Literature review · 718 term(s)
Glossary
Every entry has two definitions. The first is in plain language — the one we'd give a curious non-specialist. The second, in lighter type beneath, is the same idea stated the way you'd hear it in a paper or a conversation between researchers. The popovers you tap on transcripts and review pages show both lines stacked the same way.
A
- A-Mem
- ablation
- accessibility tree
- activation patching
- activation steering
- Adamczyk and Bailey
- adapter
- ADAS
- AddressSanitizer
- Admitted
- ADR
- AEvo
- AFlow
- agent
- Agent A
- Agent Flayer
- agent harness
- Agent JIT
- agent meltdown
- AgentBench
- AgentDojo
- agentic misalignment
- AIMD
- AIME
- AIRA-Compose
- AIRA-Design
- AIRA-dojo
- AIRAformer
- Aletheia
- ALFWorld
- Algorithmic Lovelace Bound
- alignment collapse
- alignment faking
- alignment tax
- alignment training
- alpha
- AlphaEvolve
- AlphaGo
- AlphaProof
- ambient persuasion
- AMC
- amortized inference
- AndroidWorld
- AnswerBench
- API
- Apollo
- append-only memory
- approximate nearest neighbor
- AQuA
- AR1
- ARC-AGI
- ARC-AGI-2
- argmax
- Argus
- associative recall
- associative scan
- asyncio
- ATHENA
- attention
- attention head
- attractor basin
- Attractor Model
- audience design
- AUROC
- Autellix
- Auto-Dreamer
- autoformalization
- AutoGen
- AutoRater
- autoregressive
- Autoresearch
- AWM
- AWS Bedrock
B
- BABILong
- back-chaining
- backbone
- backdoor
- BAR
- basin shift
- Bayesian network
- Bayesian optimization
- BBH
- belief-flow
- benchmark contamination
- Bicameral Model
- BIG-Bench Hard
- BigCodeBench
- bilinear
- binary reward
- binding
- bits-per-byte
- Bitter Lesson
- black box
- Boundless DAS
- branch protection
- Brier score
- BrowseComp
- Browser-Use
- BS-Bench
C
- capability
- capability buffer
- Capability Paradox
- captcha
- Cauchy-Schwarz
- causal consistency
- CCS
- certificate
- CFAA
- CFG
- CFG stride trick
- chain of custody
- chain of thought
- chain rule
- chain-of-thought faithfulness
- chain-of-thought monitoring
- Chapar
- Cholesky factorization
- chwen
- Cialdini's principles
- circle packing
- Claude
- Claude Code
- CLI
- Cline
- closed loop
- closure chain
- CLS theory
- CodeQL
- Codex
- cognitive load
- Cohen's kappa
- COM
- commitment
- commitment failure
- commitment sharpening
- compaction
- Complementary Learning Systems
- Complex-B field
- compliance gap
- COMRace
- concept-aware decoding
- concurrency
- Conformer
- context fatigue
- context rot
- context window
- contextual integrity
- Continuum
- contractive
- contrastive reward
- controllability
- copy-on-write
- copy-up
- counterfactual
- counterfactual utility
- CPU
- CPU offloading
- credit assignment
- credit-assignment SFT
- CrewAI
- critical switching step
- CRIU
- CRPS
- crystal lattice
- CUA-World
- CUDA-graph capture
- CUPMEM
- Cursor
- CVE
- CWM
- Cypher
D
- DAIRA
- DAPO
- data processing inequality
- data-flow
- Daytona
- DeepMind
- DeepSeek
- DeepSeekMath
- DeepSpeed
- deliberative alignment
- DeltaBox
- DeltaCR
- DeltaFS
- denial-of-wallet
- DEQ
- derived Brier
- directional metric
- directive weighting error
- discriminative utility
- distillation
- diversity reward
- Docker
- DOM
- double-free
- DPO
- drift probe
- dwell time
E
- E2B
- ECC
- Echo
- EDR
- EDRM
- eigenvalue
- Elo
- embedding
- emergent capability
- emotion vector
- empathic
- end-to-end learning
- entropy
- entropy phase transition
- epoch
- Epoch Capabilities Index
- equilibrium internalization
- evaluator
- evidence DAG
- evidence graph
- evidence-carrying agent
- EvoLM
- exfiltration
- exfoliation
- exokernel
- experience replay
- exploration hacking
- Explore-then-Act
F
- F1
- FAIR
- faithfulness
- False Compliance Sycophancy
- fast/slow split
- FFmpeg
- fine-tuning
- FineWeb-Edu
- FinOps
- Firecracker
- Firefly
- First Proof
- first-order
- fixed point
- FlashAttention
- Flavell
- Fleiss's kappa
- flyce's kappa
- focal loss
- forced injection
- ForecastBench
- fork
- Format Transfer Injection
- forward pass
- FPO
- Frame Lifetime Trace
- FramePilot
- Freeciv
- frontier model
- FrontierMath
- frontoparietal loop
- FrozenLake
- FSDP
- FunSearch
G
H
- H module
- H100
- H2AC
- halfspace range searching
- hallucination
- halting probe
- hard claim
- harness
- HealthBench
- heavy-tailed
- hedge
- hedge scheduling
- HellaSwag
- Hessian
- Hex-Rays
- hexagonal boron nitride
- hidden state
- HIL-Bench
- history anchors
- HistoryAnchor-100
- homoglyph
- HotPotQA
- HPT
- HRM
- HRM-Text
- hub agent
- hub-agent
- HumanEval
- Humanity's Last Exam
- hybrid model
- hypergeometric
- hyperinflation
I
- I-map
- IDA Pro
- IDE
- illusory truth effect
- IMO
- IMO ProofBench
- implicit differentiation
- implicit function theorem
- implicit prompt injection
- in-context learning
- inclusion-exclusion principle
- indirect prompt injection
- induction head
- Inductive Deductive Synthesis
- Infercept
- inference-masking
- influence function
- instance-level routing
- interferometry
- inverse scaling
- isoFLOP
J
K
L
- L module
- Lakatos
- LAMBADA
- LangGraph
- Latent Evaluator
- latent space
- layer pivot
- Lean
- learning rate schedule
- LHAW
- Life-Harness
- lifecycle taxonomy
- LightMem
- linear probe
- linguistic certainty
- Lipschitz
- Live-SWE-agent
- Llama
- Llama Firewall
- Llama-Factory
- LLM-as-a-judge
- logit lens
- logits
- Long Range Arena
- long-term memory
- LongBench
- Longformer
- LoopTrap
- LoRA
- loss
- Louver
- LUFFY
M
- MagicNorm
- Mamba
- MaR
- Maritaca AI
- MARS
- MAST
- MATH-500
- Mathlib
- Maze-Hard
- MBPP
- MCP
- MCPMark
- MCTS
- mean of means
- mechanistic interpretability
- mediation analysis
- meltdown
- Mem0
- memory cliff
- meta-agent
- Meta-Trace
- metacognition
- metacognitive knowledge
- metacognitive regulation
- Mind2Web
- Minerva
- mini-swe-agent
- MiniMax
- MIPROv2
- MiRA
- MiroMind
- Mistral
- mixture-of-experts
- MLP layer
- MMLU
- Modal
- mode collapse
- model checker
- Model Context Protocol
- model organism
- Model Spec
- monolayer
- Monte Carlo simulation
- Mostly Basic Python Problems
- moving sofa problem
- MQAR
- MSM
- MSRC
- multi-head attention
- multi-level feedback queue
- MultiArith
- multiplicatively independent
- Muon-AdamW
- mutual information
N
O
P
- P-mass
- P1-30B-A3B
- P99
- page heap
- Parallel-Distill-Refine
- Parcae
- Pareto frontier
- pass at k
- pass-cubed
- path patching
- PCA
- PDB
- PDDL
- PDR
- PEAP
- Pearl's ladder
- peer-preservation
- perceptual hashing
- Performer
- perplexity
- phantom gradients
- phase transition
- Phi-4
- Philosophy Spec
- PIML
- PINN
- Plato's Cave
- Platonic Representation Hypothesis
- PlusCal
- policy gradient
- power-seeking
- PPO
- PR
- precondition
- predicted output
- Prefix-RFT
- PrefixLM
- premature exploitation
- premise resistance
- pretraining
- Progent
- progressive growth
- prompt injection
- Proof Commandment Module
- propensity
- propose-and-amplify
- proposer-verifier
Q
R
- RAG
- Ralph loop
- rank-1 approximation
- RAO
- ReAct
- reasoning model
- ReasonMaxxer
- recurrent
- recurrent depth
- RedSearcher
- Reflexion
- reflink
- region rewriting
- REINFORCE
- reinforcement learning with verifiable rewards
- rejected-edit buffer
- ReLIFT
- REPL
- reservoir computing
- residual
- residual connection
- residual stream
- REST API
- reverse-perplexity curriculum
- reward hackability
- reward hacking
- reward model
- RewardBench
- ridge regression
- RLHF
- RLVR
- RMA
- Rocq
- rollout
- rotary embeddings
- RTV
- rubric
- RULER
S
- SAILOR
- sandbox
- sandbox checkpoint
- SAPLMA
- satisfiability solver
- Scale-SWE
- ScienceWorld
- SDF
- SDK
- Self-Consistency
- Self-Refine
- semantic attack
- semantic collapse
- semantic entropy
- semantic gap
- semantic hijacking
- Semgrep
- separation of powers
- SFT
- SGLang
- SGO
- SHADE-Arena
- shadow IT
- shadow pass
- Shannon's chain rule
- Sheeran example
- shortcut penalty
- Show-o2
- side information
- sigma
- sigmoid
- SimpleRL-Zoo
- single-rater
- SKA
- SkillOpt
- slow update
- slyp
- SNR-aware filtering
- SOC
- Sokoban
- Sourcegraph
- sparse autoencoder
- Spearman correlation
- specializable-generalist
- speculative decoding
- SQL injection
- SRFT
- SSH
- SST
- Stackelberg game
- STALE benchmark
- State Stream Transformer
- state-space model
- steering vector
- step amplification factor
- SU-01
- subgoal-driven framework
- sudo
- Sudoku-Extreme
- sufficient statistic
- sunk cost
- Sutton
- SWE-bench
- SWIFT
- sycophancy
- symbolic execution
- synthetic document fine-tuning
- system prompt
T
- tape exfoliation
- tape rug pull
- Task Formatter
- tau-bench
- Tau2-Bench
- taxonomy of failures
- teacher forcing
- teacher-forcing replay
- temperature
- template collapse
- temporal contrast
- TensorRT-LLM
- Terminal-Bench
- termination poisoning
- test-time scaling
- TextCraft
- TheAgentCompany
- thief test
- TLA+
- token
- token compatibility
- Token Signature
- Tongyi DeepResearch
- tool call
- tool compatibility graph
- tool rug pull
- ToolCUA
- top-k
- topology monitor
- TraceFix
- TracIn
- trajectory
- trajectory regulation
- transcoder
- transformer
- TransformerLens
- transmission matrix
- Trixi.jl
- TRL
- TRM
- truncated backpropagation
- TruthfulQA
- tuned lens
- two-dimensional material