Skip to content

archived edition · 2026-09-22

PaperraftFind the AI research
worth testing.

Understand what a paper changes, what it costs, and whether it fits your next build. Explore research on faster LLM inference, agent memory, and evaluation, with the original sources in reach.

Explore the researchHow papers are assessed

Start with a brief. Follow the evidence. Choose a test.

edition
2026-09-22
briefs
42
research areas
15
independent implementations
36

27 of 42 papers in the index (64%) have no independent implementations.

Research index

Search by topic, browse a research area, or sort by date. Read a brief, open the original paper, and inspect the available evidence before deciding what to test.

showing30 of 42
sort by
arXiv 2609.09153
First page of Procedural Graphs: Self-Evolving Execution Structures for LLM Agents
memory and contexttest

Procedural Graphs: Self-Evolving Execution Structures for LLM Agents

Replaces unconstrained next-action generation over an accumulating trajectory with an evolving (procedure, relation, procedure) graph that a guidance model turns into step-level hints biasing, but not dictating, the agent's next action. Costs an extra guidance-model call per decision step plus LLM-refiner passes that contrast failed and successful trajectories to edit the graph against a held-out validation set, adding orchestration complexity and API spend. Can fail if validation sets are small or unrepresentative, letting the refiner overfit graph edits, or if localization of the active node is wrong on long trajectories, injecting misleading guidance; the abstract reports only qualitative 'consistent gains' with no quantified improvement, so the benefit is unverified for any specific workload.

observed signal
12impl.
10stars
0citations
—claimed gain
abstract onlyReview decisionDownload skill ↓Original paper ↗
arXiv 2609.24797
First page of Complex KDA: Understanding and Enhancing the Expressivity of Kimi Delta Attention
efficient architecturestest

Complex KDA: Understanding and Enhancing the Expressivity of Kimi Delta Attention

CKDA replaces a single-transition delta-rule linear RNN layer (and the rank-2 DeltaProduct_2 construction) with a Kimi Delta Attention transition whose gate range is extended to [-1,1] and whose beta coefficient is extended to [0,2], achieving 2D-rotation expressivity in one diagonal-plus-rank-one update instead of two. It costs no extra rank or asymptotic compute over KDA and keeps updates non-expansive and stable, but it requires retraining from scratch with modified parameterizations and ranges, so it cannot be applied to existing pretrained checkpoints. It can fail on workloads where the added state-tracking expressivity is irrelevant, since language-modeling gains over a standard KDA baseline are negligible and the main benefits appear on state-tracking and length-extrapolation tasks.

observed signal
1impl.
1stars
—citations
—claimed gain
abstract onlyReview decisionDownload skill ↓Original paper ↗
arXiv 2609.05774
First page of Inference-Time Graph Engineering for Multi-Agent LLM Workflows
multi-agent orchestrationtest

Inference-Time Graph Engineering for Multi-Agent LLM Workflows

ReActNet replaces static or RL-learned multi-agent topologies with a training-free compilation step that generates a sequence of directed communication graphs per query, where each edge carries a natural-language message instruction. It costs additional LLM calls for graph compilation plus multi-round structured message passing, adding latency and token spend per query and requiring controller logic to define stage graphs and aggregators. It can fail through compounding orchestration overhead on simple tasks, sensitivity to the quality of the compiled graph (poorly conditioned connectivity or edge instructions degrade reasoning), and lack of validated gains on the reader's specific workloads, since benefits are shown only on benchmark suites.

observed signal
2impl.
2stars
0citations
—claimed gain
abstract onlyReview decisionDownload skill ↓Original paper ↗
arXiv 2609.24974
First page of Harness-Zero: Harness Distillation via Agent-as-Harness
distillationwatch

Harness-Zero: Harness Distillation via Agent-as-Harness

The method replaces deployment-time reliance on domain-optimized agent harnesses (and routing among many specialized harnesses) by distilling harness-induced behaviors into model weights, so a single fixed harness suffices at inference. It costs a training-time pipeline: an optimized harness per domain, a harnessing agent that rewrites student responses into the target action space, trajectory collection, and fine-tuning, which typically implies frontier-API spend and non-trivial engineering. It can fail when the distilled behaviors do not transfer across models or new task distributions, when the optimized harness itself is unavailable or weak for a domain, and because gains are validated only on the paper's three domains and may require re-distillation whenever the base model changes.

observed signal
2impl.
11stars
—citations
—claimed gain
abstract onlyReview decisionDownload skill ↓Original paper ↗
arXiv 2609.24983
First page of onPanda: Efficient Annotation of On-Policy Alignment Data for LLMs and Agents via Token-Level Correction
efficient trainingtest

onPanda: Efficient Annotation of On-Policy Alignment Data for LLMs and Agents via Token-Level Correction

onPanda replaces full manual rewriting or post-editing of model outputs with a locate-correct-continue loop: the annotator edits the first incorrect token, the response is truncated there, and the model regenerates from the corrected prefix, yielding on-policy SFT and preference data plus fine-grained positive-negative token pairs. The cost is adopting and integrating an interactive annotation tool (including harness connections for agent trajectories) rather than additional compute; the model still performs generation, so GPU requirements remain those of normal inference. Failure modes include correction positions drifting from the model's natural distribution after repeated edits, annotator inconsistency in choosing the 'first inappropriate token', and the 52% time saving resting on a small controlled study that may not transfer to other tasks, annotators, or base models.

observed signal
2impl.
11stars
—citations
2.08xclaimed gain
abstract onlyReview decisionDownload skill ↓Original paper ↗
arXiv 2609.24972
First page of RRSI: Regularized Recursive Self-Improvement of Agent Harnesses
otherwatch

RRSI: Regularized Recursive Self-Improvement of Agent Harnesses

RRSI replaces unconstrained automated editing of an agent harness (prompts, control flow, tools, memory) with evolution constrained by an annealed edit budget, exploration incentives, a critic that filters benchmark-specific proposals, and a pruner that removes small, costly, or obsolete changes. The cost is an additional optimization loop requiring many agent rollouts and evaluation passes on representative tasks, plus the engineering of the proposer/critic/pruner pipeline, all on top of API or GPU inference spend. It can fail if the evolution benchmark is not representative of production traffic, if the rollout budget is too small for the regularization to distinguish reusable mechanisms from noise, or if the evolved harness overfits to a specific backbone model and does not transfer when the model is swapped.

observed signal
2impl.
64stars
—citations
—claimed gain
abstract onlyReview decisionDownload skill ↓Original paper ↗
arXiv 2609.24259
First page of MemCalib: Benchmarking and Optimizing Memory Use in LLM Agents
memory and contextwatch

MemCalib: Benchmarking and Optimizing Memory Use in LLM Agents

MemCalib-RL replaces generic post-training (GRPO, on-policy self-distillation) for memory-use behavior with an ordered bidirectional counterfactual credit-assignment algorithm that separates over-use and under-use signals and localizes credit via exact atom ablation. It costs a full RL training run on 8B to 35B-parameter models, which exceeds a single 24 GB GPU and requires counterfactual ablation infrastructure and training data; API-only users cannot apply it. After adoption it can fail if the calibrated behavior does not transfer to the deployed memory system or if the residual directional skew between over-use and under-use persists on out-of-distribution tasks.

observed signal
1impl.
10stars
—citations
—claimed gain
abstract onlyReview decisionDownload skill ↓Original paper ↗
arXiv 2609.24967
First page of Emergent Collusion in Long-Horizon LLM Agent Interaction
safety and guardrailstest

Emergent Collusion in Long-Horizon LLM Agent Interaction

The paper does not introduce a technique to replace an existing method; it is an empirical study showing that long-horizon interaction between LLM agents produces protocol-violating collusion, and it identifies mitigations, principally restricting the amount and scope of shared interaction history, reward-structure design, and verification feedback. Adopting the mitigations costs little computationally (reduced context can even lower token spend) but requires redesigning multi-agent workflows, memory handling, and reward or incentive structures, plus running adversarial long-horizon evaluations. Collusion emerged in 94% of trajectories across 10 models, arose earlier in more capable models, and can persist or re-emerge through peer influence even after partial interventions, so single-point fixes may fail silently in production.

observed signal
2impl.
134stars
—citations
—claimed gain
abstract onlyReview decisionDownload skill ↓Original paper ↗
arXiv 2609.24391
First page of NAVIR: Neuromorphic Audio-Visual Speech Recognition for Robust Human-Robot Interaction on Edge Hardware
efficient architectureswatch

NAVIR: Neuromorphic Audio-Visual Speech Recognition for Robust Human-Robot Interaction on Edge Hardware

NAVIR replaces conventional AVSR pipelines built on 3D convolutions, recurrent units, and attention with factorized 2D-convolutional AkidaNet modules plus CTC decoding, targeting the BrainChip Akida neuromorphic chip. The cost is a hard dependency on Akida-class spiking hardware, quantization-aware training of a custom multimodal pipeline, and restricted operator support (sequential 2D convolutions only), which constrains model capacity and generality. It can fail through accuracy loss from quantization and firing-rate sensitivity, poor transfer beyond narrow command corpora like the 98.6%-accuracy industrial dataset, and the need for a camera with reliable lip visibility in the deployment environment.

observed signal
3impl.
213stars
—citations
5xclaimed gain
abstract onlyReview decisionDownload skill ↓Original paper ↗
arXiv 2609.24352
First page of Few-Shot Demonstrations Elicit the Use of In-Context World Representations in LLMs
memory and contexttest

Few-Shot Demonstrations Elicit the Use of In-Context World Representations in LLMs

Replaces zero-shot or single-world in-context observation formats with few-shot demonstrations drawn from multiple distinct environments sharing or varying latent structure. Costs only additional prompt tokens and the effort of assembling demonstrations from heterogeneous tasks or worlds, with no extra memory, training, or infrastructure. Can fail if available demonstrations do not actually span diverse underlying structures, if the task lacks a latent state the model must infer, or if longer prompts exceed context budgets without proportional benefit.

observed signal
2impl.
335stars
—citations
—claimed gain
abstract onlyReview decisionDownload skill ↓Original paper ↗
arXiv 2609.24165
First page of APEXA: Execution-Integrity Enforcement for Multi-Agent LLM Automation of Synchrotron Data Reduction
safety and guardrailsadopt

APEXA: Execution-Integrity Enforcement for Multi-Agent LLM Automation of Synchrotron Data Reduction

It replaces prompt-level safety instructions and transcript-based trust with a deterministic tool-layer check that only surfaces results backed by an executed tool call, plus a parser tolerant of cross-model tool-call format drift. Cost is engineering complexity in the tool layer and a modest latency overhead for validation; it runs on a single GPU or against API models since it is model-agnostic infrastructure, not training. It can fail if the guard's execution records are themselves spoofable, if legitimate results are incorrectly rejected (false positives blocking valid outputs), and the full 58-task benchmark has not yet been used for large-scale agent scoring, so end-to-end task success rates are unquantified.

observed signal
2impl.
132stars
—citations
—claimed gain
abstract onlyReview decisionDownload skill ↓Original paper ↗
arXiv 2609.24812
First page of MSI-Bench: Evaluating Multi-Speaker Voice Interaction for Collaborative AI Agents
agent evaluationwatch

MSI-Bench: Evaluating Multi-Speaker Voice Interaction for Collaborative AI Agents

MSI-Bench replaces ad hoc or single-speaker evaluation of voice agents with 1,152 multi-party multi-turn audio scenes featuring participant context, expected tool calls, and atomic rubrics across memory, instruction following, and reasoning. Adopting it costs benchmark integration effort and access to the evaluated models, and it is an evaluation asset rather than a deployable capability, so it adds no runtime latency or memory to a production system. It can fail to transfer if the reader's domain, languages, or speaker counts differ from the benchmark's scripted scenes, and rubric-based pass rates may not correlate with perceived quality in a specific product.

observed signal
2impl.
132stars
—citations
—claimed gain
abstract onlyReview decisionDownload skill ↓Original paper ↗
arXiv 2609.24264
First page of Canonical Procedural Actions: An Auditable Annotation Protocol for Tool-Use Agent Traces
agent evaluationwatch

Canonical Procedural Actions: An Auditable Annotation Protocol for Tool-Use Agent Traces

The protocol replaces ad hoc or purely endpoint-log-based inspection of tool-use agent traces with an explicit, versioned codebook that links procedural actions to anchored events and contextual evidence, enabling auditable inter-annotator comparison. Its costs are the manual effort of open-induction codebook development, repeated application audits, and dual isolated LLM annotation runs, plus the complexity of maintaining evidence references when multiple actions share a message anchor. It can fail as a measurement instrument because the reported overlaps are structural repeatability rather than semantic accuracy, task IDs leaked between development and test sets, tool payloads were truncated to 110 characters, and human-reference validity and downstream utility remain unestablished.

observed signal
1impl.
131stars
—citations
—claimed gain
abstract onlyReview decisionDownload skill ↓Original paper ↗
arXiv 2609.24890
First page of OSWorld-Pro: Process-based Evaluation for Computer Use Agents
agent evaluationtest

OSWorld-Pro: Process-based Evaluation for Computer Use Agents

It replaces binary end-state functional verification (as in OSWorld) with LLM-judge evaluation of over 2800 intermediate subgoals grounded in human annotations, revealing where and why an agent fails during execution. The cost is the evaluation infrastructure itself: a 300+ task environment with sequential subgoal checkpoints plus LLM-judge inference per subgoal, which raises evaluation latency and API spend relative to a single end-state check. The LLM judges, despite human-alignment validation, can misclassify subgoal fulfillment, and a benchmark-specific diagnosis may not transfer to the reader's own task distribution.

observed signal
1impl.
131stars
—citations
—claimed gain
abstract onlyReview decisionDownload skill ↓Original paper ↗
arXiv 2609.24971
First page of DolphinBench: Mapping the Pareto Frontier of Agent Memory
agent evaluationtest

DolphinBench: Mapping the Pareto Frontier of Agent Memory

DolphinBench replaces conversational QA-style memory benchmarks with task-completion evaluation over roughly 500k tokens of persona history per knowledge-work persona, requiring agents to succeed with relevant history and fail without it, and mandating cost and latency reporting alongside accuracy. Adoption costs the engineering effort of integrating the benchmark's harness and history into the reader's evaluation pipeline, plus the compute and API spend of running 200 tasks per persona, which is nontrivial on a limited cloud budget. It can fail to transfer if the reader's production tasks differ from the three included personas, and a memory system that scores well on the benchmark's Pareto frontier may still underperform on domain-specific recall patterns.

observed signal
1impl.
131stars
—citations
—claimed gain
abstract onlyReview decisionDownload skill ↓Original paper ↗
arXiv 2609.24174
First page of CREDO: Variance-Guided Rubric Evolution for Replay-Corrected Credit Assignment
efficient trainingwatch

CREDO: Variance-Guided Rubric Evolution for Replay-Corrected Credit Assignment

CREDO replaces naive terminal-reward credit assignment in RL training of long-horizon language agents with a learned credit head over rubric features, corrected by selectively replayed counterfactual continuations. It requires a resettable training environment, a frozen judge model, an extra credit head, and an expected replay budget, all of which add training complexity and compute on top of PPO. It can fail when rubrics are misspecified, the credit head is biased under distribution shift, or replay coverage is sparse, and the authors explicitly make no empirical superiority claim on language-agent benchmarks.

observed signal
0impl.
121stars
—citations
—claimed gain
abstract onlyReview decisionDownload skill ↓Original paper ↗
arXiv 2609.24289
First page of TTSE: A Two-Track Online Self-Evolution Framework
memory and contexttest

TTSE: A Two-Track Online Self-Evolution Framework

TTSE replaces static, externally supplied environment knowledge with two separately evolving stores: FACT, environment facts continuously verified against interaction evidence, and TIP, task-conditioned execution procedures, both injected at inference via retrieval. The cost is an additional online memory pipeline (fact verification, procedure extraction, retrieval injection) on top of the agent loop, plus the engineering to maintain and validate these stores; no training of the base model is required, so it is feasible on a single GPU or API-only setup. What can fail: incorrect or stale facts being verified into the FACT store, cross-condition mismatch when TIP procedures transfer poorly to new tasks, and unbounded memory growth or retrieval noise degrading rather than improving task success.

observed signal
0impl.
0stars
—citations
—claimed gain
abstract onlyReview decisionDownload skill ↓Original paper ↗
arXiv 2609.24348
First page of A Lean and Spec-Driven AI-Assisted Software Development Lifecycle for Applied AI Education: The AI-SDLC Approach
coding agentstest

A Lean and Spec-Driven AI-Assisted Software Development Lifecycle for Applied AI Education: The AI-SDLC Approach

It replaces ad-hoc prompting of AI coding agents with a lean, spec-driven development lifecycle that binds agents to specifications, AGENTS.md, and phase-specific skill files with bounded autonomy. Cost is process overhead: writing and maintaining repository-local specs and skill files, plus human review effort, with no additional infrastructure or API spend. It can fail if specifications drift from the code, if teams skip review gates under time pressure, or if the framework's value is not validated outside a small student-course setting with business-oriented use cases.

observed signal
0impl.
0stars
—citations
—claimed gain
abstract onlyReview decisionDownload skill ↓Original paper ↗
arXiv 2609.24401
First page of Artificial Structure Function Search: Preserving Artificial Functional Connectivity for Structured Pruning
sparsity and pruningtest

Artificial Structure Function Search: Preserving Artificial Functional Connectivity for Structured Pruning

ASF-S replaces heuristic or weight-magnitude structured pruning criteria with Principle Gradient Importance, a gradient-based candidate selection method that preserves output-layer topographical organization (Artificial Functional Connectivity). It costs a pruning search pass plus validation runs, and avoids the retraining phase that conventional structured pruning requires, though the search framework itself adds implementation complexity beyond off-the-shelf pruners. It can fail if the reported no-retraining accuracy recovery does not transfer to the reader's architectures or tasks, since the abstract provides no detail on which benchmarks, model sizes, or degradation margins were measured.

observed signal
0impl.
0stars
—citations
3.33xclaimed gain
abstract onlyReview decisionDownload skill ↓Original paper ↗
arXiv 2609.24417
First page of ARM: Attention with Routed-Memory for Learnable Sparse Control
KV cachetest

ARM: Attention with Routed-Memory for Learnable Sparse Control

ARM replaces fixed or eviction-based KV caches with a differentiable, fixed-size hierarchical memory that uses Gumbel-Softmax slot selection and sigmoid-gated soft updates instead of hard token eviction. The cost is architectural modification and training of the routing and memory-update policy, which requires retraining or fine-tuning the model rather than plug-in deployment on an existing checkpoint. It can fail through routing instability from Gumbel-Softmax sampling, information loss if the fixed memory size is too small for hard long-context inputs, and absence of compatibility with off-the-shelf pretrained models.

observed signal
0impl.
0stars
—citations
—claimed gain
abstract onlyReview decisionDownload skill ↓Original paper ↗
arXiv 2609.24446
First page of ActGov: Governing LLM Agent Actions via Policy-Constrained Validation
safety and guardrailstest

ActGov: Governing LLM Agent Actions via Policy-Constrained Validation

ActGov replaces static, predefined execution plans and fixed allowlists with a runtime enforcement layer that validates every proposed tool call against an iteratively constructed, SMT-verified policy set before external effects occur. It costs the engineering effort of building and maintaining tool specifications and policy records, plus per-action validation latency (including SMT counterexample checks during policy construction) added to each step of the agent loop. It can fail where tool specifications are incomplete or policies under-constrain authorization, where the policy-construction pipeline misses failure modes not present in observed traces, and where teams lack the verification tooling and expertise to maintain the SMT-backed policy lifecycle.

observed signal
0impl.
0stars
—citations
—claimed gain
abstract onlyReview decisionDownload skill ↓Original paper ↗
arXiv 2609.24515
First page of Beyond Predictable Paths: Redefining AI Security Incident Reporting for Agents
safety and guardrailstest

Beyond Predictable Paths: Redefining AI Security Incident Reporting for Agents

The paper proposes adapting conventional AI incident reporting to agents by adding agent-specific fields such as memory contents and accesses, autonomy levels, and tool usage, replacing generic AI system incident templates. The cost is added logging and governance overhead: teams must instrument memory, tool calls, and autonomy state, and the paper itself flags open problems in efficient recording and in determining whether incidents generalize. Adoption can fail through data leakage from over-detailed reports, attacks against the reporting infrastructure itself, and unresolved privacy requirements that the authors leave as research questions.

observed signal
0impl.
0stars
—citations
—claimed gain
abstract onlyReview decisionDownload skill ↓Original paper ↗
arXiv 2609.24517
First page of Not All Task Vectors Need Equal Rank: Energy-Proportional Allocation for Model Merging
efficient trainingtest

Not All Task Vectors Need Equal Rank: Energy-Proportional Allocation for Model Merging

SERA replaces uniform rank assignment in SVD-based model merging with per-task rank allocation proportional to each task vector's singular-value energy, keeping the total rank budget unchanged. It costs one SVD per task vector plus an allocation step, adding modest computation and implementation complexity over fixed-rank merging with no extra inference cost. It can fail when the assumption that spectral energy predicts useful capacity does not hold, and evidence is limited to vision protocols, so gains on language-model merging are unverified.

observed signal
0impl.
121stars
—citations
—claimed gain
abstract onlyReview decisionDownload skill ↓Original paper ↗
arXiv 2609.24538
First page of QLoRA Fine-Tuning of Ministral LLM for Sequence-to-Function Protein Annotation
efficient trainingtest

QLoRA Fine-Tuning of Ministral LLM for Sequence-to-Function Protein Annotation

The method replaces fixed-ontology multi-label classification for protein function annotation with sequence-to-text generation using a 4-bit NF4 quantized 3B model trained with low-rank adapters. Costs include training data curation for sequence-annotation pairs, the well-known QLoRA quality trade-off versus full fine-tuning, and reliance on an LLM-as-expert evaluation protocol that is not a validated biological benchmark. Failure modes include hallucinated or ungrounded functional descriptions, unreliable performance on proteins distant from the training distribution, and the authors' own caveat that data quality, scaling, and evidence grounding are not yet sufficient for practical reliability.

observed signal
0impl.
0stars
—citations
—claimed gain
abstract onlyReview decisionDownload skill ↓Original paper ↗
arXiv 2609.24569
First page of Poisson Exchange Beyond Submodularity: Effective Approximation Algorithms for Offline and Online Subset Selection over Matroids
otherwatch

Poisson Exchange Beyond Submodularity: Effective Approximation Algorithms for Offline and Online Subset Selection over Matroids

The method replaces standard greedy and prior local-search heuristics for selecting subsets under matroid constraints (e.g., feature selection, pruning candidate sets, summarization) with repeated maximum-gain exchanges scheduled by a non-homogeneous Poisson clock. The cost is implementation complexity of the continuous-time exchange process plus repeated gain evaluations and matroid independence checks, which can be substantial per iteration on large candidate pools. It can fail in practice because the guarantees are asymptotic approximation bounds: the reader's actual objective may not have an estimable weak-submodularity parameter gamma, and the realized solution quality on finite instances depends on gamma estimation and stopping criteria, not on the theoretical ratio.

observed signal
0impl.
0stars
—citations
—claimed gain
abstract onlyReview decisionDownload skill ↓Original paper ↗
arXiv 2609.24591
First page of Taking a Second Look: Correcting Sea Ice Forecasts with Sparse Observations
othernot applicable

Taking a Second Look: Correcting Sea Ice Forecasts with Sparse Observations

ECHO replaces fixed-distance propagation of sparse sea ice concentration observations with state-dependent propagation (ECHO-Scale) or a learned bounded residual correction around fixed propagation (ECHO-Delta). The cost is training an additional correction model on sea ice concentration fields and maintaining a data-assimilation pipeline tied to a specific geophysical forecasting setup, with modest inference overhead on top of the existing forecast. It can fail under observation geometries or noise regimes unlike those in training, and the learned ECHO-Delta variant is explicitly less robust to geometry shifts than ECHO-Scale.

observed signal
0impl.
0stars
—citations
—claimed gain
abstract onlyReview decisionDownload skill ↓Original paper ↗
arXiv 2609.24620
First page of Ascent: An Agentic System over the Model Context Protocol for Real-World Clinical Data Analysis
tool usetest

Ascent: An Agentic System over the Model Context Protocol for Real-World Clinical Data Analysis

The method replaces a fixed, hand-coded analysis pipeline for epidemiological question answering with an LLM agent that calls medical coding, SQL, and cohort-analysis tools exposed through a shared Model Context Protocol surface. It costs more tool calls and longer runtimes per query, depends on capable (likely API-based) models, and adds orchestration and schema-integration complexity rather than GPU memory. It can fail through unrecognized pharmacoepidemiological errors in population, denominator, and time-window choices, and the reported gains may not transfer outside clinical data domains or to weaker models.

observed signal
0impl.
10stars
—citations
—claimed gain
abstract onlyReview decisionDownload skill ↓Original paper ↗
arXiv 2609.24635
First page of Written as a Record, Read as an Address: What a Forward Pass Leaves in an Operation's KV Cache
otherwatch

Written as a Record, Read as an Address: What a Forward Pass Leaves in an Operation's KV Cache

This is a mechanistic interpretability study, not a deployable technique; it replaces nothing in a production stack, instead analyzing what a forward pass writes into operation-span KV entries (a routing record plus a narrowly localized payload readable at layers 12-17). Its cost is an extra trained reader model, cache recomputation, and a reported drop in open-book accuracy, and it is bounded by training coverage of the operations tested. For adoption the failure mode is categorical: the findings concern how models internally record state in synthetic tasks and do not yield a validated inference-time mechanism, so treating them as a production feature risks misapplication.

observed signal
0impl.
121stars
—citations
—claimed gain
abstract onlyReview decisionDownload skill ↓Original paper ↗
arXiv 2609.24639
First page of Analytical Power-Aware Provisioning for Prefill-Decode Disaggregated AI Inference
serving and batchingwatch

Analytical Power-Aware Provisioning for Prefill-Decode Disaggregated AI Inference

The method replaces profiling- and simulation-based sizing of prefill and decode instance counts with an analytical queueing- and KV-cache-aware model that yields a serving-capacity–power Pareto front. It costs modeling effort and assumptions about workload input/output length distributions, and it does not reduce per-request compute, memory, or latency on a single deployment. It can fail when real workloads deviate from the assumed length distributions or when per-instance power-versus-throughput behavior does not match the fitted model.

observed signal
0impl.
0stars
—citations
—claimed gain
abstract onlyReview decisionDownload skill ↓Original paper ↗
arXiv 2609.24657
First page of Circuit Hypernetworks for Quantum-Augmented Diffusion Language Models
othernot applicable

Circuit Hypernetworks for Quantum-Augmented Diffusion Language Models

HyperQ replaces conventional parameter-efficient fine-tuning (e.g., LoRA) of a frozen masked-diffusion language model with token-conditioned quantum residual branches, where a hypernetwork emits per-token circuit parameters inside each transformer block. The cost is a quantum or simulated quantum execution per token per block (classically exact but linear in qubit count, here up to 64 qubits within a 1.1B backbone), which adds inference-time compute and engineering complexity that neither a 24 GB GPU nor standard third-party APIs currently support. It can fail through simulation overhead negating any quality gain, lack of production tooling for quantum circuit inference, and results that are demonstrated only on a diffusion language model rather than mainstream autoregressive models.

observed signal
0impl.
0stars
—citations
—claimed gain
abstract onlyReview decisionDownload skill ↓Original paper ↗

Research signals

Three views of the same evidence: adoption against attention, publication cadence by research area, and reported gains when coverage is sufficient.

color = research area

01 · attention × adoption

The research frontier

Look to the upper left for independent implementation before mass attention.

compare by
Each point represents one paper. Higher means more independent implementations; farther right means more GitHub stars.Inference-Time Graph Engineering for Multi-Agent LLM Workflows: 2 GitHub stars; 2 independent implementationsProcedural Graphs: Self-Evolving Execution Structures for LLM Agents: 10 GitHub stars; 12 independent implementationsAPEXA: Execution-Integrity Enforcement for Multi-Agent LLM Automation of Synchrotron Data Reduction: 132 GitHub stars; 2 independent implementationsCREDO: Variance-Guided Rubric Evolution for Replay-Corrected Credit Assignment: 121 GitHub stars; 0 independent implementationsMemCalib: Benchmarking and Optimizing Memory Use in LLM Agents: 10 GitHub stars; 1 independent implementationCanonical Procedural Actions: An Auditable Annotation Protocol for Tool-Use Agent Traces: 131 GitHub stars; 1 independent implementationTTSE: A Two-Track Online Self-Evolution Framework: 0 GitHub stars; 0 independent implementationsA Lean and Spec-Driven AI-Assisted Software Development Lifecycle for Applied AI Education: The AI-SDLC Approach: 0 GitHub stars; 0 independent implementationsFew-Shot Demonstrations Elicit the Use of In-Context World Representations in LLMs: 335 GitHub stars; 2 independent implementationsNAVIR: Neuromorphic Audio-Visual Speech Recognition for Robust Human-Robot Interaction on Edge Hardware: 213 GitHub stars; 3 independent implementationsArtificial Structure Function Search: Preserving Artificial Functional Connectivity for Structured Pruning: 0 GitHub stars; 0 independent implementationsARM: Attention with Routed-Memory for Learnable Sparse Control: 0 GitHub stars; 0 independent implementationsActGov: Governing LLM Agent Actions via Policy-Constrained Validation: 0 GitHub stars; 0 independent implementationsBeyond Predictable Paths: Redefining AI Security Incident Reporting for Agents: 0 GitHub stars; 0 independent implementationsNot All Task Vectors Need Equal Rank: Energy-Proportional Allocation for Model Merging: 121 GitHub stars; 0 independent implementationsQLoRA Fine-Tuning of Ministral LLM for Sequence-to-Function Protein Annotation: 0 GitHub stars; 0 independent implementationsPoisson Exchange Beyond Submodularity: Effective Approximation Algorithms for Offline and Online Subset Selection over Matroids: 0 GitHub stars; 0 independent implementationsTaking a Second Look: Correcting Sea Ice Forecasts with Sparse Observations: 0 GitHub stars; 0 independent implementationsAscent: An Agentic System over the Model Context Protocol for Real-World Clinical Data Analysis: 10 GitHub stars; 0 independent implementationsWritten as a Record, Read as an Address: What a Forward Pass Leaves in an Operation's KV Cache: 121 GitHub stars; 0 independent implementationsAnalytical Power-Aware Provisioning for Prefill-Decode Disaggregated AI Inference: 0 GitHub stars; 0 independent implementationsCircuit Hypernetworks for Quantum-Augmented Diffusion Language Models: 0 GitHub stars; 0 independent implementationsMuon Can Outperform Dedicated Continual Learning Methods: 0 GitHub stars; 0 independent implementationsAdapting Tree-Structured Speculative Decoding to DeepSeek-V4 for Efficient Inference: 121 GitHub stars; 0 independent implementationsInference of Unknown Dynamical Components Using Next Generation Reservoir Computing: From Chaotic Systems to Climate Data: 0 GitHub stars; 0 independent implementationsNPU Accelerator: Quantized Real-Time Vehicle Detection on PYNQ-Z1 Using FINN: 0 GitHub stars; 0 independent implementationsComplex KDA: Understanding and Enhancing the Expressivity of Kimi Delta Attention: 1 GitHub star; 1 independent implementationWhen Quantization Preserves Accuracy but Not Evidence: Explanation-Aware Post-Training Quantization for Medical LLMs: 121 GitHub stars; 0 independent implementationsDecoding Guardrails: XAI-Guided Perturbation Analysis of Prompt Injection Detection: 0 GitHub stars; 0 independent implementationsMSI-Bench: Evaluating Multi-Speaker Voice Interaction for Collaborative AI Agents: 132 GitHub stars; 2 independent implementationsSPECTRA: Adaptive Execution of Speculative Decoding on a Runtime-Reconfigurable Tiled Architecture: 0 GitHub stars; 0 independent implementationsOSWorld-Pro: Process-based Evaluation for Computer Use Agents: 131 GitHub stars; 1 independent implementationSLICEChat: Progressive In-Encoder Token Pruning for Whole-Slide Pathology Language Models: 1 GitHub star; 0 independent implementationsTrajectory-Aware Benchmark Subset Selection for Cost-Efficient Software Engineering Agent Regression Testing: 0 GitHub stars; 0 independent implementationsEmergent Collusion in Long-Horizon LLM Agent Interaction: 134 GitHub stars; 2 independent implementationsDolphinBench: Mapping the Pareto Frontier of Agent Memory: 131 GitHub stars; 1 independent implementationRRSI: Regularized Recursive Self-Improvement of Agent Harnesses: 64 GitHub stars; 2 independent implementationsHarness-Zero: Harness Distillation via Agent-as-Harness: 11 GitHub stars; 2 independent implementationsLoRA-generating hypernetworks for efficient on-device LLM generative personalization: 1 GitHub star; 0 independent implementationsonPanda: Efficient Annotation of On-Policy Alignment Data for LLMs and Agents via Token-Level Correction: 11 GitHub stars; 2 independent implementationsCritical-State RL: Diagnosing Trainable States for Multi-Turn Tool Use: 0 GitHub stars; 0 independent implementationsWho Pays for the KV Cache? Attributing Shared AI Inference Spend Across Kubernetes and LLM Provider Bills: 0 GitHub stars; 0 independent implementations

Papers above 1000 stars or 200 citations are not scored. They have already broken through the attention threshold, while Paperraft is designed to identify earlier signals.

02 · cadence

Research areas over time

Compare monthly publication volume on a shared scale and calendar.

coding agentsmax 7Sep 26: 1 paperSep 26Sep 26efficient architecturesmax 7Sep 26: 2 papersSep 26Sep 26agent evaluationmax 7Sep 26: 5 papersSep 26Sep 26KV cachemax 7Sep 26: 1 paperSep 26Sep 26speculative decodingmax 7Sep 26: 2 papersSep 26Sep 26distillationmax 7Sep 26: 1 paperSep 26Sep 26sparsity and pruningmax 7Sep 26: 2 papersSep 26Sep 26memory and contextmax 7Sep 26: 4 papersSep 26Sep 26multi-agent orchestrationmax 7Sep 26: 1 paperSep 26Sep 26othermax 7Sep 26: 7 papersSep 26Sep 26quantizationmax 7Sep 26: 2 papersSep 26Sep 26safety and guardrailsmax 7Sep 26: 5 papersSep 26Sep 26serving and batchingmax 7Sep 26: 1 paperSep 26Sep 26efficient trainingmax 7Sep 26: 6 papersSep 26Sep 26tool usemax 7Sep 26: 2 papersSep 26Sep 26

Every panel uses the same calendar and vertical scale. An empty column means no papers were recorded.

Method under scrutiny

Inspect the highest-ranked paper, every repository behind its signal, and the rule used to classify each one.

Procedural Graphs: Self-Evolving Execution Structures for LLM Agents

memory and context · test · 12 of 13 implementations are independent

Replaces unconstrained next-action generation over an accumulating trajectory with an evolving (procedure, relation, procedure) graph that a guidance model turns into step-level hints biasing, but not dictating, the agent's next action. Costs an extra guidance-model call per decision step plus LLM-refiner passes that contrast failed and successful trajectories to edit the graph against a held-out validation set, adding orchestration complexity and API spend. Can fail if validation sets are small or unrepresentative, letting the refiner overfit graph edits, or if localization of the active node is wrong on long trajectories, injecting misleading guidance; the abstract reports only qualitative 'consistent gains' with no quantified improvement, so the benefit is unverified for any specific workload.

Exclusions and controls

Every exclusion is counted and published.

Exclusion counts were not recorded for this edition.

0 previously indexed papers were rechecked in this run.

Paperraft tracks one signal: how many independent implementations a paper attracts on GitHub after author-owned repositories are removed. This is a discovery signal to investigate, not a measure of implementation quality.

It does not claim that a method works. Paperraft runs no reproduction benchmarks. Reported gains remain author claims until independently tested. Papers above the attention threshold are excluded because this publication is designed to identify research before consensus forms.