PAPERS / REPRODUCTION / ENGINEERING NOTES
Structured readings that extract architecture decisions, reproducible methods, evaluation limits, and what changes when research meets a real system.
READING LIBRARY
Filter by field, reading depth, or the paper’s publication year.
Raven composes model–harness pairs as specialists and asks a Host Agent to plan a dependency DAG. This reading unpacks harness composition, the MAOB benchmark, and what its +10.4/+10.5 percentage-point result actually measures: graph planning match, not successful worker execution.
EfficientAgent asks when moving a KV cache from GPU memory to host memory actually pays off. The answer depends not on one request alone, but on the reuse working set created by concurrent agents between uses. This reading examines capacity prediction, write admission, SWE-bench replay, and hardware boundaries.
Authentic, once-correct documents can make RAG overturn an answer the model already got right. This reading examines 317 cross-domain knowledge reversals, the difference between dates and validity intervals, causal interventions, and a reranker whose gains depend on trustworthy metadata.
A single ReVerPi source-reading campaign shows how a runner can suppress a companion arm after the first arm reaches its request cap. Finite-frame bounds and stratified cost accounting reveal what completed-pair summaries omit, without making a population claim about context projection.
A deep reading of how SpecHarness compiles agent-visible instructions into source-linked obligations and commits state only from qualified evidence, with an examination of SkillsBench and GuideBench results, runtime costs, and limits on attributing gains to sign-off alone.
A close reading of arXiv v1: ten local coding-agent harness pairs, four controlled routes to trace tampering, and the distinct evidentiary roles of native traces, an external file observer, and tool-effect records.
A close reading of the arXiv v1 comparison of a dsh plugin migration skill: what the static diagnostic scores show, how a contract counterexample qualifies them, and where task concentration, LLM judges, and artifact coverage set limits.
A close reading of arXiv 2608.16586 v1: how eight chunking strategies change retrieval quality, indexing throughput, query speed, and memory—and where this benchmark stops.
A deep reading of When Stale Constraints Go Unchecked: immutable provenance, append-only supersession, and stale-consistent decisions under a two-record verification budget. Audits four runs, the held-out correction, denominators, and the Zenodo artifact while bounding the synthetic evidence.
A deep reading of APort Vault that separates payment requests, successful payments, policy decisions, recipient membership, and unpermitted transfers while examining what a pre-action authorization layer does and does not establish.
A deep reading of Trajectory-Aware Benchmark Subset Selection: historical outcome strata, trajectory sanitization, geometric selection, 76 configurations, temporal validation, cost, and transfer limits.
A deep read of MobileCybench: 13 Android apps and 495 executable security probes used to evaluate five coding agents, with careful distinctions among probe triggers, vulnerability attribution, maintainer confirmation, and the limits of the evidence.
A critical reading of Wu et al.'s XYEval (arXiv 2609.23939 v1): controlled XY mutations across five models and six benchmark suites, with trace analyses of how misleading suggestions affect task completion, communication, and tool trajectories—and the limits of generators and judges.
A deep read of The RAT: how a joint Bayesian model separates retrieval, abstention, and answer correctness, and what its annotation-budget results, judge calibration, partial retrieval analysis, and controlled Wikipedia scope do and do not establish.
A deep read of Prime Agent (arXiv 2608.23552 v1): its L0–L3 state hierarchy, persistent REPL, recursive subagents, and Continual Harness—alongside author-reported evaluations, a specification exploit retained through refinement, and the unsandboxed execution boundary.
A deep reading of Self Improvement via Fast Tree-search (arXiv:2609.19526): pairwise LLM judging, regularized Bradley–Terry ranking, and asynchronous tree search decide which agent patches deserve expensive benchmark verification.
A deep reading of BioPhys-Bridge (arXiv:2609.19180 v1): a benchmark that makes evidence IDs, values and units, physical equations, assumptions, biological mechanisms, and next decisions explicit, while separating attribution from scientific correctness.
A deep reading of RAFT: A Stateful Retrieval-Augmented Framework for Troubleshooting Agents (arXiv:2609.20754): closed support cases become state-transition trajectories, intermediate entries are retrieved, and the parent case is returned with its matched anchor.
A deep read of Learning When to Trust via Selective Context Preference Optimization: MIST holds each reasoning item fixed while varying clean, misleading, correct-context, and irrelevant-context signals; SC2W isolates clean-correct flips, and SCOPE reduces them with balanced DPO preference quartets while retaining the paper's text-only, contamination, and deployment-prevalence boundaries.
A deep read of Bounded Agents (arXiv:2608.15888 v1): the Agentic Principal Chain, six conjunctive authorization conditions, and composition closure for constraining delegated authority and cross-step side effects—together with the complete-restriction, serialized-admission, enforcement, and utility boundaries.
A deep reading of Bridging the Question-Answer Gap in Retrieval-Augmented Generation: Hypothetical Prompt Embeddings. HyPE precomputes hypothetical questions for every chunk and turns query-to-document retrieval into question-to-question matching, with strong six-dataset results but real indexing, freshness, chunking, and single-generator boundaries.
A deep reading of After the Party’s OpenClaw and ClawHub ecosystem study: 91 days of explosive growth, download concentration, the reviewability gap, privilege evidence, scanner disagreement, and the governance method that can transfer without pretending the rates do.
A deep reading of Predicting Partial Answer Quality and Utility in Agentic Retrieval-Augmented Generation: turning intermediate answer quality, utility, and trajectory signals into an early-stopping controller, then checking the savings and transfer boundary.
A deep reading of SilentProbe (arXiv:2609.00035 v1): from OpenAPI constraint gaps and live differential probes to agent false negatives, separating how disclosure and machine-readability determine whether a tool fails honestly.
Showing 24 of 80 notes
EXPLORE BY RESEARCH TOPIC
Daily notes remain below. These topic paths collect every relevant deep dive, even when a paper belongs to more than one engineering question.
Build a foundation for reading sequence-transduction papers through classic encoder–decoder and self-attention architectures.
Question:Where does a sequence model change its control point, and which evidence still transfers to today’s systems?
Build a foundation for reading methods papers through classic vision architectures, training choices, and evaluation evidence.
Question:Which design choices and evidence from a classic model still transfer to today’s systems?
Explore engineering trade-offs in multimodal retrieval, GraphRAG, ranking, evidence discovery, and traceable answers.
Question:How can a system keep finding verifiable evidence as its corpus and tool surface grow?
Examine whether agents are actually reliable through task outcomes, trajectories, runtime signals, and benchmarks.
Question:Beyond the final answer, which signals must we observe and verify before trusting that an agent completed its work?
Understand memory layers, index updates, durable state, and safe system adaptation for long-horizon tasks.
Question:What should a system remember, update, or forget to improve over time without accumulating errors?
Examine control boundaries for agents in real execution environments: risk injection, auditing, permissions, and rollback.
Question:Beyond task completion, how do we show that an agent’s process, memory, and side effects remain controlled?
Study how tool discovery, schema context, coding tasks, and verifiable execution jointly shape agent performance.
Question:When an agent must choose tools, edit code, and execute actions, how can we reduce context and operational errors?
START WITH A PATH
Three bounded routes turn the library into a sequence. Start at your level and stop when you have the engineering answer you need.
If you already have ReAct through Reflexion in hand and need a spine for how those classics connect to later notes, start from the How to Read AI Agent Papers: From CoT and WebGPT to ReAct rather than Path 3. If you just finished DPR and Lewis RAG and need the retrieval spine onto leaves already on this site, use the How to Read RAG Papers: From Dense Retrieval (DPR) to Lewis RAG rather than treating Path 2 as that genealogy.
Start with AlexNet architecture and training evidence, then ResNet residual shortcuts for trainable depth, YOLO unified real-time detection as single-pass regression, Transformer sequence transduction with self-attention, InstructGPT instruction alignment via SFT, a reward model, and PPO, then Speculative Decoding lossless inference via draft-and-verify parallel decoding, while separating methods, experiments, and historical constraints.
+ 4 more steps
Start this pathStart from 2020 retrieval-augmented pre-training (REALM), then the cheaper dense retriever (DPR), the RAG ancestor and Self-RAG when-to-retrieve, then multimodal parsing, tool retrieval, memory, GraphRAG, scaling, and runtime controls.
+ 24 more steps
Start this pathStart with canonical 2017–2023 agent method ancestors (CoT → WebGPT → ReAct → Toolformer → Gorilla → IPI → SWE-bench → Reflexion → MemGPT → Generative Agents), then later tool/eval, runtime, safety, and platform leaves.
+ 37 more steps
Start this pathRESEARCH EXCHANGE
