Staged Reasoning Curriculum

• Reasoning curriculum is a family of training strategies that organizes learning signals through staged problem difficulty, decomposed subproblems, and dynamic reward mechanisms.
• It employs model-relative difficulty estimation and progressive reduction of hints to adapt supervision timing and structure to the model’s evolving capabilities.
• This approach enhances optimization stability and sample efficiency, demonstrating practical benefits in math, logic, vision-language, and domain transfer tasks.

Reasoning curriculum is a family of training strategies that organizes problems, supervision signals, or reward components so that a model acquires reasoning behavior under a staged or adaptive schedule rather than under uniform sampling. In recent LLM and MLLM work, the term no longer refers only to a fixed easy-to-hard ordering of examples. It now includes problem-relative staging by current model proficiency, decomposition into verifiable subproblems, reward-level stratification, difficulty-aware sampling inside RLVR, progressive reduction of hints or prefixes, and curricula over modalities, languages, or even diffusion block sizes. Across these variants, the common objective is to improve sample efficiency, stabilize optimization, and expose the model to learnable but nontrivial reasoning signals instead of trivial or hopelessly hard cases (Jiang et al., 21 May 2026, Lai et al., 29 Jan 2026, Chen et al., 20 May 2025, Hu et al., 11 Mar 2026).

1. Conceptual scope and units of curriculum

The modern literature treats a reasoning curriculum as a control policy over what the model practices, when it practices it, and how learning signal is assigned. Some frameworks define the curriculum over examples. TACLer partitions the training set into three proficiency bands—G1 for correct final answer, G2 for complete but incorrect reasoning, and G3 for incomplete response—and stages training as $G1 \cup G2$, then partial G3, then the full dataset, with repeated relabeling after each stage (Lai et al., 29 Jan 2026). CCL similarly defines difficulty through the model’s own per-sample success rate $ACC_i$, then sorts data into easy, medium, and hard partitions and uses Guided Prompting on the hardest split rather than discarding it (Wu et al., 4 Jun 2025). ADCL also rejects static difficulty rankings, periodically re-estimating difficulty for the next batch only, because “Difficulty Shift” causes a ranking computed once to drift as the policy changes (Zhang et al., 13 May 2025).

Other frameworks define the curriculum over reasoning structure rather than over whole examples. SCRL turns a single hard problem into $K$ linked, self-contained, verifiable subproblems, with the final subproblem fixed to the original problem, so that partial progress becomes trainable signal within one rollout (Jiang et al., 21 May 2026). DGRC constructs a curriculum “from atoms to chains”: it first creates high-confidence atomic Q&A pairs from teacher-student disagreements, then uses those atomic facts to verify or reject teacher chains of thought before supervised adaptation (Wang et al., 27 Jan 2026). IRIS adds a second axis: a vertical SFT curriculum over Easy $\rightarrow$ Easy+Medium $\rightarrow$ Easy+Medium+Hard, and a horizontal reverse curriculum in which progressively shorter prefixes of gold reasoning are given and the model must generate longer suffixes (Gupta et al., 27 Apr 2026).

A further shift is to define the curriculum over rewards or optimization signal. RuCL explicitly reframes curriculum learning as reward design rather than data selection, stratifying generalized rubrics into easy and hard tiers and increasing the weight on hard rubrics only after foundational rubric performance stabilizes (Chen et al., 25 Feb 2026). VL-Cogito applies stage-specific online difficulty soft weighting to advantages and activates a dynamic length reward only in the hard stage (Yuan et al., 30 Jul 2025). SEC goes further and learns the curriculum policy online as a non-stationary multi-armed bandit over problem categories, using absolute advantage as a proxy for immediate learning gain (Chen et al., 20 May 2025).

This breadth suggests that “reasoning curriculum” now functions as an umbrella term for staged control over any variable that governs reasoning acquisition: examples, subgoals, hints, reward terms, output length, modality use, or domain transfer.

2. Difficulty estimation and curriculum construction

A central issue is that curriculum quality depends on the difficulty signal being aligned with the current policy. Several papers therefore replace static, human-defined difficulty with model-relative estimates. TACLer uses greedy inference under an 8k context to label each training instance as G1, G2, or G3, thereby operationalizing “difficulty” as what the current model solves, nearly solves, or cannot finish (Lai et al., 29 Jan 2026). CCL estimates $ACC_i$ by sampling $n$ responses per problem and sorting by observed success, arguing that predefined labels such as MATH levels can disagree with actual model capability (Wu et al., 4 Jun 2025). DeReason uses an LLM judge to assign reasoning intensity on a 1–5 scale and routes $s \le 3$ to SFT and $s \ge 4$ to RL, thereby decoupling “knowledge-establishing” and “reasoning-intensive” subsets (Hu et al., 11 Mar 2026).

Online curricula often rely on rollout statistics. CLPO computes $\mathrm{Acc}(q,\pi_\theta)$ from current on-policy samples, then partitions prompts into hard, medium, and easy using $ACC_i$0 and $ACC_i$1. It then simplifies hard problems and diversifies medium problems through Adaptive Problem Restructuring, keeping only items with nondegenerate signal $ACC_i$2 (Zhang et al., 29 Sep 2025). VL-Cogito similarly measures per-prompt rollout accuracy and rescales advantages through stage-specific weighting functions $ACC_i$3, using different easy, medium, and hard variants across its three RL stages (Yuan et al., 30 Jul 2025). GanitLLM uses a stronger evaluator model to generate $ACC_i$4 solutions per Bengali math problem, then bins problems by the raw correct count $ACC_i$5 into Easy, Medium, Hard, and Olympiad, and further stratifies RL sampling by exact $ACC_i$6 (Dipta et al., 11 Jan 2026).

The most systematic attempt to compare difficulty notions is the metric-driven study of mathematical curricula that decomposes difficulty into five dimensions: Problem Difficulty, Model Surprisal, Confidence Margin, Predictive Uncertainty, and Decision Variability (Jia et al., 21 Oct 2025). In that framework, ACC and VACC are treated as task-aligned metrics, whereas SLP, TLP, LG, SLE, and TLE are treated as inner-state metrics. Its central conclusion is negative but important: no curriculum direction dominates universally, and the relative value of forward versus reverse ordering depends jointly on model capability, task complexity, and the metric used (Jia et al., 21 Oct 2025).

This suggests that curriculum design is not merely a scheduling problem; it is also a measurement problem. A curriculum built on the wrong notion of difficulty can be well ordered and still be pedagogically misaligned.

3. Credit assignment and optimization mechanisms

Reasoning curricula are tightly coupled to credit assignment because sparse outcome rewards are often the motivating failure mode. SCRL addresses this most directly. It verifies each tagged subproblem independently, applies progress-aware correction so that later subproblems receive no credit after the first miss, normalizes rewards columnwise by subproblem position, and assigns the resulting advantages back to the corresponding answer spans. The paper’s theoretical analysis formalizes the failure of outcome-only GRPO on hard problems as a “gradient dead zone,” and shows that subproblem curricula can recover useful signal with a recovery ratio that scales as $ACC_i$7 as problems become harder (Jiang et al., 21 May 2026).

SEC offers a complementary optimization view. It uses the absolute value of the estimated advantage as the bandit reward for curriculum selection. Under GRPO with binary rewards, the expected absolute normalized advantage becomes $ACC_i$8, which is maximized at $ACC_i$9; this yields an explicit “zone of proximal development” interpretation for why medium-difficulty problems are most informative for policy updates (Chen et al., 20 May 2025). PC-GRPO reaches a similar conclusion from a different direction: it defines a difficulty-aware weight $K$0 that peaks at medium difficulty and uses it to scale the GRPO surrogate, while Jigsaw partial credit increases within-group dispersion and mitigates vanishing group-relative advantages (Jeddi et al., 16 Dec 2025).

Several frameworks combine sparse correctness rewards with auxiliary process rewards. IRIS uses a composite reward

$K$1

where $K$2 is a stage-dependent semantic alignment reward over the continuation suffix, $K$3 rewards correct step numbering, and $K$4 encourages numeric answers; $K$5 decays to zero as the model is asked to generate more of the solution autonomously (Gupta et al., 27 Apr 2026). MedCCO combines strict close-ended accuracy rewards with an open-ended hybrid lexical-semantic reward and a format reward, using a two-stage curriculum from close-ended to open-ended medical VQA (Rui et al., 25 May 2025). GanitLLM uses a low-resource variant of this idea, defining $K$6 so that even wrong solutions can receive nonzero reward for structural compliance and Bengali reasoning, thereby reducing all-zero rollout groups early in training (Dipta et al., 11 Jan 2026).

RuCL makes reward stratification itself the curriculum. Its rubric reward is mixed as

$K$7

and the transition is triggered when the batch-averaged foundational reward exceeds a threshold over a sliding window (Chen et al., 25 Feb 2026). VL-Cogito applies a distinct mechanism: in the hard stage, it adds a dynamic length reward based on the average length of correct responses for the current prompt, but downweights zero-accuracy prompts by $K$8 to avoid runaway verbosity (Yuan et al., 30 Jul 2025). In multimodal reasoning, “Learning When to Look” turns the timing of <perception> insertions into an RL problem and adds a Pivotal Perception Reward that couples perception actions to lexical markers of uncertainty such as “wait,” “verify,” or “check,” with a selectivity regularizer to avoid degenerate keyword hunting (Yang et al., 19 Dec 2025).

Across these methods, the curriculum is inseparable from the granularity of reward attribution. A plausible implication is that many recent “curriculum” gains arise less from ordering alone than from making the learning signal match the structure of reasoning progress.

4. Empirical performance and training dynamics

Reported gains are heterogeneous in magnitude, domain, and mechanism, but a recurring empirical pattern is that curricula help most when baseline training wastes updates on trivial or zero-signal cases. This is especially explicit in RLVR settings, where sparse outcome rewards make sample efficiency a first-order concern. It is also visible in low-resource and multimodal regimes, where curricula often improve not only accuracy but training stability, exploration, output length, or reasoning-language fidelity.

Framework	Setting	Reported result
SCRL	Qwen3-4B-Base on seven mathematical reasoning benchmarks	35.0% vs 30.9% average accuracy over GRPO, a +4.1 point gain (Jiang et al., 21 May 2026)
DGRC	1.5B student in medical unlabeled adaptation	Avg-M 58.3 vs 54.1, a 7.76% relative improvement over the unlabeled baseline (Wang et al., 27 Jan 2026)
TACLer	R1-Qwen-1.5B on four math datasets	training compute reduced by over 50%, inference token usage reduced by over 42%, accuracy improved by over 9% over the base model (Lai et al., 29 Jan 2026)
Bidirectional Curriculum Generation	Qwen3-8B-Base with 5,873 generated samples	Average 60.03, surpassing Fast-Math at 55.76 and MegaScience at 52.50 (Hu et al., 5 Mar 2026)
MedCCO	Qwen2.5-VL-7B on medical VQA	in-domain average 79.5% vs 68.1% and out-of-domain average 54.2% vs 48.5% (Rui et al., 25 May 2025)
GanitLLM	Bengali 4B model on Bn-MGSM and Bn-MSVAMP	69.20 → 76.80 and 70.50 → 76.40; Bengali reasoning 14.79% → 88.71%; average words 943 → 193 (Dipta et al., 11 Jan 2026)
Cognivolve	GPT-2 124M four-stage curriculum	success thresholds 0.25 and 0.30 reached in 500 vs 1,000 updates; 7.8× more specialized heads per checkpoint (Fu, 16 May 2025)

Beyond headline accuracy, several papers analyze training dynamics. SEC shows that adaptive category sampling improves OOD performance across planning, inductive reasoning, and math while adding minimal overhead because it reuses advantages already computed by the RL algorithm (Chen et al., 20 May 2025). TACLer reports that the largest gains occur in Stage 1, and that hybrid Thinking/NoThinking training produces a compression effect in which Thinking responses become shorter while preserving or improving accuracy (Lai et al., 29 Jan 2026). PC-GRPO tracks Reasoning-Answer Consistency and observes that vanilla GRPO raises RAC early and then degrades, whereas difficulty-aware curriculum weighting delays this decline and the consistency-aware add-on improves RAC further (Jeddi et al., 16 Dec 2025). ADCL shows that static predefined rankings drift substantially, with Normalized Inversion Rate rising from 0.000 in Batch 1 to 0.331, 0.363, and 0.369 in Batches 2–4, which quantifies why local re-estimation matters (Zhang et al., 13 May 2025).

The empirical record is therefore not simply “curricula improve accuracy.” Many of the strongest results are actually about how they improve learning: by recovering gradient signal, reducing training truncation, stabilizing cross-domain transfer, or making reasoning traces shorter, more faithful, or more language-constrained.

5. Domain extensions, multimodality, and transfer

Mathematical reasoning remains the dominant testbed, but the curriculum literature has expanded into domain adaptation, vision-language reasoning, medical decision support, diffusion models, and multilingual reasoning. One major line treats math as a bootstrap domain. “Reasoning Curriculum: Bootstrapping Broad LLM Reasoning from Math” uses a two-stage schedule in which math-only RL first elicits subgoal setting, enumeration, backtracking, and verification, and mixed-domain RL then transfers these behaviors to code, STEM, logic, simulation, and tabular tasks (Pang et al., 30 Oct 2025). DeReason reaches a related conclusion from the opposite angle: in general STEM, SFT is more sample-efficient for broad knowledge acquisition, while RL is most useful on reasoning-intensive subsets, so the curriculum is a decoupling of labor between stages rather than a universal escalation of difficulty (Hu et al., 11 Mar 2026).

Multimodal work extends the same logic to perception. MedCCO stages close-ended medical VQA before open-ended medical VQA, arguing that sharper discrete rewards stabilize early policy updates and that later hybrid lexical-semantic rewards can then refine clinical reasoning and interpretability (Rui et al., 25 May 2025). RuCL stages reward criteria from foundational perception to advanced logic by increasing weight on hard rubrics only after the easy tier stabilizes (Chen et al., 25 Feb 2026). “Learning When to Look” first trains text-only abstract reasoning, then anchors it with Perception-Grounded Chain-of-Thought, and finally learns a policy for when to insert perception segments during multimodal reasoning (Yang et al., 19 Dec 2025). PC-GRPO uses self-supervised visual puzzles as a curriculum for VLM RL post-training, replacing external verifiers with programmatic checks in Rotation, PatchFit, and Jigsaw environments (Jeddi et al., 16 Dec 2025). VL-Cogito generalizes staged curriculum RL across mathematics, science, logic, charts, and general image understanding through its three-stage PCuRL design (Yuan et al., 30 Jul 2025).

Curriculum ideas have also entered alternative generative architectures. DreamReasoner-8B studies block diffusion LLMs and shows that large training block sizes collapse diffusion reasoning, while a block-size curriculum—three epochs at $K$9 followed by mixed $\rightarrow$0—preserves fine-grained reasoning and generalizes across inference block sizes (Wu et al., 17 Jun 2026). In multilingual reasoning, IRIS combines curriculum over problem difficulty with curriculum over reasoning horizon, and GanitLLM combines SFT plus difficulty-aware GRPO to stabilize Bengali reasoning under low-resource reward sparsity (Gupta et al., 27 Apr 2026, Dipta et al., 11 Jan 2026). DGRC extends curricula to unlabeled medical and legal adaptation by deriving atomic curricula from teacher-student disagreements and verified chain curricula from atomic fact consistency (Wang et al., 27 Jan 2026).

This breadth suggests that the field increasingly treats curriculum not as a math-specific trick but as a generic mechanism for aligning exploration, supervision, and verification with the structure of reasoning in a given domain.

6. Limitations, controversies, and design principles

The literature is unusually explicit that curriculum learning is not universally beneficial under a single recipe. The most direct statement comes from the metric-driven study of data ordering, which finds that no curriculum strategy dominates universally, that forward versus reverse depends on both model capability and task complexity, and that inner-state curricula and task-aligned curricula affect different aspects of learning (Jia et al., 21 Oct 2025). This directly challenges any simple claim that “easy-to-hard” is the correct default.

A second limitation is dependence on auxiliary structure. SCRL requires reference chains and deterministic subproblem verifiers; poorly constructed intermediate subproblems can block later credit under progress-aware correction (Jiang et al., 21 May 2026). DGRC depends on a sufficiently capable diagnostician and can miss errors when teacher and student agree but are both wrong (Wang et al., 27 Jan 2026). Bidirectional Curriculum Generation depends on difficulty tagging and verifier quality, with persistent-failure thresholds introducing an additional tuning knob (Hu et al., 5 Mar 2026). RuCL reduces reward hacking relative to outcome-only RLVR, but still depends on a capable rubric judge and static rubric stratification (Chen et al., 25 Feb 2026). GanitLLM’s Bengali-character threshold can penalize legitimate mixed-script solutions, and its evaluator-based difficulty tags may inherit evaluator bias (Dipta et al., 11 Jan 2026).

A third issue is compute and engineering overhead. Some curricula are light because they reuse signals already present in the RL loop, as in SEC (Chen et al., 20 May 2025). Others require offline generation of subproblems, atomic facts, PG-CoT traces, or staged relabeling passes over the dataset, as in SCRL, DGRC, “Learning When to Look,” and TACLer (Jiang et al., 21 May 2026, Wang et al., 27 Jan 2026, Yang et al., 19 Dec 2025, Lai et al., 29 Jan 2026). DreamReasoner and multimodal frameworks add domain-specific kernels, structured prompts, or online judging costs (Wu et al., 17 Jun 2026, Chen et al., 25 Feb 2026).

Within these constraints, several design rules recur across papers. Difficulty should usually be defined relative to the current model rather than treated as a static property of the dataset (Lai et al., 29 Jan 2026, Wu et al., 4 Jun 2025, Zhang et al., 13 May 2025). Medium-difficulty or medium-variability cases often maximize learning signal, whether formalized through absolute advantages, curriculum weights, or decision variability (Chen et al., 20 May 2025, Jeddi et al., 16 Dec 2025, Jia et al., 21 Oct 2025). Hard cases are most useful when restructured, decomposed, hinted, or otherwise converted into learnable subproblems rather than naively sampled under sparse outcome rewards (Jiang et al., 21 May 2026, Wu et al., 4 Jun 2025, Zhang et al., 29 Sep 2025). Mixed or staged training is often used to avoid prompt mismatch or catastrophic forgetting, whether by combining curriculum and original prompts, mixing no-divergence chains, or revisiting easier items in later stages (Jiang et al., 21 May 2026, Wang et al., 27 Jan 2026, Wu et al., 4 Jun 2025).

Taken together, these results suggest a restrained but durable conclusion. A reasoning curriculum is effective when it aligns supervision granularity, example difficulty, and verifier reliability with the model’s evolving competence. It is ineffective, or even harmful, when difficulty is mismeasured, when the ordering ignores model state, or when auxiliary rewards and decompositions introduce brittle artifacts. The field’s central open question is therefore not whether curriculum matters, but which curriculum variable—data, subgoals, reward components, response length, modality use, or stage allocation—should be adapted for a given model, domain, and verification regime.