General+Expert Progressive Curriculum

• General+Expert Progressive Curriculum is a staged training regime that transitions models from broad, redundant cues to specialized, synergistic competence.
• It leverages adaptive difficulty signals and multi-stage scheduling to progressively expose models to increasingly complex tasks.
• Empirical studies show significant improvements in accuracy and robustness across multimodal, reasoning, and segmentation tasks using this framework.

Searching arXiv for the cited curriculum-learning papers to ground the article with current metadata and identifiers. A general+expert progressive curriculum is a training regime in which a model is first exposed to broad, low-noise, or structurally simpler supervision and is then advanced toward specialized, high-interaction, or high-precision cases. Recent work instantiates this pattern in several distinct ways: multimodal systems move from redundant shared cues to unique modality-specific patterns and finally to synergistic fusion; mathematical reasoners move from easy problems to hard ones with adaptive hints; web agents move from imitation to discrimination and then to long-horizon consistency; and dense predictors move from small, foreground-rich patches or blurred images to full-context sharp inputs (Singh et al., 15 Jun 2026, Wu et al., 4 Jun 2025, Peng et al., 14 Apr 2026).

1. Conceptual structure

The term is best understood as a family of staged curricula rather than a single algorithm. In "SPICE" (Singh et al., 15 Jun 2026), the progression is explicitly general → specialized → expert: the general stage prioritizes redundant cross-modal cues shared by all modalities, intermediate stages emphasize unique modality-specific information, and the expert stage focuses on synergistic interactions that arise only from joint multimodal fusion. In "Progressive Mastery" (Wu et al., 4 Jun 2025), the same logic appears as easy/medium/hard partitioning together with Guided Prompting, so the model first solves simpler items and then progressively handles harder items with decreasing external support.

Other systems realize the same pattern along different axes. "ACER" organizes progression simultaneously along a content axis and a cognitive axis: persona levels advance from High school → Undergraduate → Graduate → Researcher, while training content advances from Books → easy QA → hard QA (Neema et al., 30 Oct 2025). "Triton" organizes progression at the optimizer level: Triton-SFT-32B provides basic imitation, Triton-ORPO-32B adds robust discrimination, and Triton-GRPO-32B adds long-horizon consistency (Peng et al., 14 Apr 2026). "ByteLoom" uses a three-stage HOI video curriculum: Human Pose Conditioning, then Hand-Object Interaction, then full Human-Object Interaction finetuning (Liu et al., 28 Dec 2025).

This suggests that the phrase “general+expert” is most precise when it denotes a transition from foundational competence to specialized competence under a staged increase in representational, structural, or decision complexity. What changes across domains is not the presence of progression, but the variable along which progression is defined: information type, task difficulty, support level, patch size, blur level, or optimization objective.

2. Difficulty representations and curriculum signals

A central development in recent curriculum design is the replacement of fixed, hand-labeled difficulty with model-relative signals. "Customized Curriculum Learning" defines per-sample difficulty from the model’s own empirical success rate: $$ACC_i = \frac{\sum_{j=1}^{n}\mathbf{1}\{A_{ij}=A_i^*\}}{n},$$ so lower $ACC_i$ means the sample is harder for that specific model (Wu et al., 4 Jun 2025). "VL-Cogito" uses an analogous online estimate for multimodal RL: $$\mathrm{Acc}(x) = \frac{1}{G}\sum_{i=1}^{G}\text{acc}(x,y_i),$$ and then uses stage-specific soft weighting functions over $\mathrm{Acc}(x)$ to emphasize easy, medium, or hard prompts at different stages (Yuan et al., 30 Jul 2025).

Other systems use richer notions of difficulty than scalar success probability. "SPICE" decomposes multimodal information via Partial Information Decomposition,

$$I(X_1, X_2; Y) = R(Y) + U_{X_1}(Y) + U_{X_2}(Y) + S(Y),$$

and builds curricula over redundant, unique, and synergistic information components rather than over a single easy/hard score (Singh et al., 15 Jun 2026). "TPCL" moves difficulty from the sample level to the task level: it builds per-task loss histograms and measures difficulty by the Wasserstein Optimal Transport distance between consecutive task-level score distributions,

$$\text{OT}(s^\tau_r, s^\tau_{r-1}) = \inf_{\gamma \in \Pi(s^\tau_r, s^\tau_{r-1})} \mathbb{E}_{(\mu,\nu)\sim\gamma}[d(\mu,\nu)],$$

so unstable tasks are treated as harder (Akl et al., 2024).

A parallel line appears in pseudo-label adaptation. "ElimPCL" defines trustworthy samples by low entropy and prototype consistency; only samples satisfying

$$D_{tt} = \{(x_t, \hat{y}_t) \mid \mathcal{H}(\hat{y}_t) < \tau,\ \hat{y}_t = \tilde{y}_t\}$$

enter the curriculum pool, while noisier items remain excluded until later (Cheng et al., 31 Mar 2025). In knowledge distillation, "Curriculum Extraction" treats the teacher’s own layer hierarchy as an implicit difficulty structure: the student first matches random projections of shallower teacher representations and only later matches the final logits (Gupta et al., 21 Mar 2025).

A common misconception is that a curriculum requires a universal scalar ranking of examples. These papers jointly indicate otherwise: difficulty may be typed rather than scalar, task-level rather than sample-level, online rather than static, or trust-based rather than accuracy-based.

3. Schedule design and progression mechanisms

Once difficulty is defined, systems differ in how they schedule exposure. "SPICE" provides two concrete schedules. SPICE-S bins samples by dominant PID type and advances as

$$\mathcal{D}_R \rightarrow \mathcal{D}_R \cup \mathcal{D}_U \rightarrow \mathcal{D}_R \cup \mathcal{D}_U \cup \mathcal{D}_S,$$

while SPICE-E uses the entire dataset in every epoch but reorders samples by stage-specific PID-derived probabilities. Both variants recompute PID scores every $k=5$ epochs, and the reported schedule includes a 30-epoch warm-up, followed by redundant-, unique-, and synergy-focused stages (Singh et al., 15 Jun 2026).

"Customized Curriculum Learning" partitions sorted data into $D_1, D_2, D_3$, trains in stages, and uses Guided Prompting to soften the hardest subset $ACC_i$0. A hard sample with solution steps

$ACC_i$1

is converted into an easier instance by prefixing the prompt with

$ACC_i$2

and the hint length is increased until a threshold $ACC_i$3 is reached or a maximum hint ratio $ACC_i$4 is exceeded (Wu et al., 4 Jun 2025). This schedule changes both problem difficulty and support level.

"VL-Cogito" uses a fully online RL schedule. It keeps the same multimodal dataset across stages, but changes the weighting function $ACC_i$5 over prompts and activates Dynamic Length Reward only in the hard stage. The curriculum is explicitly Easy → Medium → Hard, with about 100 optimization steps for easy, 100 for medium, and about 200 for the hard stage (Yuan et al., 30 Jul 2025). "Triton" parallels this pattern at the optimizer level, with SFT, then ORPO, then GRPO (Peng et al., 14 Apr 2026).

Dense prediction papers use input simplification as the scheduling variable. "Progressive Growing of Patch Size" starts from minimal feasible patch sizes and progressively increases them to the standard maximal nnU-Net patch, either with fixed batch size (resource-efficient mode) or with dynamically increased batch size (performance mode) (Fischer et al., 27 Oct 2025). "ObjBlur" starts from strongly blurred object/background regions and decays blur according to a schedule $ACC_i$6, with the best results reported for a sin schedule and a curriculum spanning 95% of training (Frolov et al., 2024). In medical imaging with occlusions, "See Through the Fog" similarly increases occlusion level across stages and augments cross-entropy with Wasserstein, mutual-information, or geodesic regularization terms to smooth the transition from clear to heavily occluded images (Singh et al., 2023).

Not all curricula are easy-to-hard. "TPCL" reports that hard-to-easy ordering under incremental pacing is advantageous for OOD robustness in VQA, especially in low-data settings, showing that stage ordering is domain- and objective-dependent (Akl et al., 2024).

4. Representative instantiations

The same general+expert logic now appears across multimodal learning, reasoning, dense prediction, adaptation, generation, control, and distillation.

Domain	Progression pattern	Representative paper
Multimodal interaction learning	Redundant $ACC_i$7 Unique $ACC_i$8 Synergistic	(Singh et al., 15 Jun 2026)
Mathematical reasoning	Easy $ACC_i$9 Medium $$\mathrm{Acc}(x) = \frac{1}{G}\sum_{i=1}^{G}\text{acc}(x,y_i),$$0 Hard with Guided Prompting	(Wu et al., 4 Jun 2025)
Multimodal RL reasoning	Easy $$\mathrm{Acc}(x) = \frac{1}{G}\sum_{i=1}^{G}\text{acc}(x,y_i),$$1 Medium $$\mathrm{Acc}(x) = \frac{1}{G}\sum_{i=1}^{G}\text{acc}(x,y_i),$$2 Hard with online difficulty soft weighting	(Yuan et al., 30 Jul 2025)
Robust VQA	Task-level progression over question types, often hard-to-easy	(Akl et al., 2024)
3D medical segmentation	Small patches $$\mathrm{Acc}(x) = \frac{1}{G}\sum_{i=1}^{G}\text{acc}(x,y_i),$$3 large patches	(Fischer et al., 27 Oct 2025)
Layout-to-image generation	Strong blur $$\mathrm{Acc}(x) = \frac{1}{G}\sum_{i=1}^{G}\text{acc}(x,y_i),$$4 clean images	(Frolov et al., 2024)
Web navigation	Imitation $$\mathrm{Acc}(x) = \frac{1}{G}\sum_{i=1}^{G}\text{acc}(x,y_i),$$5 Discrimination $$\mathrm{Acc}(x) = \frac{1}{G}\sum_{i=1}^{G}\text{acc}(x,y_i),$$6 Long-horizon consistency	(Peng et al., 14 Apr 2026)
HOI video generation	Human pose $$\mathrm{Acc}(x) = \frac{1}{G}\sum_{i=1}^{G}\text{acc}(x,y_i),$$7 hand-object interaction $$\mathrm{Acc}(x) = \frac{1}{G}\sum_{i=1}^{G}\text{acc}(x,y_i),$$8 full HOI	(Liu et al., 28 Dec 2025)
Source-free domain adaptation	Trustworthy pseudo-labels $$\mathrm{Acc}(x) = \frac{1}{G}\sum_{i=1}^{G}\text{acc}(x,y_i),$$9 progressive incorporation of harder samples	(Cheng et al., 31 Mar 2025)
Knowledge distillation	Projected shallow representations $\mathrm{Acc}(x)$0 deeper representations $\mathrm{Acc}(x)$1 full logits	(Gupta et al., 21 Mar 2025)

Two cross-cutting themes recur. First, many systems separate a broad foundational phase from a specialized refinement phase. Second, later stages often add either a new supervision type or a sharper optimization objective: Guided Prompting adds hints, ORPO adds preference comparisons, GRPO adds hierarchical reward, RCM adds 3D geometry control, and PID-based curricula add typed interaction structure.

This suggests a useful interpretation: a general+expert progressive curriculum is not merely “easy-to-hard data ordering,” but a staged increase in the kind of competence the model is required to exhibit.

5. Empirical profile

The empirical literature reports gains in both final performance and optimization efficiency. In multimodal interaction learning, SPICE-E exceeds the strongest non-SPICE curriculum baseline BSS-L on all four reported benchmarks: on CREMA-D it reaches 83.06% ACC and 89.07% mAP versus 82.80% ACC and 88.61% mAP; on Kinetics-Sounds, 73.99% ACC and 79.48% mAP versus 73.95% and 79.43%; on NVGesture, 87.14% ACC and 87.36% F1 versus 86.72% and 87.04%; and on VGGSound, 54.98% ACC and 56.63% mAP versus 52.80% and 56.61% (Singh et al., 15 Jun 2026).

For mathematical reasoning, CCL consistently outperforms uniform training. The largest reported gain is for Qwen2.5-Math-1.5B under GRPO, where the average score rises from 24.74 to 38.54 (+13.80 points). The same paper reports 21.34 → 22.38 for 1.5B SFT, 25.24 → 30.20 for 7B SFT, and 42.82 → 45.26 for 7B GRPO (Wu et al., 4 Jun 2025).

For robust VQA, TPCL yields large gains without explicit debiasing modules. With LXMERT on VQA-CP v2, the dynamic variant improves from 48.66% to 77.23% overall; on VQA-CP v1, from 52.82% to 76.15%; and on VQA v2, from 73.06% to 78.03% (Akl et al., 2024). For 3D medical segmentation, PGPS reports two operating points across 15 tasks: the resource-efficient mode matches constant patch size with training time reduced to 44%, while the performance mode achieves a statistically significant relative mean gain of 1.28% in Dice Score and still reduces training time to 89% (Fischer et al., 27 Oct 2025).

In web navigation, the three-stage progression is particularly pronounced. Triton-SFT-32B reaches 47.6% Step Success Rate, Triton-ORPO-32B reaches 53.2%, and Triton-GRPO-32B reaches 58.7%, surpassing GPT-4.5 (42.4%) and Claude-4.5 (41.4%) by over 16% on Mind2Web (Peng et al., 14 Apr 2026). In HOI video generation, curriculum ablations show that the full I + II + III schedule substantially improves object and hand metrics over I + III: Obj-IoU 0.8288 vs 0.7627, Obj-CLIP 0.9100 vs 0.8829, LMD 0.1427 vs 0.2054, and T-SSIM 0.5682 vs 0.4812 (Liu et al., 28 Dec 2025).

A plausible implication is that progressive curricula are most effective when the final task contains a clear asymmetry between foundational signal and specialized failure modes. The reported gains are largest precisely where static or one-shot training is weakest: multimodal synergy, noisy pseudo-label adaptation, long-horizon web control, lesion segmentation, and hard mathematical reasoning.

6. Limitations, debates, and future directions

No single progression rule is uniformly optimal. Several systems follow a canonical easy-to-hard or general-to-expert pattern, but "TPCL" reports that hard-to-easy ordering can be better for OOD robustness and limited-data VQA: with 30% of VQA-CP v2, dynamic TPCL with backward ordering reaches 72.58%, forward ordering 71.04%, and baseline 37.16% (Akl et al., 2024). This makes the direction of progression an open design choice rather than a settled principle.

A second limitation is that many curricula rely on proxies rather than exact task-theoretic quantities. "SPICE" explicitly notes that its $\mathrm{Acc}(x)$2 scores are heuristics from confidences and KL divergences rather than full Shannon PID estimators (Singh et al., 15 Jun 2026). "VL-Cogito" uses binary rollout correctness as an online difficulty proxy, which the paper itself characterizes as coarse (Yuan et al., 30 Jul 2025). "ElimPCL" depends on entropy and prototype consistency to define trustworthiness, which is powerful empirically but still contingent on the quality of the current feature geometry and pseudo-labels (Cheng et al., 31 Mar 2025).

A third limitation is computational overhead. PID-based curricula require periodic full-dataset passes; "SPICE" recomputes scores every $\mathrm{Acc}(x)$3 epochs (Singh et al., 15 Jun 2026). "TPCL" reports an OT overhead of about 50–78 ms per iteration for 65 tasks and histogram size 100–200 (Akl et al., 2024). "VL-Cogito" uses $\mathrm{Acc}(x)$4 rollouts and a 4096-token maximum sequence length during GRPO, and "Triton" adds ORPO pair synthesis and GRPO group rollouts on top of SFT (Yuan et al., 30 Jul 2025, Peng et al., 14 Apr 2026). The practical value of a curriculum therefore depends not only on accuracy but also on whether the staging mechanism yields a favorable compute/performance trade-off.

Future work in the cited literature converges on three directions. One is better difficulty estimation: more principled PID estimators, richer process-level rewards, or multi-signal difficulty models that combine correctness, uncertainty, and structure (Singh et al., 15 Jun 2026, Yuan et al., 30 Jul 2025). A second is scaling and transfer: extension to large multimodal foundation models, sequential interaction settings, larger students in knowledge distillation, and broader domain-specialization pipelines (Singh et al., 15 Jun 2026, Gupta et al., 21 Mar 2025, Neema et al., 30 Oct 2025). A third is adaptive scheduling: online re-estimation of difficulty, learned stage lengths, and uncertainty-aware use of human priors instead of static hand-designed curricula (Wu et al., 4 Jun 2025, Singh et al., 15 Jun 2026).

Taken together, the literature indicates that a general+expert progressive curriculum is most compelling when it satisfies four properties simultaneously: the early stage isolates high-signal structure, the intermediate stage targets the model’s current blind spots, the late stage exposes the model to interaction-rich or ambiguity-rich cases, and the scheduling variable is updated often enough to remain aligned with model evolution.