Staged-Competence in AI Research

• Staged-Competence is a framework that structures competence acquisition into discrete, progressive stages, used in curriculum design, model growth, and alignment.
• It employs stage-specific mechanisms such as difficulty-based sampling, progressive updates, and phase-separated operations to boost efficiency and safety.
• The approach unifies diverse methods—from transformer pretraining to human-AI co-reasoning—yielding measurable performance improvements and enhanced control.

Searching arXiv for papers directly relevant to “Staged-Competence” and adjacent formulations used in recent literature. Staged-competence denotes a family of formulations in which competence is developed, assessed, or operationalized through an ordered sequence of stages rather than treated as a monolithic property. In recent work, the term appears explicitly as a curriculum framework for safety alignment, and it also functions as an organizing interpretation for staged transformer growth, difficulty-aware reinforcement learning, scaffolded evaluation, deployment-time autonomy refinement, and assessable human–AI co-reasoning (Kumar et al., 25 May 2026, Shen et al., 2022, Ji et al., 1 Apr 2025, An et al., 20 Apr 2026, Basich et al., 2020, Apartsin et al., 4 Jun 2026). Across these settings, the common structure is progressive exposure, representation change, or control transfer such that later stages build on earlier ones rather than discarding them. This suggests that staged-competence is best understood not as a single doctrine but as a family of stage-structured mechanisms for progressive capability acquisition, diagnosis, or control.

1. Conceptual scope and main formulations

The explicit formulation named Staged-Competence is a curriculum learning framework for DPO-based safety alignment that combines difficulty-ordered preference data, competence-based sampling within stages, and progressive reference-model updates between stages (Kumar et al., 25 May 2026). Closely related but distinct formulations appear elsewhere. In transformer pretraining, staged training begins with a small model and incrementally increases the amount of compute used for training by applying a growth operator to increase the model depth and width, with the aim of preserving both loss and training dynamics across stages (Shen et al., 2022). In reliable autonomy, competence improves “over the course of its deployment” through online model updates that increase the granularity of the state representation (Basich et al., 2020). In AI-assisted reasoning, CoRe-3 decomposes productive AI use into three assessable skills—Framing, Judging, and Steering—with “the Judge→Steer cycle” described as “a monitor→control loop seeded by a task definition” (Apartsin et al., 4 Jun 2026). In benchmark diagnosis, STaD defines competence through cumulative scaffolding and the minimum support level required for success (An et al., 20 Apr 2026).

Setting	Staged unit	Representative formulation
Safety alignment	Difficulty buckets and competence schedule	Difficulty-ordered preference data, competence-based sampling, reference updates
Transformer pretraining	Growth stages in model size	Loss-preserving and training-dynamics-preserving growth operators
AI-assisted reasoning assessment	Separable skills	Framing, Judging, Steering
Benchmark diagnosis	Cumulative scaffolds	Minimum scaffolding level $k$
Autonomous systems	Iterative deployment-time refinement	Indiscriminate states, discriminators, representation update

These uses differ in what is staged: model scale, data difficulty, autonomy level, human–AI reasoning skills, or external assistance. A plausible implication is that staged-competence is a structural idea whose content depends on whether competence is treated as optimization state, policy robustness, diagnosable skill, or safe autonomy choice.

2. Architecture growth and phase-separated acquisition

In autoregressive language modeling, staged training formalizes competence growth as stagewise architectural expansion over the full training state. The training state is defined as

$\mathcal{T} = \{\theta, (m,v), \lambda(t)\},$

where $\theta$ are model parameters, $(m,v)$ are Adam first and second moments, and $\lambda(t)$ is the learning-rate schedule. A growth operator is a mapping on the entire state,

$$\mathbb{G}(\mathcal{T}_{\text{orig}}) = \mathcal{T}_{\text{grow}},$$

and its two key properties are loss preservation and preservation of training dynamics (Shen et al., 2022). Loss preservation requires that immediately after growth the enlarged model represent the same function as the smaller one. Training-dynamics preservation requires that the compute efficiency,

$\frac{\partial}{\partial C}\mathcal{L}(\cdot, C),$

match that of a same-size target model trained from scratch at the same loss level. The paper argues that preserving loss alone is insufficient because a grown model may begin at the right loss and still optimize too slowly or unstably. Width growth duplicates embeddings, layernorm parameters, biases, and feed-forward structure, with the last feed-forward layer divided by $2$ to preserve logits; depth growth inserts identity-initialized residual blocks by zeroing layernorm scale and bias and zeroing linear biases. Learning continues by transforming optimizer state consistently and shifting the effective step count so that the learning-rate schedule resumes at the matched larger-model point. On GPT-2 style autoregressive LLMs trained on C4, the framework reports up to $22\%$ compute savings in the abstract and larger savings in some tables, with the largest gains appearing at or before the point of optimality (Shen et al., 2022).

A different architectural instance appears in Phasor Agents, where competence is distributed across distinct dynamical phases rather than across increasing transformer size. The learning architecture separates wake tagging, offline consolidation, deep-sleep-like gated capture, and REM-like replay. Wake updates use fast eligibility traces, while deep-sleep-like capture commits slow traces under spindle-gated windows and REM-like replay reconstructs and perturbs stored experience for planning. The staged evidence is explicitly mechanistic: wake/sleep separation expands stable learning by $67$ percent under matched weight-norm budgets, REM replay improves maze success rate by $\mathcal{T} = \{\theta, (m,v), \lambda(t)\},$0 percentage points, and a Tolman-style latent-learning signature appears at reward onset after unrewarded exploration (Trappe, 7 Jan 2026). This suggests that staged-competence can also denote phase-separated acquisition in which stability, retention, and planning are delegated to different learning regimes rather than to a single online update process.

3. Curriculum, post-training, and alignment

Difficulty-aware reinforcement learning makes staged-competence explicit at the data-curriculum level. One study partitions math and code data into three difficulty levels using empirical pass rates of DeepSeek-R1-Distill-Qwen-1.5B, 7B, 32B, and DeepSeek-R1. The scoring functions are

$\mathcal{T} = \{\theta, (m,v), \lambda(t)\},$1

and

$\mathcal{T} = \{\theta, (m,v), \lambda(t)\},$2

The staged schedule used in the main experiment is sequential: Stage 1 trains on difficulty level 2, and Stage 2 switches to difficulty level 3 once performance plateaus, with the reported run switching at step $\mathcal{T} = \{\theta, (m,v), \lambda(t)\},$3. Stage 2 also increases maximum sequence length from $\mathcal{T} = \{\theta, (m,v), \lambda(t)\},$4k to $\mathcal{T} = \{\theta, (m,v), \lambda(t)\},$5k, excludes truncated samples from loss computation, and removes entropy loss. The paper reports that the approach enables a 1.5B parameter model to achieve $\mathcal{T} = \{\theta, (m,v), \lambda(t)\},$6 on AIME-2024 and $\mathcal{T} = \{\theta, (m,v), \lambda(t)\},$7 on MATH-500, and that switching to the second stage yields sustained performance improvements relative to continuous training in the first stage (Ji et al., 1 Apr 2025).

The safety-alignment framework explicitly named Staged-Competence begins from a preference dataset

$\mathcal{T} = \{\theta, (m,v), \lambda(t)\},$8

and assigns each pair a preference alignment margin

$\mathcal{T} = \{\theta, (m,v), \lambda(t)\},$9

where $\theta$0 is the base model’s zero-shot response. The sorted dataset is divided into $\theta$1 equal buckets, and training within each bucket uses a square-root competence schedule

$\theta$2

with eligible pool

$\theta$3

After each stage, the DPO reference model is updated by

$\theta$4

Across three model families, the paper reports average reductions of $\theta$5 in OOD harmful response rates and $\theta$6 in jailbreak attack success rates, while preserving general capabilities with near-zero over-refusal; it also reports that the method matches baseline safety with only $\theta$7 of the training data (Kumar et al., 25 May 2026).

Two-stage post-training for vision-LLMs provides a narrower counterpoint. In that setting, Stage-1 warm-starts—SFT or OPD—lead to a narrow $\theta$8–$\theta$9 Geometry3K validation band after RL, and the paper’s main conclusion is that Stage-1 is “strongly associated with the entropy regime in this setup,” not with materially different final in-domain outcomes. OPD enters RL with policy entropy $(m,v)$0–$(m,v)$1 early and $(m,v)$2–$(m,v)$3 later, compared with $(m,v)$4–$(m,v)$5 for SFT, and shows initialization pass@16 gains of $(m,v)$6 to $(m,v)$7 over SFT that disappear after RL (Shen, 8 Jun 2026). This indicates that some staged procedures alter exploration, uncertainty, and diversity properties more strongly than final benchmark endpoints.

4. Assessment, scaffolding, and diagnosable stages

In AI-assisted reasoning, staged-competence is formulated as an assessable competency model rather than as a training curriculum. CoRe-3 decomposes productive AI use into Framing, Judging, and Steering. Framing is “a problem-structuring skill exercised prior to generation,” Judging is “an evaluative, epistemic skill,” and Steering is “a control skill exercised after generation, in a loop with Judging” (Apartsin et al., 4 Jun 2026). The learner workflow is explicitly phased: first an ill-defined problem is framed; then the system produces a plausible but deliberately flawed solution; then Judge/Steer cycles may repeat across multiple rounds. In simulated learners, the skills dissociate: own-effects on the 3-point grade scale are $(m,v)$8 for Framing, $(m,v)$9 for Judging, and $\lambda(t)$0 for Steering, with average off-diagonal effect $\lambda(t)$1. Inter-skill grade correlations are $\lambda(t)$2 for Framing–Judging, $\lambda(t)$3 for Framing–Steering, and $\lambda(t)$4 for Judging–Steering (Apartsin et al., 4 Jun 2026). These results operationalize staged-competence as separable but partially dependent skill phases.

STaD instead treats competence as staged recoverability under external support. An original question is decomposed as

$\lambda(t)$5

with intermediate answers

$\lambda(t)$6

and scaffolded variants

$\lambda(t)$7

The key diagnostic quantity is the minimum scaffolding level

$\lambda(t)$8

with $\lambda(t)$9 if no scaffold works. This induces three competence categories: $$\mathbb{G}(\mathcal{T}_{\text{orig}}) = \mathcal{T}_{\text{grow}},$$0 independently solvable, $$\mathbb{G}(\mathcal{T}_{\text{orig}}) = \mathcal{T}_{\text{grow}},$$1 solvable only with scaffolding, and $$\mathbb{G}(\mathcal{T}_{\text{orig}}) = \mathcal{T}_{\text{grow}},$$2 unsolvable even with full scaffolding (An et al., 20 Apr 2026). Across ToT Arithmetic, GSM8K, and Math-Hard, scaffolding reveals model-specific bottlenecks not visible in aggregate accuracy, and a leakage-control ablation replacing injected intermediate values with placeholders reduces accuracy from $$\mathbb{G}(\mathcal{T}_{\text{orig}}) = \mathcal{T}_{\text{grow}},$$3 on chosen scaffold-success cases to about $$\mathbb{G}(\mathcal{T}_{\text{orig}}) = \mathcal{T}_{\text{grow}},$$4 on average. This suggests that staged-competence can be measured either as separable skill phases or as the minimum amount of structured assistance required for success.

5. Competence refinement during deployment

In competence-aware autonomy, competence is defined not as generic ability but as the optimal autonomy level for executing action $$\mathbb{G}(\mathcal{T}_{\text{orig}}) = \mathcal{T}_{\text{grow}},$$5 in state $$\mathbb{G}(\mathcal{T}_{\text{orig}}) = \mathcal{T}_{\text{grow}},$$6 given the human’s true feedback model: $$\mathbb{G}(\mathcal{T}_{\text{orig}}) = \mathcal{T}_{\text{grow}},$$7 A competence-aware system is an augmented stochastic shortest path model with factored states and actions, $$\mathbb{G}(\mathcal{T}_{\text{orig}}) = \mathcal{T}_{\text{grow}},$$8, together with a domain model, autonomy model, and human feedback model (Basich et al., 2020). The staged aspect is iterative and deployment-time. The system begins with a partial active feature space $$\mathbb{G}(\mathcal{T}_{\text{orig}}) = \mathcal{T}_{\text{grow}},$$9, detects indiscriminate states where no feedback signal has sufficiently high predicted probability,

$\frac{\partial}{\partial C}\mathcal{L}(\cdot, C),$0

searches the inactive feature space $\frac{\partial}{\partial C}\mathcal{L}(\cdot, C),$1 for a discriminator $\frac{\partial}{\partial C}\mathcal{L}(\cdot, C),$2 that improves feedback prediction by at least $\frac{\partial}{\partial C}\mathcal{L}(\cdot, C),$3, retrains the feedback predictor on $\frac{\partial}{\partial C}\mathcal{L}(\cdot, C),$4, selects

$\frac{\partial}{\partial C}\mathcal{L}(\cdot, C),$5

and updates the active representation by

$\frac{\partial}{\partial C}\mathcal{L}(\cdot, C),$6

The process repeats over deployment, and the paper reports that the modified CAS reached nearly $\frac{\partial}{\partial C}\mathcal{L}(\cdot, C),$7 level-optimality in the random task experiment across all states and all visited states, whereas the standard CAS improved only slightly and in the fixed-task domain was unable to reach even $\frac{\partial}{\partial C}\mathcal{L}(\cdot, C),$8 level-optimality across all visited states (Basich et al., 2020). It also used significantly fewer feedback signals over time, and the expected cost converges to be almost identical to the incurred cost. Here staged-competence is neither fixed curriculum nor architectural growth; it is successive refinement of representation and autonomy choice under human feedback.

6. Theoretical decompositions, limits, and open questions

A formal theory of the “need for competence” in computational intrinsic motivation decomposes competence into four facets: effectance (C1), skill use (C2), task performance (C3), and capacity growth (C4) (Lintunen et al., 11 Feb 2025). Candidate formalisms differ by facet. For effectance, RIDE uses

$\frac{\partial}{\partial C}\mathcal{L}(\cdot, C),$9

For skill use, VIC uses

$2$0

and DIAYN uses

$2$1

For intended task performance, RIG uses

$2$2

For challenge-calibrated performance and growth, CURIOUS selects modules by

$2$3

The paper’s broader point is that computational modeling exposes preconditions that SDT leaves implicit, including causal attribution, skill individuation, goal representation, progress estimation, novelty, diversity, and challenge sensitivity (Lintunen et al., 11 Feb 2025). This suggests that staged-competence should not automatically be reduced to one scalar maturity axis.

The literature also places clear limits on stage-based claims. Staged transformer growth depends on carefully engineered architecture-specific growth operators, optimizer-state transfer, and learning-rate schedule transfer, with zero-shot transfer after width growth showing transient degradation (Shen et al., 2022). Difficulty-aware staged RL relies on manually engineered stage boundaries based on reference-model pass rates, and its second stage changes more than just data difficulty by also increasing rollout length and removing entropy loss (Ji et al., 1 Apr 2025). Safety-alignment Staged-Competence uses a particular difficulty score and $2$4 equal buckets, and its evidence is concentrated on roughly 8B-scale LoRA fine-tuning (Kumar et al., 25 May 2026). CoRe-3’s empirical validation is currently based on simulated learners, with human-rater agreement identified as the next step (Apartsin et al., 4 Jun 2026). STaD remains a black-box diagnostic method and cannot prove internal causal mechanisms or whether scaffolded recovery reflects latent skill rather than opportunistic use of intermediate information (An et al., 20 Apr 2026). Two-stage VLM post-training further shows that stage structure may control entropy regime, answer diversity, and pass@16 at initialization without materially changing the final in-domain endpoint (Shen, 8 Jun 2026).

Taken together, these results support a restrained interpretation. Staged-competence is a technically diverse research program in which stages can refer to architecture growth, curriculum difficulty, reference-policy progression, assessable reasoning skills, scaffold levels, or deployment-time representation refinement. What unifies these lines is not a single metric of competence, but a shared claim that competence can be better acquired, measured, or controlled when optimization, evaluation, or autonomy is organized into explicit stages rather than left monolithic.