---
title: Staged-Competence in AI Research
url: https://www.emergentmind.com/topics/staged-competence
type: topic
---

# Staged-Competence in AI Research

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 [2605.26315][2203.06211][2504.00829][2604.18177][2007.11740][2606.05983]. 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 [2605.26315]. 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 [2203.06211]. In reliable autonomy, competence improves “over the course of its deployment” through online model updates that increase the granularity of the state representation [2007.11740]. 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” [2606.05983]. In benchmark diagnosis, STaD defines competence through cumulative scaffolding and the minimum support level required for success [2604.18177].

| 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 [2203.06211]. 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 language models 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 [2203.06211].

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 \(+45.5\) percentage points, and a Tolman-style latent-learning signature appears at reward onset after unrewarded exploration [2601.04362]. 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
\[
\text{score\_math} =
\begin{cases}
0, & \text{if model's response format is incorrect}, \\
0, & \text{if model's answer is incorrect}, \\
1, & \text{if model's answer is correct},
\end{cases}
\]
and
\[
\text{score\_code} = \frac{\#\passed\_test\_case}{\#\test\_cases}.
\]
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 \(1600\). Stage 2 also increases maximum sequence length from \(16\)k to \(24\)k, excludes truncated samples from loss computation, and removes entropy loss. The paper reports that the approach enables a 1.5B parameter model to achieve \(42.3\%\) on AIME-2024 and \(89.5\%\) on MATH-500, and that switching to the second stage yields sustained performance improvements relative to continuous training in the first stage [2504.00829].

The safety-alignment framework explicitly named **Staged-Competence** begins from a preference dataset
\[
\mathcal{D} = \{(x_i, y_i^+, y_i^-)\}_{i=1}^N
\]
and assigns each pair a preference alignment margin
\[
m_i = \cos(e_{\hat{y}_i}, e_{y_i^+}) - \cos(e_{\hat{y}_i}, e_{y_i^-}),
\]
where \(\hat{y}_i\) is the base model’s zero-shot response. The sorted dataset is divided into \(K\) equal buckets, and training within each bucket uses a square-root competence schedule
\[
c(t) = \sqrt{\frac{(1-c_0^2)t}{T} + c_0^2},
\]
with eligible pool
\[
P_t = \{ i \in B_k : d_i \le c(t) \}.
\]
After each stage, the DPO reference model is updated by
\[
\pi_{\mathrm{ref}}^{(k+1)} \leftarrow \pi^{(k)}.
\]
Across three model families, the paper reports average reductions of \(16\%\) in OOD harmful response rates and \(20\%\) 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 \(75\%\) of the training data [2605.26315].

Two-stage post-training for vision-language models provides a narrower counterpoint. In that setting, Stage-1 warm-starts—SFT or OPD—lead to a narrow \(53\)–\(54\%\) 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 \(0.10\)–\(0.16\) early and \(0.067\)–\(0.11\) later, compared with \(0.015\)–\(0.04\) for SFT, and shows initialization pass@16 gains of \(+2.0\) to \(+5.2\) over SFT that disappear after RL [2606.09059]. 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” [2606.05983]. 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 \(+0.62\) for Framing, \(+2.00\) for Judging, and \(+0.43\) for Steering, with average off-diagonal effect \(+0.01\). Inter-skill grade correlations are \(\rho=-0.03\) for Framing–Judging, \(\rho=-0.12\) for Framing–Steering, and \(\rho=+0.25\) for Judging–Steering [2606.05983]. 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
\[
q \rightarrow S = \{s_1, s_2, \dots, s_K\},
\]
with intermediate answers
\[
SA = \{sa_1, \dots, sa_K\},
\]
and scaffolded variants
\[
q_{\text{scaf}^{(j)}} = g\!\left(q,\; \{(s_1,sa_1), \dots, (s_j,sa_j)\}\right).
\]
The key diagnostic quantity is the minimum scaffolding level
\[
k = \min \Big\{ j \mid \mathrm{Cons}(\hat{p}^{(j)}, a) = 1 \;\wedge\; \mathrm{Cons}(\hat{p}^{(l)}, a) = 0,\; l < j \Big\},
\]
with \(k=-1\) if no scaffold works. This induces three competence categories: \(k=0\) independently solvable, \(k>0\) solvable only with scaffolding, and \(k=-1\) unsolvable even with full scaffolding [2604.18177]. 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 \(100\%\) on chosen scaffold-success cases to about \(12\%\) 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 \(a\) in state \(s\) given the human’s true feedback model:
\[
\chi_{S}(s, a) = \argmin_{l \in L} Q(s, (a, l) ; \lambda^{H}).
\]
A competence-aware system is an augmented stochastic shortest path model with factored states and actions, \(\langle S, A, T, C, s_0, s_g \rangle\), together with a domain model, autonomy model, and human feedback model [2007.11740]. The staged aspect is iterative and deployment-time. The system begins with a partial active feature space \(\hat F\), detects **indiscriminate** states where no feedback signal has sufficiently high predicted probability,
\[
\exists \, a \in A \;\forall \sigma \in \Sigma: \hat{\lambda}(\sigma | s, a) \leq 1 - \delta,
\]
searches the inactive feature space \(\breve F\) for a discriminator \(D\) that improves feedback prediction by at least \(\alpha\), retrains the feedback predictor on \(\hat F \times D\), selects
\[
D^* = \argmax_D \text{EvaluateClassifier}(\lambda_D, D_{\text{val}}),
\]
and updates the active representation by
\[
\hat{F} \leftarrow \hat{F} \times D^*.
\]
The process repeats over deployment, and the paper reports that the modified CAS reached nearly \(100\%\) 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 \(40\%\) level-optimality across all visited states [2007.11740]. 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)** [2502.07423]. Candidate formalisms differ by facet. For effectance, RIDE uses
\[
R_t(s_t, s_{t+1}) := \frac{\left\lVert\phi(s_{t+1})-\phi(s_t)\right\rVert_2}{\sqrt{N(s_{t+1})}}.
\]
For skill use, VIC uses
\[
R_t(s_0,s_f,\Omega) := \log q_\phi(\Omega \mid s_0,s_f) - \log p_\psi(\Omega \mid s_0),
\]
and DIAYN uses
\[
R_t(s_{t+1}, z) := \log q_\phi(z \mid s_{t+1}).
\]
For intended task performance, RIG uses
\[
R_t(s_{t+1}, g) := -\left\lVert\phi(s_{t+1})-\phi(g)\right\rVert_A.
\]
For challenge-calibrated performance and growth, CURIOUS selects modules by
\[
p(M_i) := \epsilon \times \frac{1}{N} + (1-\epsilon) \times \frac{|LP(M_i)|}{\sum_{j=1}^N |LP(M_j)|}.
\]
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 [2502.07423]. 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 [2203.06211]. 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 [2504.00829]. Safety-alignment Staged-Competence uses a particular difficulty score and \(K=3\) equal buckets, and its evidence is concentrated on roughly 8B-scale LoRA fine-tuning [2605.26315]. CoRe-3’s empirical validation is currently based on simulated learners, with human-rater agreement identified as the next step [2606.05983]. 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 [2604.18177]. 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 [2606.09059].

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.

Source: https://www.emergentmind.com/topics/staged-competence