2026-08-05·Daily digest

🔬 Research Pulse

Daily Digest

August 05, 2026

🤖 AI

🧠 LLMs

1. Test-Time Scaling in Reasoning LLMs: Inference Regimes, Evaluation, and Reproducibility

Authors: Mohsen Hariri, Weicong Chen, Nahal Shahini... Published: 2026-08-04 | Citations: 0 arXiv | PDF

Research Question: How can test-time scaling in reasoning LLMs be systematically formalized, evaluated, and reproduced when current practice conflates structurally distinct inference algorithms under a single scalar 'budget' and reports accuracy without the inference protocol that produced it?

Summary: The paper argues that 'test-time scaling' conflates fundamentally different inference algorithms and provides a unified framework along three axes: a structural taxonomy of three inference regimes over the model's prefix tree, an evaluation profile that separates system-level from candidate-bank metrics, and reproducibility requirements distinguishing exact from distributional replay. It backs this with a release of over 2 billion reasoning traces across broad-knowledge, symbolic, and math benchmarks.

Key Results: The paper develops a three-axis framework: (1) formalizes test-time scaling as budgeted inference over an autoregressive model's implicit prefix tree, distinguishing three regimes — single-trajectory sequential scaling, leaf-level scaling with terminal reduction, and prefix-level scaling; (2) introduces an evaluation profile whose functionals recover/bound common repeated-sampling metrics with protocol-matched compute and uncertainty reporting; (3) specifies reproducibility requirements distinguishing exact replay from distributional reproducibility. Applied to broad-knowledge, symbolic-reasoning, and competition-math benchmarks, and releases 2B+ full reasoning traces with verifier and token-level signals.

Key Findings:

Test-time scaling algorithms fall into three structurally distinct regimes (sequential single-trajectory, leaf-level with terminal reduction, prefix-level search) that cannot be meaningfully compared under a single scalar compute budget
The evaluated object should be the entire inference system, not just the base model — separating end-to-end performance from candidate-bank diagnostics changes what conclusions are supportable
Reproducibility of inference protocols requires distinguishing exact replay from distributional reproducibility, each demanding different artifacts (seeds, sampler state, verifier configs, token-level signals)

Technical Novelty: Prior work reported test-time compute as a single scalar and treated best-of-N, chain-of-thought extension, and tree search as interchangeable. This paper is the first to formalize them as structurally distinct regimes over a prefix tree with different statistical structures and failure modes, and to separate end-to-end system evaluation from candidate-bank diagnostics with an evaluation profile that unifies existing metrics.

What's New: First systematic formalization of test-time scaling as budgeted inference over a prefix tree with three distinct regimes, combined with the largest public release of full reasoning traces (2B+) instrumented with verifier and token-level signals for reproducibility research.

Extension Opportunities:

Build a benchmark harness that auto-detects a paper's inference regime (sequential vs leaf-level vs prefix-level) and re-scores results under protocol-matched compute budgets to enable fair cross-study comparison
Use the released 2B reasoning trace corpus with token-level and verifier signals to train process reward models or distill efficient verifiers that outperform pure majority voting
Develop a hybrid inference controller that dynamically routes between the three scaling regimes per-query based on estimated difficulty and remaining budget, rather than committing to one regime upfront

Replicability: The paper explicitly releases 2B+ full reasoning traces with progressively richer verifier and token-level signals. Reproducibility requirements are formalized (exact replay vs distributional). Reproducing the trace generation itself would require substantial GPU compute (likely thousands of GPU-hours across open-weight reasoning models), but downstream analysis on the released traces should be feasible on modest hardware.

Research Gaps:

Cross-study comparability of test-time scaling results is currently broken because papers report accuracy without specifying inference regime or protocol
The open-weight reasoning ecosystem lacks a shared taxonomy of model-side vs interface mechanisms, making it hard to attribute gains to the model, the sampler, or the verifier

2. Video-DeepResearch: Towards the Next-Generation Multimodal Deepresearch Agent

Authors: Zhen Fang, Yu Zeng, Wenxuan Huang... Published: 2026-08-04 | Citations: 0 arXiv | PDF

Research Question: How can multimodal agents be extended from static images to continuous video streams while overcoming modality bias (agents skipping visual tools for text search) and parametric knowledge leakage (relying on internal memory rather than tool-augmented execution)?

Summary: Video-DeepResearch extends deep research agents to video by decoupling visual perception from web exploration and gating tools stage-wise, forcing genuine cross-frame grounding before search. Trained via SFT+GRPO, the 35B-A3B model sets a new SOTA of 64.0% on the authors' Video-DR-Bench, beating Claude-4.5-Sonnet, GPT-5, and Gemini 2.5 Pro.

Key Results: Video-DeepResearch-35B-A3B achieves 64.0% average accuracy on Video-DR-Bench (200 multi-hop VQA instances), surpassing Claude-4.5-Sonnet (59.0%) by 5.0 points, GPT-5 (52.5%) by 11.5 points, and Gemini 2.5 Pro (57.5%) by 6.5 points. The compact 30B-A3B variant reaches 59.3%, competitive with Claude-4.5-Sonnet, validating the training paradigm at smaller scale.

Key Findings:

Current multimodal agents exhibit two systemic failures: modality bias (bypassing visual tools) and parametric knowledge leakage (answering from memory rather than tool use)
Architecturally enforcing perception-before-exploration via stage-wise tool unlocking yields substantial accuracy gains over unconstrained tool-use agents
GRPO on top of SFT enables autonomous exploration that exceeds the imitation-learning ceiling, and the recipe scales down — 30B-A3B remains competitive with frontier proprietary models

Technical Novelty: Decoupled perception-exploration pipeline with stage-wise tool unlocking that forces exhaustive cross-frame visual grounding before web retrieval is permitted — directly counteracting modality bias by architecturally gating tools. Combined with a two-stage SFT+GRPO training recipe that pushes past imitation-learning ceilings for autonomous exploration.

What's New: Prior multimodal agents treat tools as freely-selectable options, allowing shortcuts through text search. This paper identifies modality bias as a measurable failure mode and introduces a mechanism (stage-wise tool unlocking) that structurally prevents it, paired with GRPO training and a purpose-built video deep-research benchmark.

Extension Opportunities:

Extend the decoupled perception-exploration pipeline to real-time/streaming video for live agent applications (surveillance, sports analytics, live broadcasting) rather than pre-recorded clips
Apply the stage-wise tool unlocking mechanism to other modalities (audio, 3D scenes, sensor fusion) where modality bias likely also causes agents to shortcut through text
Build domain-specific Video-DR variants for verticals like medical imaging video (endoscopy, ultrasound) or scientific video (microscopy, lab recordings) with specialized visual grounding tools

Replicability: Code released at github.com/Osilly/Vision-DeepResearch. Video-DR-Bench (200 instances) appears to be released. Reproducing the 35B-A3B model likely requires multi-node GPU clusters (A3B suggests MoE with ~3B active params), plus significant compute for the GRPO RL stage. SFT reproduction is more tractable for well-funded academic labs.

Research Gaps:

Video-DR-Bench is small (200 instances) and human-AI curated — external validation on larger, independent video reasoning benchmarks is needed
The paper doesn't fully characterize failure modes on very long videos (hours), streaming input, or adversarial visual scenes where perception itself is unreliable

3. ReflectRL: Learning from Golden Negative Trajectories via Reflective-to-Direct Reasoning

Authors: Jinhe Bi, Chennan Zhou, Zengjie Jin... Published: 2026-08-04 | Citations: 0 arXiv | PDF

Research Question: How can on-policy RL training extract useful supervision from expert trajectories that fail on hard problems (Golden Negative Trajectories), instead of discarding them?

Summary: ReflectRL is a plug-and-play framework that turns failed expert-model trajectories ('Golden Negative Trajectories') into training signal by first training the policy to reflect on the flawed reasoning, then transferring that skill back to direct problem solving. It consistently improves reasoning across 9 benchmarks, 4 backbones, and 4 on-policy methods.

Key Results: ReflectRL demonstrates consistent reasoning improvements across 9 benchmarks, 4 LLM backbones, and 4 on-policy training methods with minimal overhead. The paper identifies a 'Reflection Advantage' where reflecting on flawed expert trajectories is easier than solving hard problems from scratch. Specific accuracy deltas are not disclosed in the abstract.

Key Findings:

Expert failures on hard problems still contain exploitable reasoning signal when treated as reflection targets rather than discarded
A 'Reflection Advantage' exists: reflecting on a flawed trajectory is easier and more effective than direct problem solving on hard instances
Reflective reasoning skills can be transferred to direct reasoning via a policy transition stage, improving performance broadly with minimal overhead

Technical Novelty: Reframes failed expert trajectories as reflective training signal rather than negative samples to discard, and introduces a two-stage 'Reflective-to-Direct Policy Transition' that first elicits reflective reasoning from flawed traces then transfers it to direct solve-from-scratch behavior.

What's New: Prior trajectory-guided on-policy methods depend on successful expert demonstrations and discard failures as negatives; ReflectRL is the first to systematically extract training value from expert failures via reflect-then-transfer, exploiting an asymmetry between reflection difficulty and direct solving difficulty.

Extension Opportunities:

Apply ReflectRL to code generation or theorem-proving domains where expert models frequently fail on hard instances and reflective critique is well-defined
Combine ReflectRL with process reward models to grade the quality of reflection steps, potentially amplifying the Reflection Advantage
Extend Reflective-to-Direct Policy Transition to multimodal reasoning (VLMs) where flawed vision-language traces could seed self-correction

Replicability: Abstract does not mention public code or data release. Reproduction likely requires access to a stronger expert LLM for trajectory generation plus standard on-policy RL infrastructure (PPO/GRPO-scale compute across 4 backbones and 9 benchmarks — moderate to high GPU budget).

Research Gaps:

Abstract does not quantify the Reflection Advantage or specify which reflection failure modes are most exploitable
Unclear how quality of the expert model or trajectory selection heuristics affect the gains

🦾 ROBOTICS

1. ETA: A New Agentic Paradigm for Embodied Tasks

Authors: Yitong Chen, Zezheng Huai, Sixian Li... Published: 2026-08-04 | Citations: 0 arXiv | PDF

Research Question: How can embodied robotic systems achieve general-purpose capability, long-horizon controllability, and experiential learning — beyond the limits of current end-to-end observation-to-action architectures whose generalization is bottlenecked by training data coverage?

Summary: ETA proposes a new agentic paradigm for embodied AI that structures a robot as a Planner choosing one Tool call at a time against an Interface and World that return observations, enabling verification, adaptation, and experience reuse. OpenETA is the open-source reference implementation, offering swappable Planners, composable Tools/Skills, auditable memory, replayable trajectories, and a minimal Codex plugin.

Key Results: The paper is primarily a paradigm/framework contribution rather than a benchmark study — the abstract reports no quantitative metrics, datasets, or benchmark scores. It demonstrates feasibility by releasing OpenETA as an open-source implementation with replaceable Planners, composable Tools/Skills, auditable memory, replayable trajectories, and interfaces for both simulation and real robots, plus a lightweight Codex plugin exposing three primitives (observe, mark_point, move_to).

Key Findings:

End-to-end embodied systems generalize only as far as their training data covers and are hard to control/inspect over long horizons.
A digital-agent-style Planner-Tool-World loop can be transplanted into robotics to give verifiability, adaptivity, and reusable experience.
A very small action interface (observe, mark_point, move_to) is sufficient to expose ETA as a lightweight Codex plugin.

Technical Novelty: Reframes embodied control as a digital-agent-style Planner→Tool→Interface→World loop with one-tool-at-a-time execution, verifiable outcomes, and reusable experience — explicitly separating high-level planning from low-level skills, rather than the dominant end-to-end observation-to-action VLA paradigm. The minimal three-primitive Codex plugin (observe/mark_point/move_to) is a notably austere action interface.

What's New: Unlike prior VLA/end-to-end policies that map observations directly to actions, ETA imports the LLM-agent architecture (planner + tools + memory + trajectories) into embodied control, and ships a modular open-source stack designed to be swappable across planners, skills, simulators, and real robots.

Extension Opportunities:

Build domain-specific Tool/Skill libraries on top of OpenETA (e.g., kitchen manipulation, warehouse picking) and measure whether the Planner-Tool decomposition improves long-horizon success vs. an end-to-end VLA baseline.
Swap in different Planner backbones (Claude, GPT, open-weight models) via the replaceable-Planner interface and benchmark planning quality, latency, and cost on standard embodied benchmarks.
Exploit the auditable memory + replayable trajectories to build an experience-replay fine-tuning pipeline that distills successful/failed rollouts back into either the Planner or a learned skill policy.

Replicability: Yes — OpenETA is released as open-source with simulation and real-robot interfaces, replaceable components, and replayable trajectories. Compute requirements are not specified in the abstract but should be modest for the Planner (LLM API calls) plus whatever the underlying robot/simulator demands; reproducing real-robot results requires physical hardware.

Research Gaps:

The abstract provides no quantitative comparison against end-to-end VLA baselines or established embodied benchmarks.
How the framework's experiential learning loop actually updates future behavior (in-context memory vs. weight updates) and its scaling behavior are not characterized.

2. EvoHIL: Self-Evolving Reward and Flow-Matched Policy Optimization for Robust Human-in-the-Loop Reinforcement Learning

Authors: Shuoqin Zhang, Tongtong Cheng, Xiru Gao... Published: 2026-08-04 | Citations: 0 arXiv | PDF

Research Question: How can human-in-the-loop reinforcement learning for contact-rich robotic manipulation remain robust when the reward model, action sampling, and visual perception all degrade under real-world scene and lighting shifts?

Summary: EvoHIL is a unified human-in-the-loop RL framework for contact-rich robot manipulation that jointly adapts the reward classifier, action generator, and visual encoder across three staged mechanisms. It combines self-evolving reward learning, flow-matched action chunk generation, and retention-aware offline fine-tuning to remain robust to appearance shifts without requiring additional robot interaction.

Key Results: EvoHIL was evaluated across six manipulation tasks on two robot arms (Franka FR3 and SO-101) under a controlled lighting shift, demonstrating improvements over HIL-RL and imitation baselines across four metrics: task success rate, agreement with human-confirmation labels, motion smoothness, and completion time. Specific numeric deltas are not stated in the abstract.

Key Findings:

A single success classifier can be self-improved online using only human-confirmed positives and weak negatives, avoiding manual reward re-engineering
Flow matching over action chunks conditioned on executed prefixes produces smoother, temporally consistent motions versus independently sampled actions
Offline replay of relit interaction data with actor-critic anchoring adapts vision-based policies to lighting shifts without new robot rollouts

Technical Novelty: The unified three-stage adaptation loop is new: (1) self-evolving reward using human-confirmed positives plus provisional weak negatives to update the success classifier online, (2) flow-matching-based Action Flow Stabilization that conditions on executed action prefixes for temporal coherence, and (3) retention-aware offline fine-tuning on relit replay data anchored to prior actor-critic behavior — no additional robot interaction needed for visual domain shift.

What's New: Prior HIL-RL work typically treats reward learning, action sampling, and visual robustness as separate problems. EvoHIL is the first to couple all three into a single staged human-in-the-loop pipeline, and it introduces flow matching as the action-generation mechanism in this setting.

Extension Opportunities:

Replace the controlled lighting shift with broader domain randomization (backgrounds, occlusions, camera pose) to test whether retention-aware fine-tuning generalizes beyond relighting
Substitute flow matching in AFS with diffusion policies or consistency models to benchmark temporal-coherence tradeoffs against inference latency on real hardware
Extend the self-evolving reward classifier to multi-stage tasks by learning subgoal-conditioned success signals rather than a single binary classifier

Replicability: A project page is provided (anonymous4366.github.io/EvoHIL) suggesting anonymized supplementary material, but code/data availability is not confirmed in the abstract. Reproduction requires access to Franka FR3 and/or SO-101 arms, controlled lighting rigs, and moderate GPU compute for flow-matching policy training — likely a single high-end GPU for policy and reward updates.

Research Gaps:

Evaluation is limited to a single controlled lighting shift — robustness to more aggressive distribution shifts (novel objects, viewpoints, dynamics) is untested
The abstract omits concrete success-rate numbers and does not compare against recent diffusion-policy or VLA baselines

3. Designing Social Robots for Inclusive Child Wellbeing Assessment: Insights from Communities Supporting Developmental Language Disorder and Forced Migration

Authors: Fethiye Irmak Dogan, Yue Lou, Alva Markelius... Published: 2026-08-04 | Citations: 0 arXiv | PDF

Research Question: How can social robots be designed to support inclusive, appropriate, and ethically acceptable wellbeing assessment for children with communication barriers — specifically those with Developmental Language Disorder (DLD) and forced migration backgrounds — where traditional self-report questionnaires fail due to language comprehension demands?

Summary: The paper uses candidate child-robot interaction activities as design probes in focus groups with parents and professionals supporting children with DLD and forced migration backgrounds, then applies thematic analysis to derive ethical and inclusive design recommendations for robot-mediated wellbeing assessment. It contributes a design-guidance framework rather than a working system, foregrounding communication-barrier populations that have been largely excluded from prior child-robot HRI wellbeing research.

Key Results: The paper is qualitative/formative — no benchmarks or quantitative metrics reported in the abstract. It demonstrates via thematic analysis of focus groups with parents and professionals that design considerations cluster around four themes (robot role/capabilities, interactional dynamics, individual differences, child agency) plus population-specific concerns, yielding a set of derived ethical and inclusive design recommendations.

Key Findings:

Design considerations organize around four cross-cutting themes: robot role and capabilities, interactional dynamics, individual differences, and child agency
Population-specific considerations emerge distinctly for DLD versus forced-migration children, shaped by their differing communication needs and lived experiences
Stakeholders (parents/professionals) emphasize ethical acceptability and child agency as prerequisites — not add-ons — for inclusive robot-mediated assessment

Technical Novelty: Not a technical/algorithmic contribution — novelty is methodological: using child-robot interaction activities as design probes in focus groups with two distinct communication-barrier communities (DLD and forced migration) simultaneously, and synthesizing cross-cutting plus population-specific inclusive-design guidance.

What's New: Prior HRI wellbeing research has rarely centered children with communication barriers; this work is novel in jointly engaging DLD and forced-migration communities through participatory focus groups with design probes, and in translating the findings into a concrete inclusive-design recommendation set.

Extension Opportunities:

Prototype and empirically evaluate a robot-mediated assessment session with DLD or refugee children implementing the derived design recommendations, measuring engagement and validity against clinician judgment
Develop multimodal (gesture, gaze, drawing, picture-card) input pipelines on platforms like NAO/Pepper that reduce verbal demand, and benchmark comprehension outcomes against verbal questionnaires
Build a co-design toolkit or open library of the candidate interaction activities used as design probes so other researchers can replicate the focus-group methodology across additional communication-barrier populations (e.g., autism, hearing impairment)

Replicability: No code or datasets mentioned. Reproducing would require IRB-approved focus groups with parents/professionals in the two communities, a social robot platform (e.g., NAO/Pepper) for design probes, and qualitative analysis tooling — compute is negligible; the cost is participant recruitment and ethics approval.

Research Gaps:

No empirical validation with children themselves — recommendations are derived from proxy stakeholders (parents/professionals), not children's direct interactions
Lack of implemented, evaluated robot prototypes that operationalize the recommendations and measure assessment validity against clinical ground truth

💻 COMPUTE

1. Real-time decoding of quantum error correction codes using high-performance computing

Authors: Lingling Lao, Qiang Wang, Yuanqi Liu... Published: 2026-08-04 | Citations: 0 arXiv | PDF

Research Question: How can quantum error correction decoders meet the microsecond-scale real-time latency budget required to avoid syndrome backlog while scaling to large numbers of logical qubits, without being bottlenecked by the classical control stack?

Summary: The paper introduces THQLink, an HPC-integrated architecture that offloads quantum error correction decoding to a supercomputer via the TH-Express network, achieving 2.944 μs round-trip latency. Combined with a parallel-window scheduling strategy, it enables real-time (1 μs/round) matching-based decoding of surface codes up to distance 19 on CPUs, showing that quantum-centric supercomputers can serve as a scalable decoding backend for fault-tolerant quantum computing.

Key Results: Demonstrated THQLink architecture achieving 2.944 μs average round-trip latency between HPC and QPU control system (built on TH-Express interconnect), with only 130 ns incremental overhead per additional network hop. Using a parallel window strategy on CPUs with a matching-based decoder, they achieved real-time decoding at 1 μs per QEC round for surface codes up to distance d=19.

Key Findings:

Round-trip QPU↔HPC latency of 2.944 μs average with 130 ns per additional hop over TH-Express
Real-time decoding at 1 μs per QEC round achieved for surface codes up to distance d=19 using CPU-based matching decoders
Parallel window strategy is the key algorithmic ingredient that keeps decoding within budget as code distance scales
Architecture is technology-agnostic and can be adapted to different QPU control stacks

Technical Novelty: Prior real-time decoding work relied on custom FPGA/ASIC decoders tightly coupled to a single QPU. This paper instead offloads decoding to general-purpose HPC via a low-latency TH-Express network fabric (THQLink), combined with a parallel-window scheduling strategy that keeps CPU-based matching decoders within the 1 μs/round budget even at d=19.

What's New: First demonstration that a general-purpose HPC system, connected to QPU control electronics via a production supercomputer interconnect (TH-Express), can meet the microsecond-latency real-time decoding budget — an alternative to bespoke FPGA/ASIC decoders that also unlocks HPC resources for hybrid quantum-classical workloads.

Extension Opportunities:

Port the parallel-window matching decoder to GPUs or FPGAs within the THQLink framework to push decoding beyond d=19 or reduce per-round latency well below 1 μs
Adapt the TH-Express-based link to alternative qubit modalities (trapped ion, neutral atom, photonic) and benchmark control-stack latency variability across technologies
Integrate ML-based or neural network decoders into the same HPC pipeline and compare accuracy/latency tradeoffs against matching-based decoders on identical hardware

Replicability: Abstract does not mention public code or data release. Reproduction requires access to TH-Express interconnect hardware (Tianhe-family supercomputer fabric) and a QPU control system, plus an HPC cluster — making full replication effectively limited to groups with TH-series HPC access, though the parallel-window decoding strategy alone could be reproduced on commodity CPUs.

Research Gaps:

No evaluation beyond surface codes or matching decoders — behavior with LDPC codes or neural decoders is unaddressed
Scaling to many logical qubits simultaneously (rather than single-code demonstrations) and the resulting HPC resource contention is not characterized

2. When Does Disaggregation Pay? Simulating Prefill--Decode--Attention--FFN Specialization for Agentic LLM Inference

Authors: Przemyslaw Forys, Haoran Wu, Can Xiao... Published: 2026-08-04 | Citations: 0 arXiv | PDF

Research Question: In heterogeneous LLM inference systems, what is the optimal hardware specialization for each serving stage (prefill, decode, attention, FFN) under agentic workloads, and when does disaggregation actually pay off?

Summary: The paper introduces HeteroPanacea, a simulation framework for disaggregated LLM inference that jointly explores quantization, parallelization scheduling, and NPU heterogeneity across prefill, decode, attention, and FFN stages. It shows PD disaggregation raises throughput up to 75% on current GPUs, and 4-way PDAF disaggregation with custom NPUs is the most consistent gain across model architectures for agentic workloads.

Key Results: Using their HeteroPanacea simulator, the authors show Prefill-Decode (PD) disaggregation yields up to 75% higher serving throughput vs. traditional homogeneous GPU serving. They further demonstrate that 4-way Prefill-Decode-Attention-FFN (PDAF) disaggregation with custom NPUs is the most consistent throughput booster across model architectures, supported by ablations linking gains to model architecture properties.

Key Findings:

Prefill-Decode disaggregation delivers up to 75% throughput improvement over homogeneous GPU serving
4-way PDAF disaggregation is the most consistent throughput winner across model families when custom NPUs are assumed
Model architecture (attention/FFN ratio, KV-cache profile) materially changes which disaggregation split pays off, as shown via ablation studies

Technical Novelty: First simulation framework combining three axes simultaneously: disaggregated quantization, automated intra/inter-device parallelization scheduling, and PDAF NPU architectural heterogeneity — moving beyond prior PD-only disaggregation studies (e.g., DistServe, Splitwise) to a 4-way stage split with per-stage hardware co-design.

What's New: Prior work stopped at PD splits on homogeneous GPUs; this paper co-simulates architectural heterogeneity, quantization, and scheduling in one framework and extends the split to attention and FFN as first-class disaggregated stages targeting future NPU platforms.

Extension Opportunities:

Extend HeteroPanacea to model real emerging hardware (Vera-Rubin GPUs + Groq LPUs) with measured power/thermal envelopes and validate simulated throughput against physical testbeds
Add MoE-specific disaggregation dimensions (expert routing as a 5th stage) and evaluate PDAEF partitioning for sparse frontier models like DeepSeek-V3 or Mixtral variants
Layer in economic modeling — $/token and TCO across disaggregated topologies — to convert throughput wins into deployment decision rules for inference providers

Replicability: The abstract does not mention public code or datasets. Reproduction would require rebuilding the simulator (moderate SW engineering effort) plus access to model architecture specs; no GPU hours needed for the simulation itself, though validation against real hardware would demand a heterogeneous cluster.

Research Gaps:

No physical-hardware validation — results are simulator-based and hinge on modeling fidelity of hypothetical NPUs
Limited treatment of interconnect bottlenecks (KV-cache and activation transfer costs between disaggregated stages) which typically dominate real deployments

3. Heterogeneity-Aware Microscaling for Efficient Low-Bit LLM Inference

Authors: Junyi Luo, Xinting Jiang, Tai-Hao Wen... Published: 2026-08-04 | Citations: 0 arXiv | PDF

Research Question: How can microscaling (MX) low-bit LLM inference formats close the accuracy gap of MXFP4 without increasing equivalent bit width, given that fixed element formats and fixed precision-recovery schemes fail to capture heterogeneity that exists both across blocks and across operand types (weights vs. activations)?

Summary: AdaMX is a heterogeneity-aware microscaling format and 22nm accelerator that adapts the precision-recovery scheme per block and the representation per operand at fixed equivalent bit width. It nearly closes the MXFP4 accuracy gap (83%/82% recovery on commonsense/MMLU) for ~1% extra system energy, and its lower-EBW mode simultaneously improves accuracy, memory, and energy over the baseline.

Key Results: AdaMX, prototyped on 22nm FD-SOI, adds only ~1% system energy over an MXFP4 baseline while removing 83% of MXFP4 accuracy loss on commonsense and 82% on MMLU, and 43%/27% of NVFP4 loss, across LLMs from 3B to 70B. On Gemma-4 12B multimodal, it beats MXFP4 on all four vision-language benchmarks and retains up to 96% of FP16 accuracy. A lower-EBW operating point improves accuracy over baseline while cutting memory footprint and energy.

Key Findings:

Quantization heterogeneity is bi-level: preferred element format and precision-recovery scheme vary across blocks, and weights vs. activations need different encodings
Per-block adaptive selection recovers 83% of MXFP4 commonsense loss and 82% of MMLU loss across 3B–70B LLMs at ~1% energy overhead
The approach generalizes to multimodal: Gemma-4 12B keeps up to 96% of FP16 accuracy and beats MXFP4 on all four vision-language benchmarks tested

Technical Novelty: Prior MX formats (MXFP4, NVFP4) fix a single element format and a single precision-recovery scheme across all blocks and both operands. AdaMX makes two choices adaptive: (1) per-block selection of the precision-recovery scheme, and (2) per-operand selection of the representation (weights vs. activations encoded differently), while holding EBW constant. It also unifies two block sizes in one hardware design to trade accuracy vs. storage.

What's New: First MX variant to adapt both the precision-recovery scheme (per block) and the representation (per operand) simultaneously, backed by a taped-out accelerator that supports two block sizes in one datapath — moving beyond the one-size-fits-all assumption baked into MXFP4 and NVFP4.

Extension Opportunities:

Extend per-block scheme selection to training/fine-tuning (QAT) rather than only inference, learning heterogeneity assignments jointly with weights
Apply the operand-asymmetric encoding idea to KV-cache quantization, where activation-like statistics dominate and compression matters most for long-context serving
Port the decoder + compute unit to an open-source RTL/FPGA release and evaluate on emerging MoE and diffusion-LM workloads whose block statistics differ from dense transformers

Replicability: No code or dataset release is mentioned in the abstract. Reproducing the accelerator requires 22nm FD-SOI tape-out or FPGA emulation of the decoder/compute unit/quantization logic. Reproducing the accuracy numbers requires GPU inference on models from 3B to 70B (multi-A100/H100 for 70B) plus Gemma-4 12B multimodal evals on four VLM benchmarks.

Research Gaps:

No evaluation of training or fine-tuning under AdaMX — only inference is characterized
Silicon comparison is limited to an 'otherwise identical' MXFP4 baseline; head-to-head area/energy against NVFP4 or INT4 accelerators is not reported in the abstract

⚡ ENERGY

1. Circuit-Level Design, Modeling, and On-Wafer Characterization of a Coplanar THz Optoelectronic Mixer

Authors: S. Islam, V. Merupo, C. Chamek... Published: 2026-08-04 | Citations: 0 arXiv | PDF

Research Question: How can photoconductive terahertz heterodyne mixers be designed, modeled, and characterized as true monolithic integrated circuits (TMICs) using standard RF metrics, rather than the ad-hoc figures of merit typical of prior photonic-mixer demonstrations?

Summary: The paper presents the first coplanar THz photoconductive heterodyne mixer designed and validated as a monolithic integrated circuit using standard RFIC metrics. A distributed circuit model incorporating the photoconductor and CPW accesses matches on-wafer measurements to within ±1 dB conversion loss across 1–500 GHz, providing a reusable design methodology for integrated THz photonic mixers.

Key Results: The authors built a coplanar-waveguide TMIC combining a microcavity photoconductor with a Ti shunt resistor, MIM DC-blocking capacitor, and RF/IF decoupling capacitor. On-wafer VNA characterization up to 500 GHz validated the circuit model. Under illumination: input reflection coefficient <-10 dB up to 215 GHz and <-6.5 dB up to 500 GHz. Measured vs simulated conversion loss agrees within ±1 dB over 1–500 GHz. Conversion loss: 22.5–25.5 dB below 320 GHz, rising to 26–34 dB up to 500 GHz.

Key Findings:

Broadband matching achieved: |S11| < -10 dB to 215 GHz and < -6.5 dB to 500 GHz under illumination
Conversion loss stays in a 22.5–25.5 dB band below 320 GHz, degrading to 26–34 dB at 500 GHz
Circuit-level model incorporating photoconductor + distributed CPW reproduces measurements within ±1 dB across three decades of frequency

Technical Novelty: First photoconductive heterodyne mixer characterized as a true TMIC using standard RFIC metrics (S-parameters, isolation, matching, conversion loss) rather than optics-centric figures of merit, plus a validated distributed circuit model that couples the photoconductor to CPW accesses.

What's New: Prior photoconductive heterodyne work reported optical/optoelectronic metrics in isolation; this work reframes the device as an RFIC component, adds passive matching (Ti shunt, MIM caps) inside the coplanar network, and demonstrates a physics-based model that predicts S-parameters and conversion loss simultaneously up to 500 GHz.

Extension Opportunities:

Integrate the TMIC with an on-chip THz antenna or waveguide transition to enable a fully packaged wireless receiver front-end
Sweep photoconductor materials (LT-GaAs vs InGaAs/InAlAs vs Ge) within the same coplanar matching topology to benchmark conversion-loss floors
Extend the circuit-level model to include noise (shot, thermal, generation-recombination) so noise figure — not just conversion loss — becomes a design target

Replicability: No code or dataset is mentioned; reproduction requires cleanroom fabrication of a microcavity photoconductor with Ti/MIM CPW passives, a fs pulsed or two-tone laser source, and an on-wafer VNA extending to 500 GHz — a well-equipped THz photonics lab, not a solo builder.

Research Gaps:

No noise-figure characterization or model — conversion loss alone is insufficient for receiver-level system design
Illumination conditions and LO power scaling are not framed as design variables in the circuit methodology

2. From phase transformation to amorphization: damage accumulation in Yb-implanted \b{eta}-Ga2O3

Authors: Joanna Matulewicz, Renata Ratajczak, Ewa Grzanka... Published: 2026-08-04 | Citations: 0 arXiv | PDF

Research Question: How does β-Ga2O3 accumulate radiation damage under heavy-ion (Yb) implantation across a wide fluence range, and does the γ-Ga2O3 spinel phase truly exhibit the radiation stability claimed in prior reports?

Summary: The paper delivers a comprehensive, multi-technique characterization of damage accumulation in Yb-implanted β-Ga2O3 across ~3 orders of magnitude in fluence, revealing a strain-driven β→γ phase transformation at ~0.4 dpa, subsequent defect buildup within the γ phase, and ultimate surface amorphization at ~7 dpa. The results overturn prior claims of γ-Ga2O3 radiation robustness and highlight the strong ion-specific and strain-mediated nature of Ga2O3's radiation tolerance.

Key Results: Using RBS/c, PAS, HRTEM, and HRXRD on differently oriented β-Ga2O3 single crystals implanted with Yb ions from 5×10¹² to 1×10¹⁶ cm⁻² (0.04–74 dpa), the authors demonstrate a staged damage evolution: (1) strain-driven β→γ (monoclinic→defective spinel) phase transformation at ~0.4 dpa accompanied by strain relaxation, (2) subsequent defect accumulation within γ-Ga2O3 producing a dip in the damage curve and stacking faults in the subsurface region, (3) surface amorphization at ~7 dpa, and (4) progressive expansion of the amorphous layer that replaces the γ phase at higher fluence.

Key Findings:

A critical strain threshold of ~0.4 dpa triggers a monoclinic β → defective-spinel γ phase transformation, with concurrent strain relaxation
The γ-Ga2O3 phase is NOT radiation-stable as previously reported: continued irradiation produces stacking faults, a damage-curve dip, and rising vacancy-type defect densities with strong depth dependence
Surface amorphization initiates at ~7 dpa and progressively consumes the transient γ phase at higher fluence, up to 74 dpa

Technical Novelty: First multi-technique (RBS/c + positron annihilation spectroscopy + HRTEM + HRXRD) mapping of Yb-implanted β-Ga2O3 across the full damage regime (0.04–74 dpa), resolving vacancy-type defect evolution with depth dependence and directly refuting the assumed radiation robustness of the transient γ-Ga2O3 phase.

What's New: Prior work treated γ-Ga2O3 as a radiation-stable end-state; this study uses complementary depth-resolved probes to show γ itself accumulates defects and eventually amorphizes, and it identifies lattice strain (not just displacement damage) as the trigger of the β→γ transition.

Extension Opportunities:

Repeat the fluence sweep with lighter ions (e.g., He, Ar) or different implant energies to isolate ion-mass vs. electronic-stopping contributions to the 0.4 dpa β→γ threshold
Perform in-situ TEM ion irradiation to capture real-time nucleation of the γ phase and stacking fault formation, tying atomic-scale kinetics to the RBS/c damage curve dip
Use MD or DFT simulations of Yb cascades in β-Ga2O3 to predict the critical strain energy that triggers the monoclinic→spinel transition, then test the model on other oriented crystals and dopants

Replicability: No code/data availability mentioned in the abstract. Reproduction requires an ion implanter capable of Yb at fluences up to 10¹⁶ cm⁻², oriented β-Ga2O3 single crystals, and access to RBS/channeling, positron annihilation spectroscopy, HRTEM, and HRXRD instrumentation — a well-equipped materials/ion-beam lab (e.g., HZDR-class facility).

Research Gaps:

Ion-species and energy dependence of the 0.4 dpa β→γ threshold remains uncharacterized
Lack of atomistic/mechanistic modeling linking strain energy to the monoclinic→spinel transformation and to depth-dependent vacancy populations

3. Machine Learning-Guided Screening of Advantageous Solvents for Solid Polymer Electrolytes in Lithium Metal Batteries

Authors: Jiadong Shen, Junjie Chen, Xiaosa Xu... Published: 2026-08-04 | Citations: 0 arXiv | PDF

Research Question: How can we systematically screen trace residual solvents in solid polymer electrolytes (SPEs) to simultaneously optimize ionic conductivity, transference number, and electrochemical stability for lithium metal batteries, given the vast solvent chemical space and lack of general screening criteria?

Summary: The authors trained ML models on a ~10,000-solvent high-throughput DFT dataset to identify optimal trace-residual solvents for solid polymer electrolytes, deriving a universal criterion linking electronic and macroscopic descriptors. They validated the top hit (N-methoxy-N-methyl-2,2,2-trifluoroacetamide) in a PVDF-HFP SPE, achieving 5.5×10⁻⁴ S/cm conductivity, 0.78 transference number, and strong cycling performance in both LFP and high-nickel NCM lithium metal cells.

Key Results: Built an ML model on ~10,000-solvent DFT dataset linking electronic (HOMO/LUMO) and macroscopic (dielectric constant, dipole moment, polarizability) properties. Identified N-methoxy-N-methyl-2,2,2-trifluoroacetamide, which in a PVDF-HFP matrix delivered: 4.5 V electrochemical window, 5.5×10⁻⁴ S/cm ionic conductivity at 30°C, 0.78 Li+ transference number, 86.7% capacity retention over 500 cycles with LiFePO4, and 98.7% retention after 200 cycles at 2C with LiNi0.9Co0.05Mn0.05O2 — outperforming TFDMA, DMF, NMP, and DMSO baselines.

Key Findings:

HOMO/LUMO combined with dielectric constant, dipole moment, and polarizability form a transferable descriptor set for predicting SPE solvent performance
Trace fluorinated methoxy-amide solvent achieves a 4.5 V window and 0.78 Li+ transference number — high for SPEs, where transference numbers of 0.2–0.5 are typical
The SPE cell retains 86.7% capacity over 500 cycles (LFP) and 98.7% after 200 cycles at 2C on high-nickel NCM (Ni 0.9), demonstrating compatibility with high-energy cathodes

Technical Novelty: First universal ML-driven criterion that jointly maps electronic-structure descriptors (HOMO/LUMO) and bulk dielectric/dipole/polarizability properties to SPE performance targets at ~10K solvent scale — prior work relied on single-property heuristics or narrow trial-and-error screening. Also novel: the specific fluorinated methoxy-amide solvent identified (N-methoxy-N-methyl-2,2,2-trifluoroacetamide) as an SPE additive.

What's New: Combines large-scale DFT-based descriptor generation with ML screening to establish a general, physics-grounded criterion for SPE solvent selection, rather than relying on narrow empirical trial-and-error. Discovers a non-obvious fluorinated methoxy-amide additive that outperforms conventional polar aprotic solvents (DMF, NMP, DMSO).

Extension Opportunities:

Apply the same electronic+macroscopic descriptor ML framework to screen solvents for sodium/potassium/zinc-ion SPE systems where the chemistry is analogous but the target ion changes coordination requirements
Extend the ~10K solvent dataset with binary/ternary solvent mixtures and train a mixture-composition model, since real electrolytes often use co-solvent formulations for synergistic effects
Integrate the DFT descriptors with graph neural networks operating directly on solvent SMILES to enable de novo solvent design (generative screening) rather than screening existing molecules

Replicability: Abstract does not mention code or dataset release. Reproduction would require: (1) high-throughput DFT compute (~10K solvent geometry optimizations + electronic property calculations, likely thousands of CPU-hours on a cluster), (2) standard ML tooling (scikit-learn/XGBoost class, modest GPU or CPU), and (3) wet-lab battery fabrication and cycling for validation. Compute for the ML step is trivial; DFT and experimental validation are the real cost.

Research Gaps:

Long-term (>1000 cycle) stability, calendar aging, and thermal/safety behavior of the identified SPE are not reported
Mechanistic understanding of why the specific descriptor combination predicts performance — the ML criterion is empirical and the SEI-level explanation of the trifluoro-methoxy-amide's role remains to be established

🔬 MATERIALS

1. Epitaxial SiGeSn alloys for CMOS-compatible thermoelectric devices

Authors: Patrizio Graziosi, Damiano Marian, Andrea Tomadin... Published: 2026-08-04 | Citations: 0 arXiv | PDF

Research Question: Can heteroepitaxial SiGeSn alloys serve as CMOS-compatible thermoelectric materials for on-chip energy harvesting, bridging the gap between mainstream microelectronics and thermoelectric device integration?

Summary: The paper evaluates heteroepitaxial SiGeSn alloys as a CMOS-compatible thermoelectric material by combining experimental 3-ω thermal conductivity measurements (~1-2 W/m·K) with Boltzmann transport calculations. They predict ZT > 1 and power factors ~20 μW/cm·K² at 300-400 K for both n- and p-type, positioning SiGeSn as a promising material for on-chip Green-IT thermoelectric applications.

Key Results: Experimentally measured low lattice thermal conductivity of ~1-2 W/m·K in SiGeSn/Ge/Si layers via 3-ω measurements. Boltzmann transport equation calculations (accounting for inter-valley scattering) predict ZT > 1 for both p- and n-type material at 300-400 K operating temperatures. Power factor reaches ~20 μW/cm·K² in the same range.

Key Findings:

SiGeSn/Ge/Si layers exhibit low lattice thermal conductivity of ~1-2 W/m·K, favorable for thermoelectric efficiency
Predicted ZT exceeds 1 for both p-type and n-type SiGeSn in the 300-400 K on-chip temperature range
Power factor reaches competitive values of ~20 μW/cm·K², making SiGeSn viable for real device applications

Technical Novelty: First systematic evaluation of the ternary SiGeSn alloy system as a thermoelectric material using Boltzmann transport with proper treatment of inter-valley scattering peculiar to this multi-valley band structure, combined with experimentally-measured thermal conductivity as input.

What's New: Introduces SiGeSn — a relatively new Group IV ternary alloy — as a CMOS-compatible thermoelectric platform, uniquely combining silicon-fab compatibility with performance metrics (ZT > 1) that traditionally required non-CMOS materials like Bi2Te3.

Extension Opportunities:

Fabricate and experimentally validate actual SiGeSn thermoelectric devices to confirm the predicted ZT > 1 values under real operating conditions
Explore compositional tuning of Si/Ge/Sn ratios to optimize the power factor and identify the sweet spot between electronic transport and phonon scattering
Design and prototype on-chip thermoelectric generators that harvest waste heat from CMOS logic circuits at typical junction temperatures (300-400 K)

Replicability: No explicit code/data release mentioned. Reproduction requires: (1) SiGeSn epitaxial growth capability (specialty MBE/CVD), (2) 3-ω thermal measurement setup, (3) Boltzmann transport equation solver with multi-valley scattering — moderate compute (workstation-scale DFT + BTE, not HPC-intensive).

Research Gaps:

Experimental device-level validation of the predicted ZT > 1 is not yet demonstrated — the figure of merit is theoretical
Long-term stability, doping strategies, and manufacturability at scale in production CMOS lines remain unexplored

🔥 GitHub Trending

1. Sparkfetch/sparkfetch

⭐ 37 stars | TypeScript

🔥 Turn any URL into clean, structured, LLM-ready content. The open-source web fetching & extraction API.

ai api content-extraction data-extraction html-to-markdown llm

2. Sentry-LLM/SentryLLM

⭐ 34 stars | TypeScript

AI Security Monitor — Real-time threat detection, prompt injection defense, and behavioral analysis for LLM-powered systems

ai-security jailbreak-detection llm llm-security monitoring openai

3. danielbergholz/phoenix-ai-setup

⭐ 4 stars | Elixir

Six steps to make a new Phoenix app work well with AI coding agents: pin the toolchain, turn mix precommit into a real gate, split the generated AGENTS.md into trigger-loaded skills, and keep it all f

agents-md ai-agents ai-coding claude-code credo developer-tools

4. ishandutta2007/Awesome-Masked-Language-Modeling

⭐ 2 stars | Unknown

a machine learning training method that hides random words in a text and teaches an AI model to guess the missing words

ai-theory machine-learning masked-language-modeling nlp

5. yamancan/visual-scene-ontology

⭐ 2 stars | Python

An image is not a sentence: a VLM's description of one is unvalidated prose. VSON makes every claim about an image — object, property, spatial relation, action — a graph assertion SHACL can reject, so

computer-vision knowledge-graph llm ontology owl2 penman

6. j4oc/Disor

⭐ 2 stars | Python

بوت إدارة سيرفرات بالذكاء الاصطناعي 126 مهارة بالعامية - Discord

admin-bot ai ai-agents arabic discord-bot discord-py

7. ishandutta2007/Quantization-Aware-Training

⭐ 1 stars | Python

Quantization-Aware Training (QAT) is a method that models low-precision arithmetic (like INT4 or INT8) during training or fine-tuning

ai-theory deep-learning machine-learning quantization-aware-training

8. dvharshavardhan/Smart-Health-Prediction-System

⭐ 1 stars | Python

AI-powered Healthcare Prediction System built with Flask, Scikit-learn, SQLite, REST APIs, Chart.js dashboards, Docker, and PDF reporting.

anaytics-dashboard chartjs docker flask healthcare machine-learning

9. BhairavThakare08/nyc-airbnb-room-type-predictor

⭐ 1 stars | Jupyter Notebook

ML-powered NYC Airbnb listing classifier — Random Forest model (85.1% accuracy) served via FastAPI, with an animated skyline UI that visualizes prediction confidence in real time.

data-science fastapi javascript machine-learning portfolio-project python

10. abhisheknishad23/Sentiment-Analysis

⭐ 1 stars | Jupyter Notebook

An end-to-end NLP Sentiment Analysis web application built with interactive data visualization and single review predictions.

check-resource dashboard flask-application machine-learning nlp-machine-learning review-system

11. AdityaPagare619/llm-academy-course

⭐ 1 stars | HTML

LLM Academy — build a transformer from nothing. 14 hands-on modules, construction labs, Builder's Wall rewards. NOT affiliated with the LEGO Group.

ai education interactive-course lego llm machine-learning

12. collinear-ai/reading-list

⭐ 1 stars | Unknown

The Collinear AI reading list: synthetic data, verifiers, agent environments, evaluation science, user simulation, and self-improving systems

ai-agents evals llm reading-list synthetic-data

13. Nolane-x/Nolane-Axiom

⭐ 1 stars | Python

Proof-carrying investigation for AI agents: 32 composable skills for causal debugging, evidence provenance, integrity, and release-grade QA.

agent-skills ai-agents ai-safety chatgpt codex debugging

14. zhouyu033/medical_AI_security_experimentcode_v2

⭐ 1 stars | Python

[Version 2] OPEN AI Team Paper: A Unified Multimodal Privacy-Enhancing Framework for Secure Medical Data Processing in Large Language Models

llm llm-security medical-ai publicsafety

15. ZHAO0424/unity-cli-skill

⭐ 1 stars | Unknown

AI-powered Unity Editor automation for Claude Code - built on the official Unity CLI + com.unity.pipeline: 140+ editor commands, scene editing, tests, headless builds & C# eval

agentic-coding ai-agent automation claude claude-code claude-skill

Generated by Research Pulse on 2026-08-05 06:07