Hermes Wiki
TechEnglishVocab/tech_english_vocab

📚 Tech English Vocabulary

A running collection of everyday tech vocabulary — software development, AI/ML engineering, research, and professional/work usage in a tech context. Words worth knowing cold, not just recognizing. Not a general-English or non-tech vocab list.

How This Grows

  • Scope: tech only — software development, AI/ML engineering, research (papers, arXiv, whitepapers, innovations). No general-English, casual, or non-tech vocabulary.
  • Recency: automated entries only pull from sources published within the last 2 years — no scraping old/archival papers.
  • Sorting: keep the table in alphabetical order by word.
  • Multiple meanings: if a word has more than one relevant tech-context meaning, add each as its own row rather than combining senses in one cell.
  • Source link: automated entries link the word itself (first column) to the exact source it was pulled from — [Word](source-url) (type). Manually-added entries (like the seed row below) don't need one.
  • Cell format: the Meaning + Sentence column is written as **Meaning**: ...<br>**Examples**: ... — meaning and example stacked on separate lines within the same cell.
  • Source: manual entries any time, plus a weekly automated pass that reads recent (≤2yr old) tech/arXiv papers, whitepapers, and innovation coverage and appends new words it encounters, each linked back to its source. See weekend-vocab-scheduler for the scheduling setup.
  • Unlike the rest of TechResearch/, this is a living, append-only document — not read-only research. Update and append freely; just preserve alphabetical order and don't delete existing entries without a reason.

Vocabulary

Word (type) Meaning + Sentence
Adversarial (adjective) Meaning: Designed to actively probe or stress-test a system by anticipating and exploiting its weaknesses, rather than assuming only benign inputs.
Examples: PseudoBench is an adversarial benchmark that feeds agentic auto-research pipelines plausible-looking pseudoscientific claims to see whether they get accepted or correctly flagged.
Affinity (noun) Meaning: The tendency for related requests or tasks to be consistently routed to the same resource so as to exploit locality and avoid redundant setup cost.
Examples: SAGA's session-affinity batching keeps correlated agent requests on the same GPU so cached state can be reused across steps.
Alignment (noun) Meaning: The degree to which one thing, such as a model's generated output, correctly corresponds to or matches something else it is meant to reflect.
Examples: The paper's Semantic Alignment Score measures how well a model's generated reasoning matches the topology of the retrieved knowledge graph.
Attribution (noun) Meaning: The process of identifying which specific component, agent, or step in a system was responsible for causing an observed outcome, especially a failure.
Examples: Failure attribution in a multi-agent system tries to pinpoint which agent's action first sent the task off course.
Bottleneck (noun) Meaning: A point in a system where the flow of work or information is constrained, limiting the throughput of the whole pipeline.
Examples: DeepCode treats an LLM's limited context window as the central bottleneck to manage when synthesizing a full codebase from a paper.
Co-locate (verb) Meaning: To place related tasks, processes, or data on the same physical machine or resource so they can share state or reduce communication overhead.
Examples: The scheduler co-locates correlated requests from the same agent workflow while still balancing load across the cluster.
Compaction (noun) Meaning: The process of reducing the size of accumulated data, such as conversation or context history, by summarizing or discarding less-relevant parts so it fits within a fixed budget.
Examples: The agent relies on adaptive context compaction to keep long sessions within the model's token budget.
Contaminate (verb) Meaning: To introduce corrupting or misleading material into a body of data, research, or evidence, undermining its reliability.
Examples: The authors warn that autonomous research agents could rapidly generate plausible yet misleading studies that contaminate academic literature.
Contemporaneous (adjective) Meaning: Occurring or recorded at the same time as the events it describes, rather than reconstructed from memory afterward.
Examples: The case study's empirical record includes 88 contemporaneous field notes written during the 12-week engineering engagement itself.
Curate (verb) Meaning: To carefully select, organize, and vet items, such as data or examples, for a specific purpose, ensuring their quality and relevance.
Examples: PseudoBench contains 200 curated pseudoscientific claim-evidence pairs spanning five domains.
Decompose (verb) Meaning: To break a complex problem, system, or task down into smaller, more manageable parts.
Examples: Traditionally, human engineers decompose a problem into subtasks before writing the code that implements each one.
Deliberation (noun) Meaning: The process of carefully weighing or reasoning through options before acting, especially by simulating possible outcomes first.
Examples: The agent grounds its deliberation in future-state prediction from an internal world model before committing to an action.
Deterministic (adjective) Meaning: Producing the same output every time given the same input, with no randomness or unpredictability in the process.
Examples: The paper contrasts agentic software with traditional deterministic software, where the same code path always executes the same pre-written logic.
Distillation (noun) Meaning: The process of condensing a large or complex source of information down to its most essential, task-relevant content.
Examples: DeepCode compresses a lengthy source paper into a compact blueprint through a distillation step before any code is generated.
Empirical (adjective) Meaning: Based on direct observation or experiment/data rather than on theory alone.
Examples: The study takes an empirical approach, mining thousands of real rule files instead of theorizing about how developers configure AI IDEs.
Emulate (verb) Meaning: To closely imitate or reproduce the behavior, qualities, or standards of something else.
Examples: The AARR benchmark series asks whether agents can emulate the thoroughness and nuanced judgment that characterize human researchers.
Enumerate (verb) Meaning: To identify and list items one by one, typically as a systematic first step before acting on them.
Examples: After gaining folder access, the script enumerates every local file before reading and encrypting each one.
Exfiltrate (verb) Meaning: To covertly transfer data out of a system without authorization, typically as part of an attack.
Examples: The malicious script silently exfiltrated the contents of the user's folder before encrypting the originals.
Extortion (noun) Meaning: The practice of obtaining something, especially money, through coercion or threats.
Examples: Once the files were encrypted, the ransomware displayed an extortion note demanding payment for their release.
Faithfulness (noun) Meaning: In generated text, the degree to which the output accurately reflects the retrieved or source content rather than contradicting or embellishing it.
Examples: The paper targets faithfulness hallucinations, where a RAG system's answer is internally coherent but does not match what its retrieved documents actually say.
Harness (noun) Meaning: The runtime layer of an agent system that dispatches tool calls, manages context, and enforces safety constraints while the agent executes.
Examples: The paper distinguishes the scaffolding that assembles the agent before the first prompt from the harness that orchestrates tool dispatch and safety enforcement at runtime.
Heterogeneous (adjective) Meaning: Composed of parts or elements that differ from one another, as opposed to being uniform.
Examples: Dyserve schedules each step of a workflow across a heterogeneous backend pool made up of different models and hardware.
Hypothesize (verb) Meaning: To propose a tentative explanation for an observed effect, intended to be tested rather than simply assumed true.
Examples: The researchers hypothesize that an AGENTS.md file improves efficiency because agents spend less time on exploratory navigation of the codebase.
Inspectable (adjective) Meaning: Able to be examined or reviewed closely enough to verify its correctness or reasoning.
Examples: The paper argues that AI-mediated development must remain inspectable, correctable, and maintainable even as generating code itself becomes cheap.
Instrumental (adjective) Meaning: Serving as a means or tool toward an end, rather than being valuable in itself.
Examples: In agentic software, code is treated as an instrumental resource the agent generates and discards at runtime rather than a fixed artifact.
Irreversible (adjective) Meaning: Impossible to undo or reverse once it has happened.
Examples: Most misalignment episodes only cost the developer time and trust rather than causing irreversible system damage.
Lazy (adjective) Meaning: Deferring an operation, such as loading or computing a value, until it is actually needed rather than doing it upfront.
Examples: The agent uses lazy tool discovery so it only loads a tool's full schema once the agent actually decides to call it.
Long-horizon (adjective) Meaning: Spanning many steps or a long stretch of time before reaching completion, rather than being resolved in a single short step.
Examples: Agent scaffolding has grown sophisticated enough that agents now show real proficiency on long-horizon coding tasks.
Misalignment (noun) Meaning: A mismatch between a system's actions or outputs and what the user actually intended.
Examples: The authors identify seven recurring forms of misalignment between coding agents and the developers who direct them.
Orchestrate (verb) Meaning: To coordinate multiple components or steps so that they work together toward a single outcome.
Examples: DeepCode orchestrates compression, indexing, retrieval, and error correction as four coordinated stages of one pipeline.
Permutation (noun) Meaning: One specific ordering among all the possible ways a set of items can be arranged.
Examples: The researchers found that reordering the same retrieved documents into a different permutation could flip the model's final answer.
Reactive (adjective) Meaning: Responding directly to immediate inputs with minimal forward planning, as opposed to deliberating or simulating ahead of time.
Examples: The paper contrasts purely reactive policies, which respond to each state as it arrives, with agents that pause to plan before acting.
Scaffolding (noun) Meaning: The surrounding structure of prompts, tools, and control logic built around a model to support it in carrying out complex tasks.
Examples: As foundation models advance and agent scaffolding becomes increasingly sophisticated, agents have grown proficient at long-horizon coding tasks.
Sycophancy (noun) Meaning: The tendency to give excessive, uncritical agreement in order to please rather than to be accurate — in LLMs, the tendency of a model to reinforce a user's stated assumption instead of independently verifying it.
Examples: Researchers found that user-driven sycophancy led the model to endorse a flawed root-cause theory instead of asking clarifying questions.
Synthesize (verb) Meaning: To combine two or more distinct elements, ideas, or substances to form a single, complex new whole.
Examples: a laboratory synthesizing a new medication.
Taxonomy (noun) Meaning: A system for classifying things into categories and subcategories according to their shared characteristics.
Examples: Mining 7,310 rules from 83 open-source projects, the researchers built a taxonomy of 5 primary and 25 secondary rule categories.
Temporal (adjective) Meaning: Relating to time or to the order and timing of events, as opposed to their content or meaning.
Examples: The framework builds a temporal semantic model of execution traces to see how errors propagate between agents over time.
Traceable (adjective) Meaning: Able to be tracked back through a sequence of steps or transformations to its origin.
Examples: QPipe turns a natural-language requirement into a traceable quantum-application workflow, so each generated artifact can be linked back to its source requirement.
Triangulate (verb) Meaning: To cross-check or validate a finding by combining evidence from multiple independent sources or methods.
Examples: The team triangulated the mined rule files with survey responses from 99 practitioners to see whether stated priorities matched actual practice.
Work stealing (noun) Meaning: A load-balancing technique in which idle workers take queued tasks from busier workers instead of waiting for centrally-assigned work.
Examples: SAGA combines session-affinity batching with work stealing so no GPU sits idle while another is overloaded.

Hermes Wiki