Glossary: Full Names, Meanings, and Relationships¶
This glossary is a decoding aid. Every chapter must still expand an uncommon abbreviation at its first meaningful use and immediately explain what the concept does. A full name answers “what do the letters mean?”; the explanation answers the more important question, “what does this object do here?”
Learning paradigms¶
| Term | Full name | Plain meaning and boundary |
|---|---|---|
| AI | Artificial Intelligence | The broad field of machines performing tasks associated with intelligence. Machine learning is one approach within it. |
| LM / LLM | Language Model / Large Language Model | A model assigns probabilities to token sequences. “Large” has no fixed parameter threshold. |
| VLM / VLA | Vision-Language Model / Vision-Language-Action model | A VLM processes images and language; a VLA additionally produces actions, commonly for robotics. |
| RL | Reinforcement Learning | Learning a behavior policy from scalar rewards produced through interaction. |
| agentic RL | Agentic Reinforcement Learning | RL in which an LLM acts over multiple steps, often using tools and receiving delayed environment feedback. It is a setting, not one algorithm. |
| SFT | Supervised Fine-Tuning | Maximum-likelihood training on supplied target responses or trajectories. It imitates demonstrations rather than directly maximizing a reward. |
| CPT | Continual Pretraining | Additional next-token training of an existing checkpoint, often on a new domain, higher-quality mixture, or longer context. A CPT token count is an incremental stage, not automatically the model's total lifetime exposure. |
| agentic continued pretraining | — | Continued next-token training on serialized goals, actions, tool calls, observations, and revisions. It teaches interaction-shaped priors at scale but remains likelihood training rather than online environment RL. |
| RLHF | Reinforcement Learning from Human Feedback | RL in which human judgments directly or indirectly define reward, often through a learned reward model. |
| RLAIF | Reinforcement Learning from AI Feedback | RL whose labels, critiques, or rewards are produced by another AI system. |
| CAI | Constitutional AI | Uses written principles to generate critiques, revisions, or AI preference labels; the cited Anthropic pipeline combines these stages with SFT and RL. |
| RLVR | Reinforcement Learning with Verifiable Rewards | RL using mechanically checkable rewards such as unit tests, exact answers, or proof verification. “From verifiable rewards” is also used. |
| BC / IL | Behavioral Cloning / Imitation Learning | BC is supervised imitation of demonstrated actions; IL is the broader family. |
| KD / OPD | Knowledge Distillation / On-Policy Distillation | KD trains a student from a teacher. OPD evaluates the teacher on states sampled by the student, reducing state-distribution mismatch. |
| MOPD | Multi-Domain / Multi-teacher On-Policy Distillation | NVIDIA uses “Multi-Domain” in Cascade 2 and “Multi-teacher” in Nemotron 3 Ultra for closely related student-state distillation: the student samples states and a domain teacher supplies token distributions on those states. |
| CoT | Chain of Thought | Intermediate language or latent reasoning before an answer. Long-CoT denotes deliberately extended traces. |
| TTRL | Test-Time Reinforcement Learning | Target-specific policy adaptation performed during evaluation; AlphaProof uses generated variants around a target theorem. It differs from merely sampling more candidates. |
| NLL | Negative Log-Likelihood | For a target sequence, \(-\log p_\theta(y\mid x)=-\sum_t\log p_\theta(y_t\mid x,y_{<t})\); minimizing it performs maximum-likelihood imitation. Adding NLL on successful RL samples directly increases the likelihood of their tokens. |
MDP, POMDP, and trajectories¶
Markov Decision Process (MDP)¶
An MDP is the mathematical model
\[
\mathcal M=(\mathcal S,\mathcal A,P,r,\gamma,\rho_0),
\]
where \(\mathcal S\) is the state space, \(\mathcal A\) the action space, \(P(s'\mid s,a)\) the transition law, \(r\) the reward, \(\gamma\) the discount factor, and \(\rho_0\) the initial-state distribution. “Markov” means the current state contains everything needed to predict the next state once the action is known.
Partially Observable Markov Decision Process (POMDP)¶
A POMDP is an MDP in which the agent cannot directly observe the complete state. Instead it receives \(o_t\sim O(\cdot\mid s_t)\) and must infer what matters from its interaction history or learned memory. Tool-using LLM agents are normally closer to POMDPs: hidden web state, omitted files, asynchronous events, and truncated context make the true environment state only partially visible.
| Term | Meaning |
|---|---|
| state \(s_t\) | All environment information that makes future dynamics independent of the earlier past, conditional on the next action. It may be hidden. |
| observation \(o_t\) | Information exposed to the agent, such as a prompt, terminal output, or screenshot. |
| action \(a_t\) | A decision affecting the environment: a token, message, tool call, or structured command. |
| policy \(\pi_\theta(a\mid h)\) | A parameterized probability distribution over actions given observable history \(h\). |
| trajectory \(\tau\) | One ordered record of observations, actions, rewards, and termination signals. |
| episode / horizon | One complete or truncated interaction / its maximum or realized number of decision steps. |
| task-completion time horizon | The human-task duration at which an evaluated agent is predicted to achieve a stated success probability over a specified task distribution. It measures task difficulty under one protocol, not agent wall-clock duration or universal job automation. |
| return \(G_t\) | Discounted future reward, \(G_t=\sum_{k\ge0}\gamma^k r_{t+k}\). |
| value \(V^\pi(s)\) | Expected return from a state under policy \(\pi\). |
| action value \(Q^\pi(s,a)\) | Expected return after action \(a\), then following \(\pi\). |
| advantage \(A^\pi(s,a)\) | How much better an action is than the policy average: \(Q^\pi(s,a)-V^\pi(s)\). |
| credit assignment | Determining which earlier actions deserve responsibility for a delayed outcome. |
| on-policy / off-policy | Data comes from the current policy / from a teacher, replay buffer, older checkpoint, or other behavior policy. |
Policy optimization and estimators¶
| Term | Full name | Plain meaning and relationship |
|---|---|---|
| REINFORCE | A Monte Carlo policy-gradient estimator; not an acronym in the original paper | Multiplies sampled log-probability gradients by returns or advantages. It is simple and often high variance. |
| PG | Policy Gradient | Direct differentiation of expected return with respect to policy parameters. |
| A2C | Advantage Actor-Critic | A synchronous actor-critic policy-gradient method; an advantage estimate subtracts a learned value baseline. |
| TRPO | Trust Region Policy Optimization | Constrains updates by a Kullback–Leibler divergence trust region. |
| PPO | Proximal Policy Optimization | Reuses rollouts while clipping or penalizing policy-ratio movement. Actor-critic PPO learns a value function. |
| GAE | Generalized Advantage Estimation | A geometrically weighted mixture of temporal-difference residuals, trading bias against variance through \(\lambda\). |
| TD | Temporal Difference | Bootstraps from a later value estimate instead of waiting for the complete return. |
| RLOO | REINFORCE Leave-One-Out | Uses the other responses to the same prompt as a reward baseline; no critic is required. |
| GRPO | Group Relative Policy Optimization | DeepSeek's publicly introduced critic-free PPO-like method; it derives relative advantages from a response group for one prompt. |
| Dr. GRPO | A named correction in Understanding R1-Zero-Like Training | Removes response-length and group-standard-deviation normalizations identified as bias sources. The label has no standard expanded phrase. |
| DAPO | Decoupled Clip and Dynamic sAmpling Policy Optimization | A GRPO-family recipe with asymmetric clipping, dynamic sampling, token-level loss, and overlong-reward shaping. |
| VAPO | Value-based Augmented Proximal Policy Optimization | A long-reasoning actor-critic recipe designed around value-model bias, heterogeneous lengths, and sparse rewards. |
| ProRL | Prolonged Reinforcement Learning | NVIDIA's study of continuing verifiable RL for more than two thousand steps with KL control, curriculum changes, and reference/optimizer resets. |
| GSPO | Group Sequence Policy Optimization | Uses a sequence-level importance ratio rather than independently clipping token ratios. |
| SAPO | Soft Adaptive Policy Optimization | Replaces hard ratio clipping with a sigmoid-shaped token gate whose gradient decays smoothly away from the behavior policy; it uses separate positive/negative-advantage temperatures. |
| SAO | Single-Rollout Asynchronous Optimization for Agentic Reinforcement Learning | Consumes each long trajectory when it finishes, uses direct double-sided importance masking, and restores a critic for group-size-one training. The 2026 authors report deployment in GLM-5.2. |
| CompactionRL | Reinforcement Learning with Context Compaction for Long-Horizon Agents | Jointly optimizes ordinary agent actions and model-generated context summaries under downstream task reward, using token-level PPO loss and cross-segment GAE for variable compacted traces. |
| DIS | Direct Double-Sided Importance Sampling | SAO's direct rollout-to-current-policy importance ratio with strict two-sided rejection outside a permitted interval. It is masking, not PPO's saturated clipping. |
| PARL | Parallel-Agent Reinforcement Learning | Moonshot's Kimi K2.5 method for learning an orchestrator that delegates to parallel subagents and aggregates their results. |
| RoC / GRPO-RoC | Resample-on-Correct / GRPO with Resample-on-Correct | rStar2-Agent oversamples a rollout group, preserves a failure subset, and selects cleaner successful tool trajectories before the GRPO update. |
| GiGPO | Group-in-Group Policy Optimization | Adds a second, step-relative comparison group from repeated environment states observed across trajectories, without sampling new counterfactual actions. |
| MAGIC | Multi-stage Adaptive entropy scheduling for GRPO In Convergence | Skywork-OR1's recipe combining staged length budgets, strict policy freshness, token-global loss, data filtering, and active entropy control. |
| RPO | Reward-aware Preference Optimization | NVIDIA's framework for expressing preference objectives and reward information under a shared formulation; record the specific RPO variant because the acronym is not one universal loss. |
| DPO | Direct Preference Optimization | Turns preference pairs into a classification-style policy objective without an online RL loop or separate scalar reward model. |
| IPO | Identity Preference Optimization | A preference objective designed to avoid some overfitting behavior of logistic DPO. |
| KTO | Kahneman–Tversky Optimization | A prospect-theory-inspired objective that can train from unpaired desirable/undesirable labels. |
| ORPO | Odds Ratio Preference Optimization | Adds an odds-ratio preference term to supervised likelihood, normally without a separate reference model. |
| PRIME | Process Reinforcement through IMplicit rEwards | Learns token-level process rewards from outcome-labeled rollouts. |
| V-trace | A truncated importance-sampling return estimator | Corrects trajectories from lagging actors toward the current learner policy. |
| IMPALA | Importance Weighted Actor-Learner Architecture | A distributed actor/learner architecture that introduced V-trace. |
Reward and evaluation vocabulary¶
| Term | Full name | Meaning |
|---|---|---|
| RM / ORM / PRM | Reward Model / Outcome Reward Model / Process Reward Model | An RM predicts quality; an ORM scores the final outcome; a PRM scores intermediate steps. |
| GenRM / GRM | Generative Reward Model | Generates a critique or structured judgment instead of only a scalar. Both abbreviations occur in primary sources. |
| RBR | Rule-Based Rewards | OpenAI's system of human-authored propositions graded by a fixed language model and linearly combined into a reward; it is not merely deterministic if statements. |
| IBT | Inter-Temporal Bradley-Terry | A preference model over progress at two different prefixes of the same embodied trajectory. |
| ESS | Effective Sample Size | For nonnegative, not-all-zero importance weights, \((\sum_i w_i)^2/\sum_i w_i^2\); it equals the sample count when weights are equal and falls toward 1 as a few samples dominate. |
| IPT | Isomorphic Perturbation Test | A metamorphic reward audit that applies a meaning-preserving renaming or transformation and checks whether a genuinely general solution transforms consistently. |
| LLM judge | Large Language Model judge | A model prompted or trained to evaluate another output; it is fallible and not automatically ground truth. |
| instruction following | — | A family of capabilities for carrying out explicit or implied directives under scope, order, format, context, and trust constraints. Always identify whether the claim concerns atomic rules, multi-turn retention, hierarchy, tools, or end-to-end task behavior. |
| steerability | — | How predictably a model's behavior changes in response to legitimate instructions or controls. It includes instruction following but also sensitivity, literalness, style, refusal boundaries, and stability across runs. |
| instruction pretraining | — | Adding instruction–response structure during from-scratch or continued pretraining rather than waiting for post-training SFT. It operates at a different scale and can change both later tunability and base-model regressions. |
| instruction hierarchy | — | A trust ordering among instruction sources. In OpenAI's documented hierarchy, system instructions outrank developer, user, and tool content; correct behavior can require ignoring a conflicting lower-priority command. |
| instruction contract / task ledger | — | A versioned representation of active requirements, prohibitions, order, scope, authority, completion evidence, supersession, and approvals. It helps context management but must be checked against the original request. |
| selective clarification | — | Choosing whether to ask a minimal decision-changing question, proceed, refuse, or use a tool. It must balance silent guessing against unnecessary interaction and be judged by downstream task success. |
| automatic prompt optimization | — | Searching instruction wording, demonstrations, or an LM pipeline against a development metric. The optimized prompt and retriever are learned artifacts that can overfit or leak evaluation cases. |
| activation steering | — | Adding a learned direction to internal model activations at inference to change behavior. It requires white-box access and does not provide a hard policy or authorization guarantee. |
| non-interference / scope adherence | — | Satisfying a request without changing unrelated content, state, files, tools, or policy. A substantively correct result can still fail by causing an unauthorized extra effect. |
| constraint-level accuracy | — | The fraction of individual instructions satisfied. It gives partial credit and must be paired with all-constraints accuracy when one missed rule can invalidate the task. |
| all-constraints / prompt-level accuracy | — | The fraction of prompts for which every required constraint passes. It exposes conjunctive failure but does not show which rule failed. |
| constrained decoding | — | Inference-time masking of tokens that would violate a regular expression, automaton, grammar, or schema. It can guarantee supported syntax, not semantic correctness or authorization. |
| task / instruction / state oracle | — | Separate checks for whether the requested outcome is correct, every active rule was followed, and only authorized external state transitions occurred. |
| tool relevance | — | Whether a supplied tool is applicable and necessary for the request. Function-call syntax may be perfect while the model calls the wrong tool or should call none. |
| IFEval | Instruction-Following Evaluation | A reproducible benchmark of roughly 500 prompts and 25 program-verifiable instruction types. Its strict and loose scores measure precise constraints, not complete product reliability. |
| IFBench | Instruction-Following Benchmark | Ai2's harder benchmark with 58 new out-of-distribution, program-verifiable constraints combined with held-out user prompts; the standard score remains narrower than multi-turn, hierarchy, or tool-policy reliability. |
| MultiChallenge | — | Scale's multi-turn benchmark covering instruction retention, inferred user memory, versioned editing, and self-coherence; it uses instance-specific rubrics and an LLM judge. The authors do not expand the name as an acronym. |
| KL | Kullback–Leibler divergence | A directional mismatch, \(D_{\mathrm{KL}}(p\Vert q)=\mathbb E_p[\log p-\log q]\), often used to limit policy drift. |
| TV | Total Variation distance | A symmetric distance, \(\tfrac12\sum_x\lvert p(x)-q(x)\rvert\). |
| pass@k | pass at \(k\) | Probability that at least one of \(k\) sampled candidates passes; it requires a sampling protocol. |
| IoU / F1 | Intersection over Union / harmonic mean of precision and recall | Overlap metric for regions / class metric \(2PR/(P+R)\). |
Factuality, grounding, and uncertainty¶
| Term | Full name | Meaning |
|---|---|---|
| hallucination | — | An imprecise umbrella for plausible but false, unsupported, or source-unfaithful generation. A useful report names the failed reference: world fact, supplied context, citation, tool result, image, code, or another authority. |
| factuality | — | Whether a claim agrees with the best available evidence about the world. This differs from faithfulness to a supplied source. |
| faithfulness / grounding | — | Whether an output is supported by the supplied document, retrieved passage, image, database value, or tool observation. A response can be grounded in a source that is itself false. |
| attribution | — | A trace from a generated claim to a source and location. Attribution is useful only when the source exists, the cited span entails the claim, material claims are covered, and the source is appropriate. |
| parametric knowledge | — | Information represented in model weights. It lacks the simple record identity, update, access-control, and provenance properties of a database. |
| non-parametric memory | — | An external corpus, index, database, or other store consulted at inference; “non-parametric” does not mean error-free or unlearned. |
| evidence sufficiency | — | Whether the available evidence directly covers every material part of a request with the correct entity, date, version, authority, and qualifiers. Retrieval relevance alone is insufficient. |
| citation entailment | — | Whether the meaning of a cited span actually supports the nearby claim, including its negation, magnitude, population, date, and qualifiers. |
| calibration | — | Agreement between stated or scored confidence and empirical correctness over a defined distribution. A confident tone is not a calibrated probability. |
| selective prediction / abstention | — | A policy that answers only above a calibrated evidence or confidence threshold and otherwise asks, abstains, or escalates; reliability improves by trading some coverage. |
| coverage / selective risk | — | The fraction of requests answered / the error rate among those answered. Report them together so refusing everything cannot appear perfectly reliable. |
| semantic entropy | — | Uncertainty measured over clusters of semantically equivalent sampled answers rather than surface word sequences. It detects instability-like confabulations but not every consistent falsehood. |
| RARR | Retrofit Attribution using Research and Revision | A post-generation method that searches for supporting sources and minimally revises unsupported text. |
| Self-RAG / CRAG | Self-Reflective Retrieval-Augmented Generation / Corrective Retrieval-Augmented Generation | Self-RAG trains retrieval and reflection decisions into generation; CRAG evaluates retrieval quality and triggers correction or broader search. |
| FActScore | Factual precision in Atomicity Score | Decomposes long-form output into atomic claims and measures the fraction supported by a knowledge source. |
| SAFE | Search-Augmented Factuality Evaluator | Uses a search-capable LLM agent to decompose and verify long-form claims; search and evaluator errors still require audit. |
Transformer architecture¶
| Term | Full name | Meaning |
|---|---|---|
| MHA / MQA / GQA | Multi-Head / Multi-Query / Grouped-Query Attention | MHA has separate heads; MQA shares one key/value head; GQA shares fewer key/value heads across groups of query heads. |
| MLA | Multi-head Latent Attention | DeepSeek's compressed latent key/value representation for lower inference cache and bandwidth. |
| DSA | DeepSeek Sparse Attention | DeepSeek's long-context mechanism that selects a limited set of relevant tokens. |
| KDA | Kimi Delta Attention | Moonshot's gated delta-rule linear attention. |
| MoBA | Mixture of Block Attention | Moonshot's parameter-free sparse routing of each query to selected context blocks. |
| AttnRes | Attention Residuals | Moonshot's learned attention over earlier residual streams instead of a fixed cumulative sum. |
| LatentMoE | Latent Mixture of Experts | Projects routed token activations into a narrower shared latent space before expert dispatch and computation, reducing routed communication and expert-weight bandwidth. |
| Quantile Balancing | — | An MoE load-allocation family that derives routing offsets or thresholds from router-score quantiles rather than a hand-tuned auxiliary-loss coefficient or heuristic update rate; K3's exact algorithm was not yet public at the cutoff. |
| SiTU | Sigmoid Tanh Unit | K3's named activation-control component; its exact public equation was still pending at the 2026-07-22 cutoff. |
| MoE | Mixture of Experts | Routes each token to a small subset of feed-forward experts, increasing total parameters without proportional active compute. |
| FFN / MLP | Feed-Forward Network / Multilayer Perceptron | The per-token nonlinear sublayer / the general stack of linear layers and nonlinearities. |
| MTP | Multi-Token Prediction | Predicts several future tokens from one position, potentially improving representations and speculative decoding. |
| FIM | Fill-in-the-Middle | Generates a missing span conditioned on both its prefix and suffix. |
| RoPE | Rotary Position Embedding | Encodes relative position by rotating query/key channels. |
| YaRN | Yet another RoPE extensioN method | Extends RoPE context through frequency scaling and attention-temperature adjustment. |
| ALiBi | Attention with Linear Biases | Adds head-specific distance biases to attention logits. |
| RMSNorm | Root Mean Square Layer Normalization | Normalizes magnitude without subtracting the mean. |
| SwiGLU / GeGLU | Swish-Gated / Gaussian Error Linear Unit Gated Linear Unit | Two gated feed-forward activations. |
| QK-Norm | Query-Key Normalization | Normalizes attention queries and keys before the dot product. |
| KV cache | Key-Value cache | Stores earlier attention keys/values to avoid recomputing them during decoding. |
| QAT | Quantization-Aware Training | Simulates low precision during training so the model adapts before quantized inference. |
| Muon | MomentUm Orthogonalized by Newton-Schulz | A matrix optimizer that applies approximate orthogonalization to gradient momentum; Moonshot's scalable variant adds weight decay and shape-aware update scaling. |
Data and representation¶
| Term | Full name | Meaning |
|---|---|---|
| training-token exposure | — | The number of token positions processed by an optimizer stage. Repeated epochs and resampled documents count again, so exposure is not necessarily unique corpus size; fresh pretraining, CPT, multimodal stages, and post-training must be identified separately. |
| BPE / BBPE | Byte Pair Encoding / Byte-Level Byte Pair Encoding | Frequent-pair subword merging / the byte-based variant that can represent any input. |
| OCR | Optical Character Recognition | Converts pixels or scanned pages into text. |
| VLM / LMM | Vision-Language Model / Large Multimodal Model | A model that jointly processes language and one or more non-text modalities; the exact input and output modalities must still be stated. |
| ViT | Vision Transformer | Divides an image into patches, embeds them as a sequence, and applies transformer attention to produce visual representations. |
| visual token | — | An overloaded term for a continuous visual embedding occupying a sequence position or a discrete image-code identifier; always state which. |
| projector / adaptor | — | Maps a modality encoder's feature width and sometimes token count into the language backbone's embedding space. |
| Q-Former | Querying Transformer | Uses a small learned query set to cross-attend to a larger visual feature grid, producing a fixed-size bridge into a language model. |
| pixel unshuffle / space-to-depth | — | Deterministically moves each local spatial block into the channel dimension, reducing spatial positions without averaging them; it is the downsampling inverse of pixel shuffle. |
| NaViT | Native-resolution Vision Transformer | A variable-resolution approach that packs patch sequences from differently sized images rather than forcing every image to one square resolution. |
| M-RoPE | Multimodal Rotary Position Embedding | Extends rotary position coordinates across modality-appropriate axes such as time, height, and width while retaining a text sequence axis. |
| SigLIP | Sigmoid Loss for Language-Image Pre-training | A contrastively trained vision-language representation model using independent sigmoid pair losses rather than a batch softmax. |
| optical context compression | — | Renders or receives text-rich content as pixels and represents it with fewer continuous visual positions before a language decoder reconstructs or reasons over it; compression is lossy and does not eliminate decoding cost. |
| VQ / VQ-VAE | Vector Quantization / Vector-Quantized Variational Autoencoder | Maps continuous image/audio latents to discrete codebook entries and decodes predicted codes back into media. |
| VAE | Variational Autoencoder | Encodes media into a compact continuous latent distribution and decodes latent samples or generated latents back into media. |
| DiT | Diffusion Transformer | Uses a transformer as the denoising network over noisy image or video latent patches; it is not automatically a language model. |
| rectified flow | — | Learns a continuous velocity field that transports noise toward data, then generates by integrating that field, often in a compressed latent space. |
| CFG | Classifier-Free Guidance | Combines conditional and unconditional predictions during media sampling to strengthen prompt alignment; it usually requires both branches and therefore adds inference work. |
| native multimodal | — | An ambiguous integration claim; specify API, sequence, training, shared-backbone, and media-generation levels instead of treating it as a binary architecture. |
| LSH / MinHash | Locality-Sensitive Hashing / Minimum Hashing | Approximate similarity indexing / Jaccard-similarity estimation, often used in deduplication. |
| PII | Personally Identifiable Information | Data that can identify or link to a person. |
| AST | Abstract Syntax Tree | A tree representation of program syntax. |
| QA / STEM | Question Answering / Science, Technology, Engineering, and Mathematics | A task family / a broad collection of technical subject domains; neither term identifies a reward or evaluation protocol by itself. |
| SQL | Structured Query Language | A language for querying and mutating relational databases; database-agent environments must distinguish read-only queries from state-changing actions. |
Distributed training and serving¶
| Term | Full name | Meaning |
|---|---|---|
| CPU / GPU | Central Processing Unit / Graphics Processing Unit | The general-purpose host processor / the dominant parallel accelerator for modern LLM work. |
| CUDA | Compute Unified Device Architecture | NVIDIA's GPU programming platform and execution model; the name now covers compilers, runtimes, libraries, and kernel APIs. |
| FLOP / FLOPs | Floating-Point Operation / Floating-Point Operations | Arithmetic work; “per second” must be stated explicitly when it denotes a rate. |
| MFU | Model FLOPs Utilization | Achieved model arithmetic divided by theoretical accelerator peak under a stated precision convention. |
| DP / DDP | Data Parallelism / Distributed Data Parallel | Replicates parameters, splits examples, and synchronizes gradients. |
| TP / PP | Tensor Parallelism / Pipeline Parallelism | Shards within layers / places consecutive layer groups on different devices. |
| EP / SP / CP | Expert / Sequence / Context Parallelism | Shards experts / sequence work / long-context attention across devices. |
| FSDP | Fully Sharded Data Parallel | Shards parameters, gradients, and optimizer state across data-parallel workers. |
| ZeRO | Zero Redundancy Optimizer | Removes replicated optimizer state, gradients, and then parameters in stages. |
| RDMA | Remote Direct Memory Access | Moves data between registered memory on machines with little CPU involvement. |
| RPC / HTTP / API | Remote Procedure Call / Hypertext Transfer Protocol / Application Programming Interface | Common ways to invoke a remote component or exposed service. |
| TITO | Token-In, Token-Out | GLM-5's asynchronous design that exchanges exact tokens and behavior log-probabilities to reduce rollout/training mismatch. |
| SRS | Streaming Rollout System | ByteDance Seed's system for continuing incomplete long generations while training only the newly generated segment. |
| TAD | Training-Agent Disaggregation | Agent Lightning's separation of arbitrary agent execution/tracing from the learner and optimization runtime. |
| SPMD | Single Program, Multiple Data | Every worker runs the same program over a different shard; distributed model operations synchronize at declared collectives. |
| veRL | Volcano Engine Reinforcement Learning | A distributed post-training framework originating in HybridFlow; it coordinates rollout generation, reward/reference/value inference, and policy updates under configurable resource placement. |
| TRL | Transformer Reinforcement Learning | Hugging Face's post-training library/project for supervised, preference, and RL workflows; a TRL configuration is an implementation artifact, not an algorithm. |
| WSD | Warmup-Stable-Decay | A learning-rate schedule that warms to a peak, holds a long stable phase, then decays near the end of a planned token budget. |
| TTFT / ITL / QPS | Time to First Token / Inter-Token Latency / Queries per Second | First-token delay / decode spacing / request throughput. |
| 1F1B | One Forward, One Backward | A pipeline schedule alternating forward and backward microbatches after warmup. |
| Mooncake | — | Moonshot's KV-cache-centric serving architecture, separating prefill and decode pools and scheduling around reusable cache placement, load, and latency objectives. |
| WAL | Write-Ahead Log | A durable record written before mutation for replay and recovery. |
Precision and hardware¶
| Term | Full name | Meaning |
|---|---|---|
| FP32 / FP16 / FP8 / FP4 | 32-/16-/8-/4-bit floating point | Bit width alone does not specify exponent layout, mantissa layout, or scaling. |
| BF16 / TF32 | Brain Floating Point 16 / TensorFloat-32 | A 16-bit format with FP32-like exponent range / NVIDIA's reduced-mantissa matrix format. |
| E4M3 / E5M2 | 4 exponent + 3 mantissa bits / 5 exponent + 2 mantissa bits | Two FP8 range/precision trade-offs. |
| MXFP4 / MXFP8 | Microscaling 4-/8-bit floating point | Low-precision values sharing scale metadata in small blocks. |
| NVFP4 / E2M1 | NVIDIA 4-bit floating point / 1 sign bit, 2 exponent bits, and 1 mantissa bit | NVIDIA's block-scaled FP4 training/inference format and its 4-bit element encoding; shared scaling metadata is part of the numerical format. |
| A100, A800, H100, H800 | NVIDIA product names, not acronyms | Accelerator variants with distinct compute, memory, and interconnect limits. |
| NVLink / NVSwitch | NVIDIA device interconnect / its switch fabric | High-bandwidth GPU links / an all-to-all fabric connecting them. |
| RoCE | RDMA over Converged Ethernet | Remote Direct Memory Access transported over Ethernet. |
Agents and software¶
| Term | Full name | Meaning |
|---|---|---|
| agent scaffold | — | Runtime prompt construction, tool schema, context management, retries, stopping, and persistence around the model. |
| agent harness | — | The executable loop and policy substrate that connects a model to context, tools, state, and host effects. “Harness” often emphasizes a reusable or modifiable runtime rather than a sealed end-user product. |
| context compaction | — | Replaces older interaction history with a shorter generated or structured state so an agent can continue within a bounded context window. It is lossy unless the original session remains recoverable. |
| context reset / structured handoff | — | Starts a fresh model context while carrying forward an explicit task-state artifact, such as active requirements, completed work, evidence, failures, and next steps. It trades continuity for a clean attention state. |
| ACP | Agent Client Protocol | A protocol boundary through which an editor or other client can start, observe, and steer an agent implemented by a separate process. It is distinct from MCP, which primarily exposes tools and context to the agent. |
| MCP | Model Context Protocol | An open protocol exposing tools and contextual resources to AI applications. |
| provider abstraction | — | A runtime interface that normalizes model endpoints and streaming responses. A common interface does not erase vendor differences in reasoning fields, tool calls, caching, limits, or error behavior. |
| project trust | — | A decision about whether project-local configuration, instructions, plugins, or executable extensions may load. It is an input-loading boundary, not automatically an operating-system sandbox. |
| GUI / CLI | Graphical User Interface / Command-Line Interface | Visual point-and-type interaction / textual command interaction. |
| sandbox | — | Isolation limiting filesystem, network, process, or credential access. |
| prompt injection | — | Untrusted content that attempts to redirect an agent away from authorized instructions. |
| RAG | Retrieval-Augmented Generation | Retrieves external passages or records and conditions generation on them; one-shot retrieval is not automatically an interactive agent environment. |
| reward hacking | — | Raising the measured reward through unintended behavior rather than the desired outcome. |
| policy lag | — | Difference between the rollout-generating policy and the newer learner policy. |
| JSON / XML | JavaScript Object Notation / Extensible Markup Language | Two structured text formats. |
| UUID / SHA-256 | Universally Unique Identifier / Secure Hash Algorithm 256-bit | Decentralized identifiers / cryptographic artifact digests. |
| RNG | Random Number Generator | Produces pseudorandom values for initialization, shuffling, and sampling; reproducibility needs the algorithm, seed, device, and counter/state, not a seed alone. |
| FAIL_TO_PASS / PASS_TO_PASS | failure-to-passing / passing-to-passing tests | Tests fixed by a patch / already-passing tests that detect regressions. |
| SWE | Software Engineering | The domain named in benchmarks such as SWE-bench. |
First-use example¶
Bad:
We train with GRPO in a POMDP and use OPD afterward.
Good:
We train with Group Relative Policy Optimization (GRPO), a critic-free method that compares responses sampled for the same prompt. The interaction is a Partially Observable Markov Decision Process (POMDP) because the agent observes tool results rather than the complete environment state. We then apply On-Policy Distillation (OPD) on states sampled by the student.