Teacher forcing
Algorithms AIMech:TeacherForcing DRAFT
A sequence-training procedure that supplies reference outputs from earlier steps as feedback for subsequent predictions, in place of the model outputs that would otherwise be fed back.
Identity Notes & Concept Scope
Compared all 442 pinned AIO and 582 native IDs, names, aliases and scoped synonyms; no normalized identity match. Ignored- and hidden-inclusive fixed-string searches of the recovery, older discovery and original Documents checkouts returned 70, 10, 10 lines. All distinct match contents were inspected. Semantic matching covered 75 AIO and 167 native definitions, with unchanged preceding rows verified and new matches read. 4 full annotated AIO node/edge/generated contexts and 3 full native records were inspected. Matches concern the AIO DifferentiableLLM definition, the RTRL configuration and the pending broader teacher-forcing lead. Teacher forcing specifies the feedback supplied during training, not a particular architecture, gradient algorithm or trained instance. RTRL is one algorithm that can use it. Mean Teacher averages student parameters for consistency targets; DocumentNADE is a particular autoregressive architecture. Neither is this general feedback-substitution procedure. The pinned Distillation compression definition is also distinct. Scheduled sampling changes the feedback-selection rule and is not an alias; the all-reference endpoint coincides with full teacher guidance.
Claims & Grounded Evidence
6 Claims
Mechanism
#reference-feedback
During training, substitute the supplied reference output for the model output on the feedback path used by subsequent computations. In autoregressive token prediction this means predicting y_t given reference tokens y_<t and any conditioning input, rather than feeding the current target y_t as its own predictor input. The reference may be a token or a continuous output; the term does not by itself specify a neural architecture, optimizer or decoding policy. Merely placing a reference in the loss without substituting a feedback input is insufficient to define this procedure.
Mechanism
#observed-prefix-likelihood
For the cited autoregressive sequence model, log p_theta(Y|X)=sum_t log p_theta(y_t|y_<t,X). Evaluating these conditionals on the observed prefix gives the exact observed-sequence log likelihood for that model. Conditioning on reference tokens during likelihood training is therefore not inherently an invalid likelihood objective. This identity does not establish that a trained model generates correct sequences or that every nonlikelihood use of teacher forcing optimizes this objective.
Mechanism
#derivative-paths
When supplied reference values are parameter independent, future computations have no derivative through the model-output values they replace. Unreplaced recurrent hidden-state paths can still carry gradients through time. In the Williams–Zipser RTRL configuration, the current output sensitivity first contributes to its current update, then the forced output is treated as having zero sensitivity for subsequent propagation. Thus teacher forcing does not generally detach every hidden state or remove all temporal credit assignment.
Limitation
#generation-distribution-boundary
At generation time the preceding reference output may be unavailable, so an autoregressive generator uses its own selected outputs, for example through sampling, greedy selection or beam search. Those prefixes can differ from training prefixes and permit errors to compound. This is a conditional distribution-mismatch concern, not a proof that every teacher-forced model fails or that low reference-prefix loss guarantees accurate autonomous rollouts.
Scope
#scheduled-sampling-boundary
The cited scheduled-sampling procedure chooses between the reference previous token and a model-generated token during training using a scheduled probability. Its all-reference endpoint is full teacher guidance; mixed-input training and the all-generated endpoint are distinct feedback rules, not aliases for full teacher forcing. The 2015 experiments do not backpropagate through the sampled token-selection decision. These choices do not redefine the exact observed-prefix likelihood identity for the all-reference baseline.
Scope
#empirical-scope
The 2015 comparisons cover specified image-captioning, parsing and speech setups. Training entirely on generated tokens from the start is not uniformly successful in those experiments; the parsing discussion reports failure to learn correct tree structure in that condition. The speech next-step and free-decoding frame-error results use different prediction conditions, and must not be treated as the same metric or as general word-recognition accuracy. These selected experiments do not establish a universal ordering between teacher forcing, scheduled sampling and autonomous training.
⚠️ Curated Limitations & Negative Results
- The explicit teacher-forcing name is supported by Williams and Zipser1989; Bengio et al2015 supplies the broader true-prefix sequence-training baseline. Jordan1986 was not accessible for inspection, so no origin or priority claim is made.
- The supplied target is assumed independent of model parameters in the stated derivative rule. Learned targets or differentiable target generation need their own dependency analysis.
- The reference-prefix likelihood claim is scoped to the cited autoregressive likelihood formulation; not all continuous-output or task-loss configurations have that likelihood objective.
- Target substitution and gradient truncation are separate choices. Keeping recurrent hidden paths does not guarantee that a particular implementation differentiates every path or that long dependencies are learned.
- Neither good reference-prefix prediction nor optional teacher-forced training establishes matched free-running performance, universal robustness or superiority over other training procedures.
- All ten1989 manuscript pages and nine2015 proceedings pages were read. Constructed algebraic examples were run without fitting; no author implementation, model training or benchmark replication was performed.
Artificial Intelligence Ontology (AIO) Grounding
Mapped AIO Classes
| AIO Target Class | Mapping Relation | Mapping Rationale |
|---|---|---|
AIO:TrainingStrategy
TrainingStrategy
|
related_match | Teacher forcing specifies feedback supplied during model training. The pinned target describes preprocessing and general training strategies rather than this sequential feedback rule. A contextual related match preserves that abstraction difference without asserting exact identity or a subclass edge. |
Evidence Sources & Primary Literature
2 Sources| Source ID | Publication / Document | Type / Access | Reference & Checksum |
|---|---|---|---|
williams1989
|
Ronald J. Williams, David Zipser
|
primary paper
full_text
|
ID: https://doi.org/10.1162/neco.1989.1.2.270
sha256:348ca1d75163…
|
bengio2015
|
Samy Bengio, Oriol Vinyals, Navdeep Jaitly, Noam Shazeer
|
primary paper
full_text
|
ID: https://proceedings.neurips.cc/paper/2015/hash/e995f98d56967d946471af29d7bf99f1-Abstract.html
sha256:ab71fd38a31d…
|
Curation Provenance
Curator: Codex
Method: ai_assisted
Model: GPT-6
Created: 2026-10-13T06:44:59Z
Updated: 2026-10-13T06:48:49Z
Source File:
data/curated/TeacherForcing.yaml