Paperraft · Edition 2026-10-20

Oct 2026arXiv 2610.26100

othertest

Multilingual Voicemail and Answering Machine Detection via Pretrained Audio Embeddings and Gradient Boosting Machines

The method replaces proprietary answering machine detection systems and computationally expensive real-time models (including ASR-plus-LLM pipelines) with frozen pretrained audio embeddings feeding a gradient boosting classifier. Cost is low: no GPU or fine-tuning is required for inference, training reduces to fitting a GBM on extracted embeddings, and the authors release a 500+ recording, 7-language dataset (OpenAMD) for replication. What can fail: the dataset is small for a production telephony system, per-language and per-gender performance differences were found statistically significant, and robustness degrades under additive noise and packet loss typical of real carrier networks.

observed signal

2impl.

0stars

—citations

—claimed gain

abstract onlyReview decision Download skill ↓Original paper ↗

Oct 2026arXiv 2610.25964

agent evaluationadopt

Token-Set Choice Confounds POPE: A Systematic Audit of Yes/No Extraction in VLM Hallucination Evaluation

It replaces hard-coded vocabulary-ID yes/no comparisons in POPE-style logit readouts with a tokenizer-derived token set covering the surface forms the model actually emits (e.g., ' Yes' vs 'Yes'). The cost is a one-time audit of generated tokens per model/tokenizer plus a slightly larger lookup set; no additional memory, latency, or compute in production. It can fail if adopted naively: tokenizer readouts must still be re-derived per model family, and cross-paper score comparisons remain invalid until readout and quantization precision are matched, since measured baseline shifts exceeded most published mitigation gains.

observed signal

4impl.

2stars

—citations

—claimed gain

abstract onlyReview decision Download skill ↓Original paper ↗

Oct 2026arXiv 2610.26394

agent evaluationtest

When Coding Agents Outpace Review: Evaluating PRD-Driven Development in Mature Enterprise Repositories

PACE replaces ad hoc or issue-resolution-style evaluation of coding agents with evidence-based scoring of full PRD fulfillment against 750 expert-audited Code Points in mature repositories. It costs evaluation engineering effort: tasks must be adapted or the benchmark adopted as-is, expert-style auditing is labor-intensive, and running 43 enterprise tasks against candidate model-scaffold configurations consumes API budget. It can fail as a selection signal because even the best configuration achieves only a 14.0% complete-delivery rate, so benchmark ranking may not translate into dependable autonomous delivery on the reader's own repositories.

observed signal

1impl.

1stars

—citations

—claimed gain

abstract onlyReview decision Download skill ↓Original paper ↗

Oct 2026arXiv 2610.26861

agent evaluationwatch

Can It Actually Work? Synthesizing Functional CAD Mechanisms

FunCAD-Harness replaces visual plausibility checks and generic simulation passes with candidate-specific physical tests that translate abstract functional requirements into executable verification of LLM-generated CAD mechanisms. Adoption costs a physics simulation stack (CAD modeling, kinematic/dynamic simulators), test-generation logic per candidate design, and iterative agent-simulation loops that increase latency and API token spend per design. It can fail where requirements are too underspecified to operationalize as tests, where simulators diverge from real manufacturing tolerances, and it transfers only partially to domains outside mechanical CAD.

observed signal

2impl.

7stars

—citations

—claimed gain

abstract onlyReview decision Download skill ↓Original paper ↗

Oct 2026arXiv 2610.26802

efficient trainingnot applicable

Scaling Parameter-Efficient Fine-Tuning for Brain Encoding

The method replaces fitting fixed readouts on frozen pretrained backbones (or costly full fine-tuning) with LoRA adapters plus a lightweight readout on large multimodal backbones, exploiting causal structure to predict multiple timepoints in one pass instead of sliding windows. Costs include access to large backbones (up to 104B parameters, requiring API access or multi-GPU inference beyond a single 24 GB card for the largest models), training complexity of joint multi-LoRA training, and the need for expensive fMRI/MEG-style neural recording data. It can fail if the reader lacks aligned neural-stimulus datasets, if the backbone's causal structure does not match the temporal sampling of the neural data, or if LoRA rank and readout capacity are insufficient for a given brain region or modality.

observed signal

1impl.

2stars

—citations

—claimed gain

abstract onlyReview decision Download skill ↓Original paper ↗

Oct 2026arXiv 2610.26825

memory and contextwatch

MemT-1.0: Towards Memory-Native Evolution

The method replaces full-context KV caching and external retrieval augmentation with memory internalized in model state via Gated DeltaNet-2 updates and a Mixture-of-Storage layer. It costs a full training pipeline (continued pre-training plus post-training on 34B tokens over a prepared backbone), a nonstandard architecture with custom training and inference infrastructure, and reported parity rather than superiority against retrieval-augmented methods. It can fail because the claimed gains are tied to the authors' released MemT checkpoints and infrastructure, long-horizon memory quality may degrade on workloads unlike their evaluations, and there is no drop-in path for existing API-based models.

observed signal

4impl.

304stars

—citations

1.86xclaimed gain

abstract onlyReview decision Download skill ↓Original paper ↗

Oct 2026arXiv 2610.26342

agent evaluationwatch

VibeTraining-Bench: Data-Trainer-Harness Co-Design for Iterative System Improvement

VibeTraining-Bench replaces ad hoc manual tuning of data, trainer, and inference harness with a standardized benchmark that measures whether research agents can perform that co-design iteratively under a fixed budget. It costs substantial compute per run, since each episode requires repeated training and evaluation cycles inside isolated containers with anti-reward-hacking safeguards. It can fail to transfer: strong results by frontier research agents on this benchmark do not imply a small team with one 24 GB GPU can replicate the iterative training loops, and residual reward-hacking or overfitting to validation feedback remains possible.

observed signal

2impl.

123stars

—citations

—claimed gain

abstract onlyReview decision Download skill ↓Original paper ↗

Oct 2026arXiv 2610.26836

memory and contextadopt

Same Simulation, Different Evidence: Where Retrieved Tool Output Costs LLM Agents Accuracy

The study evaluates replacing in-context tool outputs with file-based retrieval (section indexes) in agent harnesses, and proposes mitigations: explicitly naming the authoritative output file or inlining outputs that fit. Retrieval saves context window space, but for GPT-5.x models it costs 8-11 accuracy points; the naming prompt costs essentially nothing and recovers most of the loss. The residual failure mode is evidence substitution: the agent computes correctly but answers from its own higher-precision intermediate files, misrounding to a neighboring option, and prompting for correct rounding does not fix this.

observed signal

2impl.

132stars

—citations

—claimed gain

abstract onlyReview decision Download skill ↓Original paper ↗

Oct 2026arXiv 2610.26233

tool usetest

Heartbeat Thinking: Speculative Reasoning and Tool Execution While Listening in Realtime Voice Agents

It replaces chunk-bounded speculative execution, where generation is truncated or restarted at each speech update, with periodic Thinker invocations that let ongoing reasoning and dispatched tool calls run to completion against the latest partial transcript and a shared results record. The cost is additional inference calls during listening (raising token/API spend), plus orchestration complexity for managing concurrent invocations and a shared state record. It can fail when partial transcripts mislead speculative tool calls into wasted or incorrect executions, and the paper notes final task quality effects are benchmark-dependent, so gains may not transfer to a given workload.

observed signal

1impl.

10stars

—citations

1.57xclaimed gain

abstract onlyReview decision Download skill ↓Original paper ↗

Oct 2026arXiv 2610.26324

agent evaluationtest

The Control Layer Behind Rule-Following Agents Trading Across Market Conditions

The method replaces informal judgment of agent trustworthiness with reproducible trade-log tests that measure rule adherence (stops, targets, position limits, regime-conditional risk) before capital is committed. Cost is modest: the published trade snapshot, Bitcoin prices, and regime labels suffice to rerun the tests, but the methodology assumes agents that declare strategies and risk parameters upfront, and building equivalent instrumentation for a new agent is real engineering work. It can fail by producing false confidence: rule-following is shown to be orthogonal to profitability, sample sizes are small (2,627 positions, one unprofitable agent), and results are trade-level, not account-level returns.

observed signal

1impl.

10stars

—citations

—claimed gain

abstract onlyReview decision Download skill ↓Original paper ↗

Oct 2026arXiv 2610.26632

memory and contexttest

When the harness decides success: state that cannot be re-observed

The method replaces drop-oldest-turn context truncation in agent harnesses with summarization of prior turns plus an explicit invitation for the model to take notes. It costs additional tokens per turn for summary generation and added harness complexity, and provider-native reasoning replay may be needed to recover the remaining gap. It can fail when a required fact is never re-observable and the summary or notes omit it, and the ARC-AGI-3 numbers are single-run point estimates that may not transfer to other benchmarks or models.

observed signal

2impl.

122stars

—citations

—claimed gain

abstract onlyReview decision Download skill ↓Original paper ↗

Oct 2026arXiv 2610.26101

multi-agent orchestrationwatch

Reliable Before Profitable: QuantHarness for Governing Financial Strategy Development with LLM Agents

It replaces ad-hoc single-agent or loosely scripted LLM pipelines for turning natural-language research requests into backtested financial strategies, substituting a structured workflow (specification, data grounding, construction, testing, review) with enforced stage-level information, action, and commitment boundaries. The cost is additional orchestration complexity, structured run-record storage, and extra LLM calls across stages, which raises latency and API spend per strategy developed. What can fail: the benchmark (FinRelBench) is introduced by the same authors, so the 90.72% figure may not transfer to real trading desks; boundary rules can block legitimate repairs or add friction without catching novel leakage modes; and the paper demonstrates violation containment only under its own controlled stress tests, not under production data or live-market conditions.

observed signal

2impl.

132stars

—citations

—claimed gain

abstract onlyReview decision Download skill ↓Original paper ↗

Oct 2026arXiv 2610.26369

multi-agent orchestrationwatch

Autonomous intelligence for mining productivity: a multi-agent AI approach to operational value creation

EI2ROC proposes a three-tier multi-agent architecture (LLM orchestrator plus specialised monitoring and production-capacity agents over an MCP-style tool interface) intended to replace manual, siloed supervision in mining Integrated Remote Operation Centres, alongside a six-dimension benchmarking framework for industrial agentic systems. It costs orchestration complexity, LLM inference cost and latency per decision in a safety-critical setting, and the paper itself marks most components as reference design or future work, with only preliminary cost and latency observations from one site. It can fail through orchestrator errors propagating to SCADA-connected actions, unmet real-time and regulatory-auditability requirements, and the absence of validated benchmarks for the industrial blind spot the paper identifies.

observed signal

2impl.

132stars

—citations

—claimed gain

abstract onlyReview decision Download skill ↓Original paper ↗

Oct 2026arXiv 2610.26469

agent evaluationtest

VeriCUA: Learning Proactive Verification for Reliable GUI Agent Rewards

It replaces prompted foundation-model judges (4B to GPT-5.4 scale, plateaued at 68%-70% agreement) with a small trainable Judge, as small as 1B, that learns what to check and when to stop via online RL while a frozen Probe executes checks in the live environment. The cost is an online RL training loop requiring interactive OSWorld-style environments, single-attempt verification per terminal state, and return normalization across states, plus engineering to couple the Judge to a live GUI sandbox. It can fail when target environments differ from the training distribution despite the reported cross-benchmark transfer, when reward agreement gains do not translate to Actor task success on a specific workload, or when the team cannot maintain the live-environment training infrastructure.

observed signal

2impl.

132stars

—citations

—claimed gain

abstract onlyReview decision Download skill ↓Original paper ↗

Oct 2026arXiv 2610.26826

agent evaluationtest

RefCraft: Benchmarking Visual-Reference-Conditioned Agentic Generation through Web Development

RefCraft replaces idealized webpage-generation evaluations (single-screenshot reconstruction or text-only instructions) with 300 multimodal tasks combining instructions and one to five reference screenshots, scored via keypoints plus programmatic and VLM judges. It costs nothing in model deployment but requires API or GPU budget to run candidate models through render-inspect-refine loops, and the VLM-judge component adds evaluation variance and API cost. It can fail as a selection signal if the reader's production tasks differ from webpage development, if VLM judge scores are noisy or gameable, or if benchmark overfitting substitutes for real user evaluation.

observed signal

2impl.

132stars

—citations

—claimed gain

abstract onlyReview decision Download skill ↓Original paper ↗

Oct 2026arXiv 2610.26828

efficient trainingtest

Agent World Model in the Wild: Computer-Use Agent Environment Synthesis

WildAWM replaces reliance on human-collected computer-use trajectories and hand-built training environments with an open-source-model pipeline that synthesizes executable GUI/MCP/CLI environments, tasks, and execution-based verifiers. It costs a full environment-generation and agent-training workflow: heterogeneous rollouts, cross-judge consistency checks, counterfactual probes, and fine-tuning runs, which is feasible with open models but exceeds casual experimentation on a single 24 GB GPU. It can fail if synthetic tasks drift from real-world distributions, if verifiers reward incorrect trajectories that pass the quality gates, or if gains do not transfer to the reader's specific target applications.

observed signal

2impl.

132stars

—citations

—claimed gain

abstract onlyReview decision Download skill ↓Original paper ↗

Sep 2026arXiv 2609.17509

efficient architecturestest

LACE: Layer-Wise Compression for Dynamic Frame Rate Codecs

LACE replaces single shared segmentation boundaries in dynamic frame-rate neural audio codecs with an independent compression step at each quantization layer, plus union alignment and boundary anchors so layer token streams stay duration-consistent for TTS. The cost is added codec complexity: per-layer compression modules and alignment logic must be trained and maintained, and adoption requires integrating the ESPnet3 codec recipe and retraining or fine-tuning the codec and downstream TTS model. It can fail if layer-specific boundaries interact poorly with an existing TTS stack, if reconstruction gains do not transfer to your data domain beyond LibriTTS, or if the alignment mechanisms erode the frame-rate reduction under different speech characteristics.

observed signal

2impl.

674stars

0citations

—claimed gain

abstract onlyReview decision Download skill ↓Original paper ↗

Oct 2026arXiv 2610.26511

agent evaluationtest

RouterWorld: Scaling Conversational AI with Latency-Aware LLM Routing

RouterWorld replaces ad hoc, accuracy-only router evaluation with a standardized benchmark that treats latency as a first-class metric and models realistic partial observability (cold-start logs and streaming data). Adopting it costs integration effort against its LLM-as-a-judge protocol and does not itself provide a production router; the reader still needs to select or train one of the baseline routing methods. It can fail to transfer if the reader's query distribution, candidate model set, or hardware latency profile diverges from the benchmark's assumptions, and its LLM-as-a-judge quality labels may not match the reader's task-specific quality bar.

observed signal

1impl.

122stars

—citations

—claimed gain

abstract onlyReview decision Download skill ↓Original paper ↗

Oct 2026arXiv 2610.26531

speculative decodingtest

ReAnchor: Reusing Rejected Verifier States for Diffusion Drafting

ReAnchor replaces the fixed single-block diffusion draft by reanchoring a pretrained drafter at shifted positions using verifier hidden states that standard speculative decoding discards after the first rejection, generating parallel draft blocks that extend coverage beyond the original block endpoint. It is training-free, so the cost is limited to implementation complexity and modest extra compute for the additional reanchored draft blocks, with no retraining or extra model weights. Gains can shrink or vanish at sampling temperatures above zero, depend on a compatible block diffusion drafter such as DFlash, and may not transfer to workloads or target models outside the evaluated set.

observed signal

1impl.

122stars

—citations

1.13xclaimed gain

abstract onlyReview decision Download skill ↓Original paper ↗

Oct 2026arXiv 2610.26636

coding agentstest

Revision-Based Learning from Interactive Experience and Feedback

RELIEF replaces training pipelines that ignore post-generation interactive behavior by having a coding agent revise generated interfaces from feedback (code inspection, automated interaction, human comments, and interaction recordings) and folding those revisions back into the original generation traces as fine-tuning supervision. Its cost is a feedback-collection effort (the study used 200 websites, 355 interaction recordings, 24 users), agent revision runs, and fine-tuning compute, all feasible on a single 24 GB GPU only for small models or via API-based fine-tuning. It can fail when feedback is noisy or sparse, when the base model is too weak to benefit (effectiveness depends on base capability), and when improvements in revised traces do not transfer to zero-shot generation of new interfaces.

observed signal

1impl.

122stars

—citations

—claimed gain

abstract onlyReview decision Download skill ↓Original paper ↗

Oct 2026arXiv 2610.26714

agent evaluationtest

HoloCosm: Beyond Single-Metric Benchmark Compression Across Multiple Metrics

HoloCosm replaces full-benchmark evaluation and single-metric sample-selection methods by selecting a small subset of samples that jointly minimizes score-reconstruction error across multiple metrics while constraining ranking distortion. It costs an offline optimization pass over existing model scores on the full benchmark, meaning it requires score data from a population of source models and adds selection complexity before it saves any compute. It can fail when applied to models outside the source population used for selection, when metric correlations shift, or when too few historical model scores exist to calibrate the compressed subset, in which case random sampling may match or beat it.

observed signal

1impl.

122stars

—citations

—claimed gain

abstract onlyReview decision Download skill ↓Original paper ↗

Oct 2026arXiv 2610.25997

othernot applicable

Federated Community Detection in Bipartite Networks

FLICKER replaces centralized spectral decomposition of a horizontally partitioned matrix with federated power iteration that exchanges only matrix-vector products on local blocks, and BiFLICKER applies this to bipartite community detection. The cost is iterative cross-server communication rounds, a recursive grid search for spectral targets, and reliance on a reference server whose rescaled leading singular values must be valid global targets. It can fail when local sample sizes are too small for the reference server to estimate targets, when the data do not follow the assumed sparse bipartite stochastic block model, or when the horizontal partitioning assumption does not match the deployment.

observed signal

0impl.

0stars

—citations

—claimed gain

abstract onlyReview decision Download skill ↓Original paper ↗

Oct 2026arXiv 2610.26083

quantizationwatch

MX-Like Tensor Core for LLM Inference with NPU Integration via Hardware-Software Co-Design Exploration

The work does not replace a software technique; it replaces ad hoc selection of block-scaled quantization formats (OCP MX, NVFP4) with a systematic HW-SW co-design exploration of scale, weight, and activation format combinations on a specialized NPU MAC datapath. For a software practitioner the actionable finding is narrow: scale format choice matters mainly under aggressive FP4 quantization, where UE5M3 scales improve accuracy at negligible silicon cost. Adoption fails at the reader's constraints because the gains are realized only in custom NPU silicon (SNAX platform), cannot be enabled on a commodity 24 GB GPU or third-party API, and the perplexity results may not transfer across models or calibration pipelines.

observed signal

0impl.

0stars

—citations

1.34xclaimed gain

abstract onlyReview decision Download skill ↓Original paper ↗

Oct 2026arXiv 2610.26102

safety and guardrailstest

Provable Adversarial Detection: Prime Quantization meets Gromov--Wasserstein

The method replaces heuristic adversarial detectors (e.g., feature-density or statistics-based screens) with a detector that compares sample geometry across a CNN embedding space and a prime-quantized space, using Gromov–Wasserstein discrepancy with claimed separation guarantees. It adds per-inference cost: an extra embedding pass, a quantization mapping, and a GW-style distance computation, plus implementation complexity, without quantified latency or memory figures in the abstract. It can fail against adaptive attacks crafted to preserve geometry in both spaces, and its guarantees likely depend on the specific threat model, attack strength bounds, and embedding choices validated in the paper's experiments.

observed signal

0impl.

0stars

—citations

—claimed gain

abstract onlyReview decision Download skill ↓Original paper ↗

Oct 2026arXiv 2610.26131

serving and batchingwatch

A Hybrid GPU and Processing-Near-Memory Architecture with Shared DRAM for Efficient Prefill-Decode Disaggregation

DuoPD replaces GPU-side prefill-decode disaggregation and SM partitioning by executing full decoding on Processing-Near-Memory units in the HBM logic die while prefill runs on the GPU, sharing the HBM interface via TSV-level bandwidth allocation. The cost is entirely hardware-side: it requires custom PNM silicon, a new microarchitecture, and a hierarchical data-mapping scheme, with no software-only path to deployment. It can fail commercially if no memory vendor ships such parts, and technically if bandwidth sharing between GPU and PNM does not hold under real serving skew between prefill and decode load.

observed signal

0impl.

0stars

—citations

2.23xclaimed gain

abstract onlyReview decision Download skill ↓Original paper ↗

Oct 2026arXiv 2610.26160

otherwatch

Holistic Buoyancy-Driven Cloud Native Resource Management

The method replaces static resource requests and prior workload profiling in Kubernetes scheduling with runtime 'buoyancy' signals that measure each workload's sensitivity to shared resources such as last-level cache and memory bandwidth, driving adaptive scheduling, node-local allocation, and limited rescheduling. It costs a Kubernetes-native control plane extension, continuous runtime instrumentation of workloads, and a multi-node cluster to provide rescheduling headroom, plus ongoing maintenance of the sensitivity-measurement framework. It can fail when buoyancy estimates are inaccurate or unstable under rapidly shifting workload mixes, when rescheduling triggers migration overhead or disruption, and its benefits are unproven outside the evaluated congestion regimes.

observed signal

0impl.

0stars

—citations

—claimed gain

abstract onlyReview decision Download skill ↓Original paper ↗

Oct 2026arXiv 2610.26258

sparsity and pruningtest

AudioScorer: Learning Question-Aware Token Relevance for Large Audio Language Models

It replaces heuristic audio-token pruning (attention, similarity, temporal sampling) with a lightweight learned scorer that ranks tokens by relevance to the user question, distilled from the frozen LALM's answer-gradient saliency plus answer supervision. The cost is a one-time training/distillation stage that needs gradient access to a self-hosted frozen LALM, an added per-question scoring step in inference, and retained quality that depends on the chosen retention rate. It can fail on out-of-distribution questions or acoustic content where the scorer mis-ranks tokens, is unavailable for API-only models without gradients, and may degrade tasks that depend on dense temporal detail at aggressive retention.

observed signal

0impl.

0stars

—citations

—claimed gain

abstract onlyReview decision Download skill ↓Original paper ↗

Oct 2026arXiv 2610.26340

First page of A Generic On-Board AI Processing Chain for Earth Observation: From Application Development to Embedded Deployment and Mission-Level Orchestration

otherwatch

A Generic On-Board AI Processing Chain for Earth Observation: From Application Development to Embedded Deployment and Mission-Level Orchestration

The method replaces ground-station processing of full downlinked scenes with an on-board chain of int8-quantized inference on FPGA, a reusable tiler-inference-mosaicer workflow, and unikernel-packaged steps orchestrated by a cloud-native on-board orchestrator. The cost is substantial embedded-engineering effort: model quantization and hardware porting, plus adoption of specialized unikernel packaging and orchestration infrastructure, none of which runs on a conventional GPU workstation. Adoption can fail through quantization-induced accuracy loss on new tasks, hardware-specific toolchain friction, and orchestration overhead on constrained space-grade compute.

observed signal

0impl.

0stars

—citations

—claimed gain

abstract onlyReview decision Download skill ↓Original paper ↗

Oct 2026arXiv 2610.26370

efficient trainingtest

An Empirical Study of Parameter-Efficient Fine-Tuning for Mixture-of-Experts LLMs

The recipe replaces full fine-tuning of MoE models (and adapters that bypass experts) with LoRA/DoRA applied directly to native expert weights, trained without a load-balancing auxiliary loss since no nonzero coefficient Pareto-dominated the zero baseline. Cost is limited to a small trainable-parameter budget (17% fewer than the module-bypass baseline) within a single-GPU fine-tuning run, plus a measured 3.2-point average downstream regression for the redesigned variant. Failure modes include expert routing drift degrading general-knowledge retention, results transferring poorly across the three studied backbones to other MoE architectures, and task-specific sensitivity to which experts and ranks are adapted.

observed signal

0impl.

0stars

—citations

—claimed gain

abstract onlyReview decision Download skill ↓Original paper ↗

Oct 2026arXiv 2610.26379

efficient architecturestest

CTLite: Lightweight Irregular Multivariate Time Series Forecasting with Reusable Continuous-Time Summaries

CTLite replaces learned temporal encoders for irregular multivariate time series with a parameter-free continuous-time moment encoder whose summaries are cached and reused across optimization steps and repeated queries, leaving only a small damped-extrapolation plus low-rank cross-variable readout to train. It costs almost nothing in parameters or latency and runs on a single small GPU or even CPU-class hardware, but it adds a preprocessing and caching layer and its accuracy claim is only best-or-second-best rather than uniformly superior. It can fail where the moment-based local level and trend summaries are a poor fit for the dynamics, on datasets unlike the four evaluated benchmarks, or when histories change frequently enough to negate the reuse benefit.

observed signal

0impl.

0stars

—citations

1.23xclaimed gain

abstract onlyReview decision Download skill ↓Original paper ↗

Oct 2026arXiv 2610.26437

otherwatch

R2-GP: Variance-Calibrated Priors for Additive Gaussian Processes

R2-GP replaces independent priors on component amplitudes in Matérn-kernel additive Gaussian process models with a single global R2 prior that allocates each component's variance contribution after rescaling by length-scale and observed inputs. It costs a reparameterisation of the prior and, for large datasets, adoption of Hilbert space kernel approximations, adding modelling complexity without reducing compute. It can fail if the chosen global R2 prior is misspecified, if the spectral approximation degrades kernel fidelity at small length-scales, or in regimes where the independent-prior baseline was already adequate.

observed signal

0impl.

0stars

—citations

—claimed gain

abstract onlyReview decision Download skill ↓Original paper ↗

Oct 2026arXiv 2610.26438

safety and guardrailswatch

Fake News, Real Crashes: Information Injection Attacks on LLM Financial Trading Agents

The paper does not replace an existing method; it is an attack benchmark that supersedes informal red-teaming of news-consuming LLM agents, and it evaluates candidate defenses including hierarchical aggregation of evidence. Adopting its lessons costs added inference calls and latency (hierarchical aggregation and credibility-cue checking require extra model passes) plus engineering effort to validate and annotate untrusted feed items. Defenses can fail because attack success is state-dependent (91% ASR against an agent about to hold, 12% against one about to buy), generalizes across model vendors, and remains effective when fake items are blended unlabeled into real feeds, while surface credibility cues alone drive success regardless of claim content.

observed signal

0impl.

0stars

—citations

—claimed gain

abstract onlyReview decision Download skill ↓Original paper ↗

Oct 2026arXiv 2610.26470

sparsity and pruningtest

On Learning Expand-and-Sparsify Representations

The method replaces a dense ReLU layer (and its random-map expand-and-sparsify variant) with a learned linear expansion followed by top-k sparsification, letting the expansion be much narrower than the exponential widths random theory requires while matching the dense layer's accuracy at 10% activation density. The cost is an extra learned linear map and top-k selection per layer, plus the need to tune expansion width and sparsity per task; the evidence base is 73 OpenML classification and 48 regression tasks, all small tabular benchmarks, not production-scale deep networks. It can fail because the learned-map advantage vanishes at wide expansions, gains over dense ReLU are parity rather than improvement, and results on small tabular data may not transfer to the reader's actual workloads.

observed signal

0impl.

0stars

—citations

—claimed gain

abstract onlyReview decision Download skill ↓Original paper ↗

Oct 2026arXiv 2610.26474

safety and guardrailstest

AfriGuard: Safety and Robust Refusal For African Languages

AfriGuard replaces general-purpose or larger off-the-shelf guardrail models for safety classification and refusal detection in African languages with a culturally grounded evaluation dataset plus supervised fine-tuned smaller classifiers. Costs include fine-tuning on the released dataset, maintaining per-deployment classifiers in the inference path (added latency and a failure point), and dependence on dataset coverage of the 10 included languages and safety categories. It can fail on languages or dialects outside the dataset, on culturally specific harm expressions not represented in training data, and because the paper shows models often produce linguistically inadequate responses rather than proper refusals, a detector may report incidental non-assistance as safe behavior.

observed signal

0impl.

0stars

—citations

—claimed gain

abstract onlyReview decision Download skill ↓Original paper ↗

Oct 2026arXiv 2610.26676

multi-agent orchestrationtest

Which Model for Which Role? Cost-Efficient Pareto Recommendation for LLM Agentic Workflows

CC-Gittins replaces exhaustive or ad-hoc evaluation of model-to-role assignments in multi-model agentic workflows with a cost-aware multi-objective bandit search that reuses each evaluation for both quality-seeking and cost-seeking. It costs the implementation of a Gittins-index search loop plus uncertainty-adjusted quality and cost tracking per candidate, with a small inference overhead per batch but no additional model training. It can fail when the search budget is too small to clear the frontier, when uncertainty estimates are miscalibrated for the user's task distribution, or when the workflow has so few roles and candidate models that exhaustive evaluation is already cheap.

observed signal

0impl.

0stars

—citations

10xclaimed gain

abstract onlyReview decision Download skill ↓Original paper ↗

Oct 2026arXiv 2610.26693

speculative decodingwatch

Parallel Language Models via Variational Decoding

The method replaces single-token autoregressive decoding with parallel generation of mutually consistent tokens in one forward pass, trained via a variational objective combining autoregressive and parallel decoding losses. It requires pretraining the model with this objective from scratch, so it cannot be applied to existing pretrained checkpoints, and the reported speedups depend on latent-variable routing that adds training and inference complexity. Failure modes include quality degradation on tasks where token dependencies are poorly captured by the latent variables, and the power-law speedup claim is validated only up to 3B parameters, leaving behavior at production model scales unverified.

observed signal

0impl.

0stars

—citations

5.7xclaimed gain

abstract onlyReview decision Download skill ↓Original paper ↗

Oct 2026arXiv 2610.26747

efficient architectureswatch

recAMPLIFY: Large-Scale Pretraining of Recursive Protein Language Models

recAMPLIFY replaces parameter scaling in masked protein language models with a weight-tied recursive Transformer core that refines hidden states over a variable number of cycles, optionally with a Gated Mixture Recurrence decoder for adaptive depth. The recipe requires large-scale pretraining (2T tokens), which exceeds the reader's infrastructure and budget; only released checkpoints, if published, would make GMR's adaptive-depth inference usable, and GMR's accuracy trade-off is task-dependent and can be negative. Recursive weight-tying can become unstable beyond the trained depth (the paper's own gating analysis addresses this), and the method is specific to masked protein encoders, not general LLM workloads.

observed signal

0impl.

0stars

—citations

—claimed gain

abstract onlyReview decision Download skill ↓Original paper ↗

Oct 2026arXiv 2610.26754

sparsity and pruningtest

TokenGate: Adapting LLMs for Fast and Memory-Efficient Prefill with Learned Token Pruning

TokenGate replaces full-sequence prefill, in which every input token passes through all layers, with a learned gating component that scores and prunes tokens via token-gated attention added to an existing pretrained model. The cost is a lightweight adaptation step comparable to parameter-efficient fine-tuning, plus the integration and maintenance of a nonstandard attention mechanism in the serving stack. It can fail on workloads where pruned tokens carry task-critical information, where the pruning distribution shifts from the adaptation data, or where short-to-medium contexts make the TTFT gain too small to justify the added complexity.

observed signal

0impl.

122stars

—citations

11.3xclaimed gain

abstract onlyReview decision Download skill ↓Original paper ↗

Oct 2026arXiv 2610.26817

othertest

Next Best Action: Causal Constrained Offline RL for Business Action Targeting

This replaces rule-based or myopic (single-step uplift/response-model) next-best-action targeting with a Constrained MDP solved via linear programming, integrating causal off-policy correction of customer dynamics and yielding shadow prices for constraint trade-offs. The cost is not GPU compute but data and modeling effort: it requires logged customer-intervention data with adequate coverage, a defensible state representation (the paper uses RFM stages), causal estimation of transition dynamics, and LP machinery rather than standard supervised pipelines. Failure modes include misspecified causal dynamics that invalidate the off-policy correction, OPE that relies on a learned model rather than a randomized holdout (so the 19% figure may not transfer to live traffic), state aggregation that hides heterogeneity, and constraint shadow prices that mislead if the LP formulation omits real operational couplings.

observed signal

0impl.

0stars

—citations

1.19xclaimed gain

abstract onlyReview decision Download skill ↓Original paper ↗

Oct 2026arXiv 2610.26833

othernot applicable

Supervised Learning with Quantum Hamiltonian Embeddings

The method replaces a classical supervised classifier with a parameterized data-dependent Hamiltonian whose approximate ground state defines the decision rule, trained and executed on quantum processors. It costs access to quantum hardware (demonstrated on IBM Heron processors), a quantum ground-state approximation subroutine, and a hybrid training pipeline, none of which fit a 24 GB GPU or standard API budget. It can fail through noise and limited qubit counts on current hardware, unproven practical advantage over classical baselines on real workloads, and dependence on cryptographic assumptions for its classical-hardness claim.

observed signal

0impl.

0stars

—citations

—claimed gain

abstract onlyReview decision Download skill ↓Original paper ↗

Oct 2026arXiv 2610.26847

tool usetest

WeAct: Dynamic Weight Switching as Agent Actions

WeAct replaces joint fine-tuning and static adapter merging by training independent capability-specific LoRA adapters on a frozen backbone and letting the agent select adapters at runtime through its native tool-calling interface. The cost is maintaining and serving multiple adapters (each small, but adding routing complexity), plus training a selection policy whose adapter-switching actions must be learned from trajectories. It can fail if the model selects the wrong adapter mid-trajectory, if switching breaks contextual coherence across stages, or if the benchmark gains do not transfer to workloads with fewer or less distinct capability boundaries.

observed signal

0impl.

0stars

—citations

1.062xclaimed gain

abstract onlyReview decision Download skill ↓Original paper ↗

Research index

Research signals

Method under scrutiny

Exclusions and controls