L

Late Fusion

• Category: Multimodal systems
• What it actually means: Processing modalities through separate encoders or predictors and combining their high-level representations, scores, or decisions near the task output.
• Why it matters: Separate branches can use modality-specific architectures and tolerate missing inputs, though they may miss fine-grained interactions available to earlier fusion.
• In practice: Calibrate each branch, define how missing modalities affect the merge, compare score-level and feature-level combinations, and evaluate each branch alone as an ablation.
• Common confusion: Late fusion describes the position of combination. It does not mean simple averaging or guarantee that the modalities contribute equally.
• Learn it: Cross-Attention Fusion
• Related terms: Early Fusion, Multimodal Fusion, Modality, Evaluation (Eval)
• Sources: Multimodal Deep Learning; Multimodal Machine Learning: A Survey and Taxonomy

Latent Space

• Category: Data & representations
• What people say: A model's hidden representation space.
• What it actually means: A learned representation space whose coordinates encode factors useful to a model. It may be lower-dimensional than the input, but compression is not required for every latent representation.
• Common confusion: Nearby points are only meaningfully similar according to what the model and training objective learned.
• Related terms: Embedding, VAE (Variational Autoencoder), Feature

Learning Rate

• Category: Math & training
• What people say: How large each optimization step is.
• What it actually means: A scale factor used by an optimizer to control parameter-update magnitude. Values that are too large can destabilize training; values that are too small can make useful progress impractically slow.
• Common confusion: The effective update also depends on the optimizer, schedule, gradient scale, batch, and parameter history.
• Related terms: Optimizer, Gradient Descent, Batch Size

Learning Rate Schedule

• Category: Math & training
• What it actually means: A policy that changes the optimizer's learning rate as training progresses according to steps, epochs, metrics, or a predefined curve.
• Why it matters: Different training stages can benefit from different update scales, so one constant rate may be unstable early or wasteful late.
• In practice: Version the schedule with the optimizer configuration, log the actual rate at every step, and compare schedules under the same token or update budget.
• Common confusion: A scheduler controls the learning rate over time; it does not decide when an optimizer step occurs or guarantee convergence.
• Related terms: Learning Rate, Warmup, Optimizer, Epoch
• Sources: SGDR; Attention Is All You Need

Least Privilege

• Category: Evaluation & safety
• What it actually means: Giving a model, agent, tool, or user only the permissions required for the current task, for only as long as those permissions are needed.
• Why it matters: Models can make mistakes or follow malicious instructions. Narrow permissions reduce the damage any one failure can cause.
• In practice: Give a documentation agent read access to source files and write access to one branch, but no production credentials or merge permission.
• Common confusion: Authentication proves identity. Least privilege limits what that identity can do.
• Related terms: Sandbox, Approval Gate, Prompt Injection, Tool Contract

LLM (Large Language Model)

• Category: Models & inference
• What people say: The brain of an AI application.
• What it actually means: A language model with enough capacity and broad training to perform many language tasks through prompting or adaptation. Most current LLMs use transformer architectures and sequence-prediction objectives, but size thresholds, data sources, and training recipes vary.
• Common confusion: An LLM is a model component. Tools, retrieval, state, policies, and product logic live in the surrounding system.
• Related terms: Transformer, Autoregressive, Agent Harness

LLM-as-a-Judge

• Category: Evaluation & safety
• What it actually means: Using a language model to score, compare, classify, or critique another system's output against a rubric.
• Why it matters: It can scale evaluation of qualities that are difficult to express as exact-match tests, such as clarity or instruction adherence.
• In practice: Give a separate evaluator model the task, candidate answer, reference evidence, and a structured rubric, then calibrate its scores against human-reviewed examples.
• Common confusion: A judge model is not ground truth. It can be biased by order, verbosity, style, prompt wording, or shared model failures.
• Learn it: Eval-Driven Agent Development
• Related terms: Evaluation (Eval), Eval Set, Verification Gate, Precision & Recall

Load Shedding

• Category: Reliability & operations
• What it actually means: Deliberately rejecting, dropping, or cancelling selected work at one or more overload boundaries when demand exceeds the capacity available to produce useful results.
• Why it matters: Continuing to accept every request during overload can increase queueing until nearly all requests miss their deadlines and recovery becomes harder.
• In practice: Shed at the earliest informed boundary, preserve high-priority and already-admitted work when possible, identify the overloaded scope, and mark a response retryable only when the condition is transient and the request remains within its retry budget.
• Common confusion: Load shedding is not confined to work that has already been accepted. Admission control is specifically the pre-acceptance gate, while rate limiting can enforce a usage policy even when capacity remains.
• Related terms: Admission Control, Backpressure, Rate Limit, Graceful Degradation
• Sources: Google SRE: Handling Overload

Logits

• Category: Models & inference
• What it actually means: The model's unnormalized numeric scores for candidate outcomes before a normalization function or decoding rule converts them into selections.
• Why it matters: Temperature, softmax, top-k, and top-p operate on or derive from logits, so logits connect model computation to generated tokens.
• In practice: Inspect logits or log probabilities when the API exposes them, apply masks before sampling, and avoid interpreting raw magnitude as calibrated confidence.
• Common confusion: Logits are not probabilities and are not comparable across unrelated positions, models, or tasks without a defined transformation.
• Related terms: Softmax, Temperature, Token, Cross-Entropy
• Sources: Attention Is All You Need

LoRA (Low-Rank Adaptation)

• Category: Math & training
• What people say: Parameter-efficient fine-tuning.
• What it actually means: A method that keeps base weights frozen and learns low-rank update matrices for selected layers. It reduces the number of trainable parameters and can lower training memory relative to full-parameter fine-tuning.
• Common confusion: Actual memory and speed savings depend on rank, target modules, optimizer state, activation memory, quantization, and implementation.
• Learn it: Fine-Tuning and LoRA
• Sources: LoRA paper
• Related terms: Fine-tuning, QLoRA, Parameter

Loss Function

• Category: Math & training
• What people say: A number that measures training error.
• What it actually means: An objective that maps predictions and targets, sometimes with regularization terms, to a value optimization tries to reduce. The loss determines which errors training directly rewards or penalizes.
• Common confusion: A low training loss does not guarantee useful, safe, or generalizable behavior on production tasks.
• Related terms: Cross-Entropy, Gradient, Evaluation (Eval)

Lost in the Middle

• Category: Prompting & context
• What it actually means: A long-context failure pattern in which model performance changes with evidence position and can degrade when relevant information sits between the beginning and end.
• Why it matters: Fitting evidence inside the context window does not guarantee that the model will use every position with equal reliability.
• In practice: Test several evidence positions, reduce distractors, place decision-critical constraints where they remain salient, and verify answers against the source.
• Common confusion: It is an observed behavior pattern, not a fixed law that affects every model, task, or position identically.
• Related terms: Context Window, Context Engineering, Eval Set, Grounding
• Sources: Lost in the Middle