human-robot collaboration or crowd-navigation data for social humanoid behaviors

Replicability: Abstract does not mention code/data release. Reproduction would require: (1) access to video foundation models (likely closed/expensive), (2) a 3D asset library and simulator (e.g., IsaacSim/MuJoCo), (3) substantial GPU compute for VFM inference at 20K+ sequence scale, plus sim-to-real RL training, and (4) a Unitree G1 humanoid (~$16K+) for real-world validation. Likely 8+ high-end GPUs for weeks.

Research Gaps:

Manipulation diversity beyond pick-up (e.g., tool use, deformables, bimanual coordination) is not addressed and may stress VFM physical-plausibility limits
No reported comparison against equivalent teleop/mocap-trained baselines, leaving the quality gap between synthetic and human-collected demonstrations unquantified

2. Potential-Guided Flow Matching for Vision-Language-Action Policy Improvement

Authors: Yunpeng Mei, Jiakai He, Hongjie Cao... Published: 2026-06-03 | Citations: 0 arXiv | PDF

Research Question: How can VLA policies learn effectively from mixed-quality deployment data (successes, partial completions, recoverable mistakes, failures) without imitating failures (full BC), discarding useful sub-trajectories (filtered BC), or paying the cost of a separate critic (offline RL)?

Summary: ForesightFlow is a self-guided flow-matching VLA policy that jointly generates action chunks and a success-potential trajectory, enabling best-of-K inference without a separate critic. Its core trick is decoupling advantage weighting — applied only to action velocities, not potential velocities — to avoid value hallucination, plus a one-step boundary estimator that makes advantage computation cheap. It matches separate-critic offline RL on simulation, beats it in real-world bimanual tasks, and cuts training compute by 38%.

Key Results: On 5 BEHAVIOR-1K simulation tasks and 5 real-world bimanual tasks, ForesightFlow beats imitation baselines, matches the strongest separate-critic offline RL baseline in simulation success, improves real-world success, and reduces training compute by 38%. Ablations show decoupling prevents value hallucination, the one-step boundary estimator preserves candidate-ranking fidelity, and self-guided sampling improves long-horizon execution.

Key Findings:

Self-scoring via a learned potential head reaches the success rate of a dedicated critic while eliminating critic-network overhead.
Naively sharing advantage weights between action and potential heads causes value hallucination; decoupling supervision is necessary for calibrated scores.
A one-step boundary estimator yields advantages faithful enough for candidate ranking without multi-step ODE rollouts, enabling 38% training compute savings.

Technical Novelty: Three coupled ideas: (1) augmenting flow-matching action chunks with a co-generated learned success-potential trajectory so the same network proposes AND scores candidates (no external critic); (2) decoupled advantage-weighted flow matching — exponentiated advantage weights are applied only to action velocities while potential velocities are trained uniformly, preventing failure-gradient suppression and overconfident scores; (3) a one-step boundary estimator for conditional flow matching enabling advantage computation in a single stop-gradient forward pass.

What's New: Prior offline-RL-for-VLA work bolts a separate critic onto a generative policy; prior advantage-weighted flow/diffusion methods apply weights uniformly. ForesightFlow is the first to (a) make a flow-matching policy its own critic via a co-generated potential channel and (b) recognize that policy improvement and value calibration require asymmetric supervision on the two channels.

Extension Opportunities:

Apply decoupled advantage-weighted flow matching to diffusion-policy variants (e.g., Octo, RDT-1B) to test whether the action/potential decoupling trick generalizes beyond flow matching.
Replace best-of-K sampling with a learned potential-guided ODE drift term so the same network steers sampling toward high-success regions in a single pass, eliminating the K-fold inference cost.
Extend the success-potential trajectory into a multi-objective potential (success + safety + efficiency) to enable controllable trade-offs at inference for safety-critical bimanual manipulation.

Replicability: Abstract doesn't mention code/data release. Reproduction would require BEHAVIOR-1K simulator access, a bimanual real-robot setup (likely 2x 6-7 DoF arms), and VLA-scale training compute — though the claimed 38% compute reduction vs. separate-critic baselines suggests it's cheaper than typical offline-RL VLA work. Likely needs 8+ A100/H100 GPU-days for the flow model.

Research Gaps:

Abstract reports only aggregate success on 10 tasks — no analysis of how the method scales with the proportion of failure data or with chunk length / horizon.
No theoretical characterization of when the one-step boundary estimator's advantage approximation degrades vs. multi-step rollouts, or how K trades off against potential calibration quality.

3. WAM-Nav: Asymmetric Latent World-Action Modeling for Unified Visual Navigation

Authors: Ning Yang, Yan Huang, Kaiwen Peng... Published: 2026-06-03 | Citations: 0 arXiv | PDF

Research Question: How can visual navigation policies combine anticipatory visual foresight with action generation in a unified model, avoiding the error accumulation and inference latency of modular predict-then-act approaches while supporting multiple goal modalities?

Summary: WAM-Nav is a unified embodied navigation policy that jointly learns action generation and latent visual foresight via asymmetric joint diffusion in a shared Diffusion Transformer, eliminating the error accumulation of modular predict-then-act pipelines. It supports Image-Goal, Point-Goal, and No-Goal exploration within a single model and achieves strong gains on ClutterScenes/InternScenes plus 85% zero-shot sim-to-real success.

Key Results: WAM-Nav improves success rates by 15.7% on Image-Goal and 3.3% on Point-Goal navigation on ClutterScenes and InternScenes benchmarks. Zero-shot sim-to-real transfer achieves 85% average task success rate across diverse indoor and outdoor real-world environments using a single unified policy for Image-Goal, Point-Goal, and No-Goal exploration.

Key Findings:

Asymmetric joint diffusion (long-horizon actions + short-horizon visual latents) reduces inference latency and visual error accumulation versus autoregressive rollouts
Dual-stream conditioning combining episode-level ego-motion with sequential visual observations produces smoother, more consistent trajectories
A single unified policy with goal alignment generalizes across Image-Goal (+15.7%), Point-Goal (+3.3%), and No-Goal navigation, and transfers zero-shot to real robots at 85% success

Technical Novelty: Asymmetric joint diffusion in a shared Diffusion Transformer that generates long-horizon actions and short-horizon visual latents concurrently (rather than autoregressively), combined with a dual-stream contextual conditioning mechanism (episode-level ego-motion + sequential visual observations) and a unified goal alignment module supporting three goal types in one policy.

What's New: Unlike modular methods that separate world modeling from policy learning, WAM-Nav fuses them in one shared Diffusion Transformer with asymmetric horizons — long actions, short visual foresight — sidestepping autoregressive error accumulation while preserving foresight benefits. The unified goal alignment that handles Image/Point/No-Goal in a single policy is also unusual in the visual navigation literature.

Extension Opportunities:

Extend the asymmetric joint diffusion to multi-agent navigation where each agent predicts both its own actions and neighbors' short-horizon visual states for cooperative collision avoidance
Replace the short-horizon visual foresight with a learned 3D/occupancy latent to enable navigation in dynamic environments with moving obstacles
Add language-goal conditioning to the unified goal alignment module to make WAM-Nav a fourth-mode policy supporting instruction-following navigation

Replicability: Abstract does not mention code release. Reproduction would require a Diffusion Transformer backbone, the ClutterScenes and InternScenes simulators, and substantial GPU compute typical for DiT training on embodied datasets (likely multi-GPU for days). Real-world validation requires a physical robot platform.

Research Gaps:

Reactive policies lack anticipatory reasoning, while modular world-model-plus-policy designs suffer error accumulation and inference latency
Most navigation policies are specialized per goal type, requiring separate models for Image-Goal, Point-Goal, and exploration

💻 COMPUTE

1. A thorium-229 optical nuclear clock with feedback loop

Authors: L. Toscani De Col, T. Riebner, I. Morawetz... Published: 2026-06-03 | Citations: 0 arXiv | PDF

Research Question: Can a thorium-229 nuclear optical clock be realized with active laser stabilization to the 148 nm nuclear transition, and how does its stability compare with state-of-the-art atomic clocks for fundamental physics tests?

Summary: The authors implement the first feedback-stabilized thorium-229 nuclear optical clock, locking a CW 148 nm laser to the nuclear transition in a CaF2 crystal and referencing it against a Yb+ ion clock. They achieve a shot-noise-limited instability of 3·10^-12 √(τ/s), reaching 10^-15 over a day, and use the clock to place competitive constraints on ultralight dark matter couplings to photons, the strong force, and quarks.

Key Results: Demonstrated a closed-loop Th-229 nuclear clock by stabilizing a CW laser to the 148 nm transition in a room-temperature CaF2 crystal via continuous absorption spectroscopy, comparing a subharmonic against a Yb+ single-ion clock. Achieved shot-noise-limited fractional frequency instability of 3·10^-12 √(τ/s), reaching ~10^-15 over 1 day. Used the clock to constrain ultralight dark matter couplings (timescales 20 s – 1 day) — competitive with best atomic clocks for photon coupling and surpassing prior limits on strong-force and quark couplings.

Key Findings:

Closed-loop locking to the Th-229 nuclear transition is achievable in a solid-state host at room temperature via continuous absorption spectroscopy.
Fractional frequency instability follows shot-noise scaling at 3·10^-12 √(τ/s), reaching ~10^-15 after 1 day of operation.
The clock yields ultralight dark matter constraints that match the best atomic clocks for photon coupling and improve on prior bounds for strong-force and quark couplings.

Technical Novelty: First closed-loop frequency stabilization of a laser to the Th-229 nuclear transition using continuous absorption spectroscopy in a solid-state (CaF2) host at room temperature — moving beyond prior open-loop spectroscopic detections of the isomer toward an operational clock with a real feedback servo and continuous comparison to an optical atomic standard.

What's New: Transitions Th-229 from spectroscopic demonstration to an actual operating nuclear clock with active feedback, in a solid-state platform — exploiting the nuclear transition's enhanced sensitivity to variations of fundamental constants for new-physics searches.

Extension Opportunities:

Engineer lower-defect / cryogenic Th-229-doped crystals (or alternative hosts like LiSrAlF6) to reduce inhomogeneous broadening and push instability several orders of magnitude lower as projected.
Build a compact transportable Th-229 nuclear clock for geodesy, gravitational potential mapping, and field-deployable searches for spatially-varying fundamental constants.
Extend dark-matter coupling searches to longer integration windows (months–years) and combine with global atomic-clock networks to probe lower DM masses and topological defect signatures.

Replicability: No code/data release mentioned in the abstract. Reproduction requires a Th-229-doped CaF2 crystal (rare radioactive material), a VUV (148 nm) CW laser system with frequency-comb-linked subharmonic, and a co-located Yb+ optical ion clock — accessible only to a handful of metrology labs worldwide.

Research Gaps:

Systematic uncertainty budget (crystal-induced shifts, temperature/strain dependence) and absolute accuracy remain immature compared with the demonstrated stability.
Projected orders-of-magnitude instability improvements depend on host-material engineering that has not yet been realized.

2. Phase-correlation-free quantum key distribution source operating at gigahertz rates

Authors: Shashank Kumar, Alessandro Marcomini, Loïc Millet... Published: 2026-06-03 | Citations: 0 arXiv | PDF

Research Question: How can decoy-state QKD systems achieve intrinsic phase randomization at gigahertz repetition rates without the inter-pulse correlations that plague actively phase-randomized or gain-switched laser sources, which compromise security and performance?

Summary: The paper introduces a 1.25 GHz QKD source built on a super-luminescent LED in the C-band, exploiting the spontaneous-emission nature of SLEDs to provide intrinsic, correlation-free global phase randomization between pulses. It demonstrates >99% interferometric visibility within signals while preserving phase randomization between them, offering a compact, scalable alternative to actively randomized or gain-switched laser sources.

Key Results: Demonstrated a 1.25 GHz phase-randomized QKD source using a C-band SLED, generating ~100 ps optical pulses with 400 ps pulse-to-pulse separation. Interferometric measurements show >99% visibility between adjacent time bins (confirming strong first-order coherence within signals), while spontaneous-emission dynamics guarantee global phase randomization between adjacent signals — compatible with high-speed time-bin encoding.

Key Findings:

SLED-based source achieves 1.25 GHz repetition with ~100 ps pulses and 400 ps spacing, compatible with time-bin encoding
99% interferometric visibility between adjacent time bins confirms intra-signal first-order coherence
Spontaneous-emission dynamics deliver intrinsic inter-pulse global phase randomization without active modulation

Technical Novelty: Using a super-luminescent LED — whose spontaneous-emission-driven output is intrinsically phase-randomized — as a QKD source, rather than relying on active phase modulation or gain-switching of a laser. This sidesteps the inter-pulse correlation problem at GHz rates by construction rather than by mitigation.

What's New: Prior GHz QKD sources rely on actively phase-randomized or gain-switched lasers, both of which exhibit inter-pulse correlations at high rates that undermine decoy-state security proofs. This work replaces them with a SLED whose physics inherently guarantees phase randomization, eliminating the correlation problem at the source rather than patching it.

Extension Opportunities:

Integrate the SLED source into a full prepare-and-measure QKD link and benchmark secret key rate against gain-switched laser baselines at matched repetition rates
Push beyond 1.25 GHz (e.g., 2.5–10 GHz) and characterize whether SLED spontaneous emission still guarantees intrinsic global phase randomization at shorter pulse spacings
Combine the SLED with measurement-device-independent (MDI) or twin-field QKD protocols where source-side phase randomization assumptions are especially critical

Replicability: No code/data availability mentioned in the abstract. Reproduction requires a C-band SLED, GHz-class pulse-shaping/modulation electronics, and an interferometric setup for visibility measurement — moderate optics-lab cost, no significant compute requirement.

Research Gaps:

No reported secret key rate or full QKD link demonstration — only source-level characterization
Scalability beyond 1.25 GHz and behavior of SLED phase randomization at shorter pulse separations remain unexplored

3. FlexNPU: Transparent NPU Virtualization for Dynamic LLM Prefill-Decode Co-location

Authors: Jiongjiong Gu, Jianfeng Wang, Zidong Han... Published: 2026-06-03 | Citations: 0 arXiv | PDF

Research Question: How can NPU deployments overcome the rigidity of direct device passthrough to enable phase-aware, dynamic scheduling for LLM serving — specifically reconciling the compute-bound prefill phase with the memory-bandwidth-bound decode phase without the imbalance penalties of static PD disaggregation?

Summary: FlexNPU is a transparent user-space virtualization layer that interposes on AscendCL APIs to route NPU operations through per-device daemons, decoupling applications from physical Ascend NPUs without changes to models, frameworks, or drivers. This runtime boundary enables dynamic prefill-decode co-location for LLM serving, adapting scheduling to the complementary compute/memory characteristics of the two phases and outperforming both static PD disaggregation and static PD co-location on real Ascend deployments.

Key Results: Implemented on Huawei Ascend NPUs, FlexNPU shows: (1) no measurable inference overhead vs. direct passthrough, with slight throughput gains in some cases; (2) on a 384-card Ascend 910C deployment of DeepSeek-R1, throughput improvements of 5.15% and 26.33% over static PD disaggregation; (3) on Qwen2.5-7B vs. static PD co-location, comparable throughput with >92% TTFT reduction and nearly unchanged TPOT.

Key Findings:

Transparent API interposition adds no measurable overhead vs. direct NPU passthrough, validating user-space virtualization as a viable substrate
Dynamic PD co-location beats static PD disaggregation by 5.15–26.33% throughput on a 384-card DeepSeek-R1 deployment by mitigating resource imbalance
Versus static PD co-location, dynamic scheduling cuts TTFT by over 92% on Qwen2.5-7B while keeping TPOT and throughput largely intact — i.e., latency wins without throughput cost

Technical Novelty: A user-space, API-interposition virtualization layer for NPUs that decouples application AscendCL calls from physical devices via per-device daemons — enabling operator-level dispatch control and dynamic prefill-decode co-location without modifying model code, AI frameworks, or NPU drivers. Prior NPU work either exposes devices directly or relies on static PD disaggregation.

What's New: Combines three rarely co-occurring properties: (1) full transparency (no model/framework/driver changes), (2) operator-level dispatch control via per-device daemons, and (3) phase-aware dynamic PD co-location — moving beyond the prevailing static disaggregation paradigm and applied to a non-NVIDIA accelerator (Ascend).

Extension Opportunities:

Port the AscendCL interposition technique to other accelerator stacks (CUDA/ROCm) to enable cross-vendor transparent virtualization for PD co-location
Add ML-driven adaptive schedulers on top of the FlexNPU runtime that predict prefill/decode resource demand from request features (prompt length, batch composition) rather than reacting to current load
Extend the virtualization layer to support multi-tenant SLO-aware QoS isolation, mixing LLM serving with conventional inference workloads on the same physical NPU pool

Replicability: The abstract does not mention public code release. Reproduction at the headline scale requires a 384-card Ascend 910C cluster (Huawei-specific hardware), which is largely inaccessible outside Huawei partners; smaller-scale validation on a handful of Ascend NPUs running Qwen2.5-7B would be more feasible but still vendor-locked.

Research Gaps:

Evaluation appears limited to Ascend hardware and a small model set (DeepSeek-R1, Qwen2.5-7B); generalization to other NPUs/GPUs and longer-context or MoE-heavy workloads is unverified
Daemon-mediated dispatch likely raises fault-isolation, security, and multi-tenant fairness questions that the abstract does not address

⚡ ENERGY

1. Low Distortion Fusion Bonding using Pneumatically Warped Wafers

Authors: Utkarsh Jain, Koen D'have, Philipp Schmidt... Published: 2026-06-03 | Citations: 0 arXiv | PDF

Research Question: How can wafer-to-wafer fusion bonding for backside power-delivery-network (BSPDN) schemes be performed without introducing high-order shape changes, localized stresses, and grid distortions that propagate to subsequent lithographic exposures on the thinned wafer?

Summary: The paper introduces a pneumatically warped wafer-to-wafer fusion bonding technique on the SUSS XBA tool that replaces localized mechanical bond initiation with uniform backside gas pressure, keeping wafers symmetric and compliant during bonding. This eliminates a dominant source of bond-induced grid distortion in BSPDN flows, achieving sub-10 nm residuals with low-order corrections and ≤3 nm with advanced CPE models, validated by both PWG and scanner metrology.

Key Results: Using the SUSS XBA tool with pneumatically warped bonding (controlled backside gas pressure throughout bonding), the authors demonstrated sub-10 nm residual grid distortion achievable with relatively low-order correction models, and ≤3 nm residual distortion using advanced CPE (correction-per-exposure) models. Patterned wafer geometry (PWG) distortion predictions were validated against dense scanner metrology on patterned-bonded wafers, confirming spatial agreement of distortion location and severity.

Key Findings:

Pneumatically warped bonding without localized force initiation yields low-distortion bonded stacks suitable for BSPDN integration
PWG-derived IPD predictions correlate with dense scanner grid measurements, validating PWG as a cheaper proxy for distortion characterization
≤3 nm residual grid distortion is achievable with advanced CPE models, and sub-10 nm with low-order correction — sufficient for advanced-node lithographic overlay budgets
Simple process tuning (gas pressure recipe) decouples the dominant distortion source from edge-related variability

Technical Novelty: Prior fusion bonding initiates the bond via a localized mechanical pin/force, creating asymmetric bond-front propagation and edge-related distortions. This work eliminates localized bond initiation entirely by applying uniform pneumatic pressure across the full wafer backside, keeping both wafers compliant and symmetric throughout bonding — decoupling the dominant distortion source from edge variability.

What's New: Most prior bonding distortion work focuses on post-bond compensation via lithography corrections or wafer-shape pre-conditioning. This paper attacks the problem at the source by changing the bonding mechanics themselves — symmetric, distributed pneumatic initiation rather than asymmetric point-contact initiation — and quantitatively links the new mechanics to lithography-relevant overlay residuals.

Extension Opportunities:

Develop closed-loop control where real-time PWG measurements during bonding feed back into gas-pressure profiles to actively cancel emerging distortion modes
Extend the pneumatic warping approach to hybrid bonding (Cu-Cu + dielectric) where alignment tolerances are even tighter (<5 nm) for advanced 3D logic stacking
Build ML-based distortion prediction models that map pre-bond wafer shape + gas-pressure recipe to post-bond IPD fields, enabling recipe optimization without physical iteration

Replicability: No code or dataset is mentioned. Reproduction requires specialized fab equipment: the SUSS XBA bonding tool, PWG metrology, and a dense scanner for grid readout — not reproducible outside a semiconductor research fab (e.g., imec-scale facility). Compute requirements (gradient-based IPD modeling, CPE fitting) are modest.

Research Gaps:

No characterization of throughput, yield, or defectivity trade-offs of pneumatic bonding vs. conventional mechanical initiation at HVM scale
Performance under hybrid bonding (with metal pads) — where Cu-Cu contact mechanics complicate the pneumatic pre-stress assumption — is not addressed

2. 20 ps Non-Destructive Read and 1 ns Write Operations at <5 V in Ferroelectric HfO2/ZrO2 Non-Volatile Memories

Authors: Alexandre Baigol, Ruben Hamming-Green, Paul Uriarte Vicandi... Published: 2026-06-02 | Citations: 0 arXiv | PDF

Research Question: How can ferroelectric HfO2/ZrO2 non-volatile memories achieve simultaneously low-voltage (<5 V), nanosecond multi-level programming AND non-destructive ultrafast read-out, given that the MemCapacitance (MC) read window collapses above 1 MHz read frequencies — a critical bottleneck for analog in-memory computing using FeCap devices?

Summary: The paper integrates HfO2/ZrO2 ferroelectric capacitors in CMOS BEOL achieving <5 V, 1 ns multilevel writes and >10^11 cycle endurance with 10-year retention. To break the 1 MHz read-frequency ceiling where the memcapacitance window collapses, the authors introduce a non-destructive readout using 20 ps pulses that probe polarization-dependent leakage current at only 14 fJ per read, enabling analog in-memory computing with ferroelectric devices.

Key Results: Demonstrated: (1) HfO2/ZrO2 ferroelectric nanolayers integrated in BEOL of CMOS and on SiO2/Si; (2) nanosecond multilevel switching at <5 V programming voltage; (3) write operations at 1 ns; (4) endurance >10^11 cycles via partial ferroelectric switching; (5) 10-year retention of MC states; (6) novel non-destructive read via 20 ps electrical pulses (below RC time constant) measuring polarization-dependent leakage current; (7) read energy of only 14 fJ per operation — overcoming the >1 MHz MC window collapse limit.

Key Findings:

Partial ferroelectric switching boosts FeCap endurance above 10^11 cycles while producing distinguishable MemCapacitance states with 10-year retention
Conventional capacitive readout fails above 1 MHz due to MC window collapse — a previously underreported speed ceiling for FeCap analog memory
20 ps pulses (below RC time constant) enable non-destructive read via polarization-dependent leakage current at 14 fJ, decoupling read speed from RC limits

Technical Novelty: The core novelty is the sub-RC-time-constant readout: using 20 ps pulses shorter than the device RC time constant to probe polarization-dependent leakage current rather than displacement/capacitance current. This sidesteps the >1 MHz MC window collapse that has limited prior FeCap memcapacitor read schemes, while remaining non-destructive — unlike conventional FeRAM destructive polarization-switching reads.

What's New: Prior FeRAM/FeCap work has either destructive reads (polarization switching) or capacitive reads limited by RC and the <1 MHz MC window. This work is the first to demonstrate ps-scale non-destructive reads by exploiting polarization-modulated leakage rather than capacitance, in a CMOS-BEOL-compatible HfO2/ZrO2 stack with simultaneous ns writes at <5 V.

Extension Opportunities:

Build a crossbar array prototype using these FeCaps for analog matrix-vector multiplication, benchmarking energy/MAC against SRAM and ReRAM in-memory computing baselines
Co-design CMOS peripheral circuits (sub-20 ps pulse generators and femtojoule-sensitive sense amplifiers) to actually exploit the 20 ps read regime at system level
Extend the polarization-dependent leakage-current readout methodology to other ferroelectric stacks (e.g., Hf0.5Zr0.5O2, AlScN, doped HfO2) to test generality of the sub-RC read technique

Replicability: No code/data availability mentioned in the abstract. Reproduction requires substantial fab infrastructure: BEOL-compatible HfO2/ZrO2 atomic-layer deposition, CMOS integration capability, and ultrafast electrical characterization equipment (sub-20 ps pulse generators, high-bandwidth sampling oscilloscopes, RF probing). Not reproducible without a semiconductor cleanroom and ps-scale electrical test setup.

Research Gaps:

No array-level or in-memory-computing-task demonstration — only device-level metrics; bit-error rate, state count, and analog MAC accuracy in a crossbar remain unproven
Variability, yield, and device-to-device matching of the leakage-based read across CMOS wafers are not addressed, which is critical for analog in-memory computing fidelity

3. Strain-induced suppression of thermochromism in divalent cobalt molybdate thin films

Authors: Kiri Van Koughnet, Joey Williamson, Kane Hill... Published: 2026-06-03 | Citations: 0 arXiv | PDF

Research Question: How does the thermochromic β→α structural transition of CoMoO4 (well-characterized in bulk powders near 230K) manifest in thin-film form, where microstrain and substrate clamping may alter the phase landscape critical for device applications?

Summary: The paper presents the first THz-VIS-UV spectroscopic and structural study of CoMoO4 thin films, showing that microstrain suppresses the bulk β→α thermochromic transition and stabilizes the β-phase across all temperatures. Through anomalous phonon softening at 42 cm⁻¹, blue-shifted Co²⁺ crystal-field transitions, and supporting theory, the authors identify strain-induced crystal-field strengthening as the mechanism, establishing strain as a thermodynamic lever for controlling phase stability in thermochromic oxides.

Key Results: Combined experiment + theory on CoMoO4 thin films showed: (1) first THz-VIS-UV spectra of β-CoMoO4 thin films; (2) complete suppression of the bulk β→α transition (normally at ~230K) — β-phase retained across all temperatures, confirmed by temperature-dependent XRD; (3) anomalous softening of a 42 cm⁻¹ cation-dominated phonon on cooling (opposite of normal mode-hardening), saturating between 200–150K; (4) blue-shift of Co²⁺ crystal-field transitions and charge-transfer excitations across 0.12–3.7 eV, indicating strain-enhanced ligand field; (5) structural refinements + DFT-type calculations identify crystal-field strengthening as the stabilization mechanism.

Key Findings:

Thin films retain the high-temperature β-phase across all measured temperatures, in contrast to bulk powders which undergo a first-order β→α transition near 230K.
A low-energy 42 cm⁻¹ cation-dominated phonon softens on cooling (anomalous vs. typical hardening) and saturates between 200–150K — signaling incipient atomic displacements analogous to the suppressed bulk transition.
Co²⁺ crystal-field transitions and charge-transfer excitations (0.12–3.7 eV) blue-shift under strain, indicating a strain-enhanced ligand field that strengthens the crystal field and stabilizes β-phase.

Technical Novelty: First THz-VIS-UV optical spectroscopy of β-CoMoO4 in thin-film form, combined with the identification of microstrain-induced crystal-field strengthening as a thermodynamic control knob that frustrates a first-order structural transition — reframing strain from a passive epitaxial side-effect into an active phase-engineering parameter for thermochromic oxides.

What's New: Prior CoMoO4 literature focused almost exclusively on powders/bulk; this is the first thin-film optical study (THz-VIS-UV) and the first to demonstrate that microstrain can completely suppress a first-order thermochromic transition. The mechanistic link — strain → enhanced ligand field → crystal-field strengthening → β-phase stabilization — is a new conceptual framework for the broader class of thermochromic oxides.

Extension Opportunities:

Systematically vary substrate lattice mismatch (or use piezoelectric substrates) to tune microstrain in situ and map a strain–temperature phase diagram, identifying the critical strain at which the β→α transition reappears.
Apply the same strain-engineering framework to other thermochromic transition-metal molybdates/tungstates (NiMoO4, MnMoO4, CuMoO4) to test generality of crystal-field-strengthening as a phase-stabilization lever.
Exploit the suppressed transition and soft 42 cm⁻¹ phonon for device prototypes — e.g., strain-tunable IR/THz modulators or memristive switches where epitaxial strain gates the thermochromic response.

Replicability: Abstract does not mention released code/data. Reproduction would require thin-film deposition capability (PLD/sputtering for CoMoO4), cryogenic THz–VIS–UV spectroscopy, temperature-dependent XRD, and DFT/lattice-dynamics compute (modest cluster-scale, not GPU-heavy). Moderate experimental barrier; theory is tractable.

Research Gaps:

Quantitative strain–phase diagram: the paper shows suppression but doesn't fully map how much strain is needed, or whether the transition can be re-induced by strain relaxation or opposite-sign strain.
Functional device demonstration: thermochromic and phononic anomalies are characterized spectroscopically, but no integration into a working optical modulator, sensor, or memory device is shown.

🔬 MATERIALS

1. Superconducting properties of Nb${0.85}$Sc${0.15}$ film deposited by magnetron co-sputtering

Authors: Paul Berezhnoy, Anna Elistratova, Zakhar Enbaev... Published: 2026-06-03 | Citations: 0 arXiv | PDF

Research Question: Can Nb-Sc alloy films deposited via magnetron co-sputtering yield superconducting properties suitable for cryogenic electronics, and what composition optimizes critical temperature and current density?

Summary: The paper demonstrates magnetron co-sputtered Nb1-xScx thin films, identifying x≈0.15 as the optimal composition with Tc = 6.35 K and Jc = 2.5 MA/cm². Combined structural, transport, and magnetic characterization establishes NbSc as a promising disordered superconductor for cryogenic electronics, with a short coherence length (10.1 nm) and moderate upper critical field (3.2 T).

Key Results: Synthesized Nb0.85Sc0.15 films with Tc = 6.35 K (optimal at ~15% Sc by Auger spectroscopy), critical current density Jc = 2.5 MA/cm² in microbridges (50 μm × 2 μm × 30 nm), upper critical field Hc2(0) = 3.2 T, electron diffusion coefficient D = 1.1 cm²/s, and Ginzburg-Landau coherence length ξ_GL = 10.1 nm. Structural characterization via XRD and XRR confirmed thickness, phase composition, and crystal structure.

Key Findings:

Optimal Sc concentration of ~15% yields the highest Tc of 6.35 K
High critical current density of 2.5 MA/cm² achieved in patterned microbridges
Hc2(0) = 3.2 T and ξ_GL = 10.1 nm indicate dirty-limit superconductivity suitable for high-kinetic-inductance devices

Technical Novelty: Application of magnetron co-sputtering from separate Nb and Sc targets to produce a controlled-composition NbSc intermetallic thin film with characterization across structural, transport, and magnetic regimes — Nb-Sc is less explored than NbN/NbTi for thin-film superconducting electronics.

What's New: While NbN and NbTiN dominate thin-film superconducting electronics, this work establishes Nb-Sc as a viable alternative intermetallic system, providing a full materials-to-device characterization pipeline with co-sputtering as the synthesis route.

Extension Opportunities:

Fabricate superconducting nanowire single-photon detectors (SNSPDs) using Nb0.85Sc0.15 films, leveraging the short coherence length (10.1 nm) for high kinetic inductance
Explore broader Sc concentration ranges (e.g., 5-30%) with finer granularity to map full Tc-composition phase diagram and search for additional optima
Integrate these films into Josephson junctions or SQUID devices to benchmark against standard NbN/NbTiN for quantum/cryogenic readout applications

Replicability: No mention of public code/data. Reproduction requires a magnetron co-sputtering system with Nb and Sc targets, XRD/XRR for structural characterization, Auger spectroscopy for composition, cryogenic transport measurement setup (sub-7 K), and microfabrication facilities for patterning 2-μm-wide microbridges.

Research Gaps:

No demonstration of integration into functional devices (SNSPDs, Josephson junctions, qubits) — only baseline material properties measured
Limited exploration of composition space and no aging/stability or substrate-dependence studies reported

2. Reinforcement Learning-Enabled Agent for Transmitter Optimization in Digital-Analog Radio-over-Fiber Fronthaul

Authors: Junhao Zhao, Huayuan Qin, Ouhan Huang... Published: 2026-06-03 | Citations: 0 arXiv | PDF

Research Question: How can transmitter parameters in digital-analog radio-over-fiber (DA-RoF) fronthaul systems be optimized autonomously when the parameter space is tightly coupled (rounding factor, scaling factor, geometric shaping factor, pre-equalization taps) and conventional grid search is computationally prohibitive, especially without a differentiable channel model?

Summary: The paper introduces an RL-enabled agent that autonomously tunes four coupled transmitter parameters in digital-analog radio-over-fiber fronthaul systems using only end-to-end SNR feedback, avoiding the need for a differentiable channel model or exhaustive grid search. Experimental validation shows ~2.7 dB SNR gains and support for up to 65536-QAM, enabling higher-order modulation formats for next-generation radio access networks.

Key Results: An RL agent achieved ~2.7-dB SNR improvement over baseline across 1- to 4-order DA-RoF transmission, reaching final SNRs of 35.8 dB, 42.9 dB, 53.8 dB, and 63.2 dB — supporting 1024-, 4096-, 16384-, and 65536-QAM formats respectively. The agent learned end-to-end from SNR feedback through sequential decision making in experimental hardware.

Key Findings:

RL agent delivers consistent ~2.7-dB SNR improvement over baseline across DA-RoF orders 1–4
Final SNRs (35.8 / 42.9 / 53.8 / 63.2 dB) are sufficient to support 1024- through 65536-QAM modulation formats
Model-free RL is viable for online parameter optimization in optical fronthaul where channel models are non-differentiable or unavailable

Technical Novelty: First application of model-free reinforcement learning to DA-RoF transmitter parameter optimization that does not require a differentiable channel model — distinguishing it from gradient-based / autoencoder approaches in optical communication. The agent treats four coupled physical-layer parameters as a joint action space and learns from experimental SNR feedback.

What's New: Prior DA-RoF optimization relied on grid search or assumed differentiable channel models for gradient-based tuning. This work demonstrates a hardware-in-the-loop, model-free RL agent that jointly handles quantization noise, nonlinear distortion, and ISI through learned sequential decisions on experimentally-measured SNR.

Extension Opportunities:

Extend the RL agent to jointly optimize receiver-side DSP (e.g., adaptive equalization, carrier recovery) alongside the transmitter parameters for true end-to-end co-optimization
Investigate transfer learning across different fiber link configurations and DA-RoF orders so a pre-trained agent generalizes without retraining per deployment
Replace the scalar SNR reward with multi-objective rewards (power consumption, latency, BER under fading) to make the agent viable for live O-RAN fronthaul deployments

Replicability: No code or dataset release is mentioned in the abstract. Reproduction requires a DA-RoF optical testbed (laser, modulator, photodetector, ADC/DAC, fiber spool) plus an RL training loop — moderate compute for the agent itself, but substantial specialized optical hardware that limits replicability outside photonics labs.

Research Gaps:

No comparison against other learning-based optimizers (e.g., Bayesian optimization, evolutionary strategies) to isolate RL's specific advantage
Convergence time, sample efficiency, and robustness to channel drift during live operation are not characterized in the abstract

🔥 GitHub Trending

1. REHXZ/AUTOML

⭐ 3 stars | Python

A Agentic Machine Learning, local-first tabular ML discovery workspace. Hand a goal to the AI Autopilot a multi-agent system that runs the full CRISP-DM lifecycle.

agentic-ai ai ai-agents machine-learning workflow

2. roby-uo/daily-report-ai-automation

⭐ 2 stars | Python

日报 AI 自动化框架：AI Agent + Workflow 混合架构，端到端实现从 Excel 面板入职到每日日报自动生成与飞书通知

ai-agent automation data-analysis excel feishu llm

3. therowdygent/monozukuri

⭐ 2 stars | Python

A free, open council of 13 voices grounded in the people who built lean and quality: Deming, Ohno, Shingo, Juran, and more. Coil with patience. Strike once.

ai continuous-improvement deming kaizen lean lean-manufacturing

4. axumquant/axum-trading-platform

⭐ 1 stars | Unknown

Architecture overview of AxumQuant — an autonomous quantitative trading platform: island-model strategy evolution, a 9-stage statistical validation gauntlet (CPCV, Deflated Sharpe), Rust compute kerne

algorithmic-trading genetic-programming machine-learning multi-agent python quantitative-trading

5. Pralishatripathy000/Semantic-cache-llm

⭐ 1 stars | Python

Semantic caching system using Meta Llama 3.3, Groq, FAISS, and Sentence Transformers to reduce LLM latency and API costs through meaning-based query matching.

faiss-vector-database groq llama3 machine-learning semantic-caching transformers

6. Noahpeng829/wikipedia-rag-retrieval-system

⭐ 1 stars | Python

A retrieval-first RAG project demonstrating knowledge base construction, vector search, retrieval evaluation, and error analysis using Chinese Wikipedia.

bge-m3 embeddings faiss fastapi information-retrieval machine-learning

7. Samriti-saini/Multipurpose-Gemini-AI-Chatbot

⭐ 1 stars | Python

A Streamlit-based AI assistant powered by Google's Gemini API featuring chatbot capabilities, image captioning, medical image analysis, question answering, and health diagnosis prediction.

artificial-intelligence chatbot gemini-ai google-generative-ai healthcare image-captioning

8. HayreBuilds/SYNAPSE

⭐ 1 stars | Python

🧠 FAANG-style technology intelligence platform leveraging autonomous AI agents and semantic search.

artificial-intelligence autonomous-agents data-visualization django machine-learning python

9. paudelsandhya/FluRisk-Net

⭐ 1 stars | Python

FluRisk-Net is an interactive, supervised Machine Learning model that automates preliminary influenza screening. Instead of relying on manual rules, it analyzes multi-feature patient data to instantly

flu machine-learning model

10. rishabh-kumar-singh72/telco-churn-analysis

⭐ 1 stars | Jupyter Notebook

End-to-end churn analysis of 7,043 telecom customers — SQL, Python, Stats, ML Model (AUC 0.85), Power BI Dashboard

churn-analysis data-analysis data-science eda logistic-regression machine-learning

11. anmol396/ai-interview-analytics-system

⭐ 1 stars | JavaScript

AI-powered HR Analytics platform for candidate evaluation, insights, and intelligent hiring decisions.

artificial-intelligence candidate-evaluation dashboard fastapi generative-ai hr-analytics

12. Monam1083/AI-Blood-Pressure-Prediction-System

⭐ 1 stars | HTML

An AI-powered Blood Pressure Prediction System using Flask, Machine Learning, BMI Calculator, SQLite, and modern healthcare UI.

ai-project blood-pressure healthcare machine-learning medical-ai python

13. Raghavendra1422/sentiment-analysis-api

⭐ 1 stars | JavaScript

Sentiment Analysis REST API using HuggingFace DistilBERT and FastAPI with single and batch prediction endpoints

deep-learning distilbert fastapi huggingface nlp python

14. aryanap07/Timeless

⭐ 1 stars | Jupyter Notebook

Building Large Language Models from Scratch.

ai artificial-intelligence attention-mechanism deep-learning fine-tuning from-scratch

15. BeamusWayne/trace-vault

⭐ 1 stars | Python

A model-agnostic record/replay reliability gate for tool-using agents — scores Determinism and Faithfulness as two independent, CI-gateable axes.

agent-evaluation agent-harness ai-agents ci-gate determinism llm

Generated by Research Pulse on 2026-06-04 06:07