Easy-to-Hard Generalization in Learning Systems

• Easy-to-hard generalization is a modeling technique that structures training from simple tasks to complex challenges by ordering examples based on difficulty.
• It employs strategies such as difficulty-controlled benchmarks, curriculum learning, and learner-relative methods to optimize performance.
• Empirical results reveal both significant transfer benefits and limitations, highlighting the need for precise difficulty calibration and task-specific curricula.

Easy-to-hard generalization denotes the ability of a learning system to acquire competence from easier instances, easier subtasks, or easier forms of supervision and then transfer that competence to harder instances that preserve the same underlying structure (Schwarzschild et al., 2021). In recent work, the term spans at least three closely related settings: difficulty-controlled benchmarking, curriculum learning that orders data from easy to hard, and scalable supervision pipelines in which evaluators or self-improvement procedures trained on easier tasks are used to tackle harder reasoning problems (Mordig et al., 28 Mar 2026, Sun et al., 2024, Lee et al., 3 Feb 2025). The common research question is whether models learn the underlying rule or merely exploit superficial shortcuts that fail as difficulty increases (Schwarzschild et al., 2021). The empirical record is mixed: some studies report strong transfer or gains from carefully designed curricula, whereas others find that difficulty-based ordering offers no robust advantage over random sampling and that transfer is often local in difficulty space rather than global (Hase et al., 2024, Kordi et al., 26 Nov 2025, Mordig et al., 28 Mar 2026).

1. Core definitions and problem formulations

One formalization of curriculum learning introduces a difficulty function

$f:\mathcal{X}\to \mathbb{N}_{>0},$

where larger $f(x)$ denotes harder examples; a curriculum strategy then specifies which difficulty levels are sampled at each training phase (Mordig et al., 28 Mar 2026). In the canonical easy-to-hard variant, training begins with low difficulty and gradually expands to higher levels. In post-training studies of LLMs, this ordering is compared directly against uniform random sampling while holding the total training budget fixed, so that the effect of ordering itself can be isolated (Mordig et al., 28 Mar 2026).

A second formulation treats the topic as cross-difficulty generalization: train on examples of one difficulty and evaluate on another (Kordi et al., 26 Nov 2025). This includes both easy-to-hard generalization, in which training uses easier examples and evaluation uses harder examples, and hard-to-easy generalization, in which the direction is reversed (Kordi et al., 26 Nov 2025). In this setting, the central object is not the curriculum schedule but the train–test difficulty gap.

A third formulation makes difficulty explicitly learner-relative. The notion of transitional problems defines a problem $p$ as transitional at level $k$ if all weaker models fail on it and all models at or above level $k$ succeed: $$\forall i \in [k-1]: \phi_p(M_i)=0 \quad \land \quad \forall i \in [k,n]: \phi_p(M_i)=1.$$ This yields a model-specific easy-to-hard partition and a “level-up” curriculum that targets problems just beyond the current competence tier (Tang et al., 14 Mar 2026). A related task-centric theory organizes reasoning problems into difficulty-indexed distributions $p_1,\dots,p_L$, with easy-to-hard curricula moving progressively across this family rather than training on a fixed mixture (Liu et al., 10 Feb 2026).

These formulations differ in what “easy” and “hard” mean. Some work ties hardness to human ability to correctly label the data, using measures such as grade level, expert rating, Bloom-style cognitive skill, or reasoning-step count (Hase et al., 2024). Other work argues that human proxies can be poor measures of what is difficult for LLMs, and instead derives difficulty from large-scale model behavior via Item Response Theory (IRT) or related systems (Kordi et al., 26 Nov 2025, Ding et al., 2024). A recurring implication is that easy-to-hard generalization is not defined by a single universal difficulty axis.

2. Difficulty measurement and benchmark construction

Difficulty-controlled benchmarking is central to the literature because claims about easy-to-hard transfer depend strongly on how hardness is operationalized. Some benchmark families use transparent structural axes. In the dataset suite introduced specifically for studying generalization from easy training examples to harder test examples, Prefix Sums uses sequence length, Mazes uses maze size, and Chess Puzzles uses puzzle Elo difficulty rating (Schwarzschild et al., 2021). In deductive-reasoning post-training, MathGAP-derived LinearDepth defines difficulty as the number of inference steps, PartWhole defines it as the number of axioms, and Knights-and-Knaves defines it as the number of characters, with the authors emphasizing that difficulty is structural rather than merely a surface proxy such as prompt length (Mordig et al., 28 Mar 2026).

Other benchmark programs derive difficulty from performance statistics. A large-scale cross-difficulty study ranks items in ARC, GSM8K, MMLU-Pro, BBH, MATH, and MuSR using thousands of LLM responses and the Rasch / 1PL IRT model,

$$P(r_{ij} \mid \theta_j, \beta_i) = \frac{1}{1 + e^{-(\theta_j - \beta_i)}},$$

then partitions each dataset into 10 equal-sized bins from easiest to hardest (Kordi et al., 26 Nov 2025). Easy2Hard-Bench extends this idea into a standardized collection of six datasets with continuous numerical difficulty scores estimated from either IRT or Glicko-2, depending on domain (Ding et al., 2024).

Benchmark family	Tasks	Difficulty definition
(Schwarzschild et al., 2021)	Prefix Sums, Mazes, Chess Puzzles	Longer strings, larger mazes, higher Elo difficulty rating
(Mordig et al., 28 Mar 2026)	LinearDepth, PartWhole, Knights-and-Knaves	Inference steps, number of axioms, number of characters
(Kordi et al., 26 Nov 2025)	ARC, GSM8K, MMLU-Pro, BBH, MATH, MuSR	10 equal-sized IRT bins from easiest to hardest
(Ding et al., 2024)	E2H-AMC, E2H-Codeforces, E2H-Lichess, E2H-GSM8K, E2H-ARC, E2H-Winogrande	Numerical difficulty scores from IRT or Glicko-2

A major theme in this line of work is the divergence between human-oriented and model-oriented hardness. The IRT-based study reports that grade level, question length, answer length, reasoning-step count, Bloom’s taxonomy, Depth of Knowledge, and expert ratings correlate weakly or inconsistently with model-based difficulty; examples include a positive correlation up to about $0.56$ for MATH answer length and about $0.49$ for GSM8K reasoning steps, but many other correlations are near zero or negative (Kordi et al., 26 Nov 2025). Easy2Hard-Bench likewise argues for continuous, unified, performance-derived scores rather than coarse labels or domain-specific heuristics (Ding et al., 2024). This suggests that a central technical issue in easy-to-hard research is not only transfer, but calibration of the difficulty axis itself.

3. Training regimes and algorithmic patterns

Static easy-to-hard curricula remain the most direct operationalization. In deductive-reasoning post-training, five sampling strategies are compared: Standard, SingleDiffInc, SingleDiffDec, UpToDiff, and DownToDiff, under fixed epoch and sample budgets across SFT, GRPO, and PPO (Mordig et al., 28 Mar 2026). This setting isolates sequencing from confounds such as adaptive pacing or varying compute. Closely related work on mathematical reasoning evaluates Forward Curriculum Learning, Reverse Curriculum Learning, and grouped variants under identical offline fine-tuning settings, explicitly reframing curriculum quality as a question about data ordering rather than dataset content (Jia et al., 21 Oct 2025).

A more competence-aware family of methods constructs the curriculum relative to the learner. Transitional-problem curricula partition the data into subsets $f(x)$0 defined by the level at which problems flip from unsolved to solved across a monotonic model series, then train in Level-Up, Level-Down, or Random order (Tang et al., 14 Mar 2026). The distinctive claim is that the useful frontier is not globally easiest-first, but the set of problems just beyond the model’s current ability.

Several lines of work replace direct easy-to-hard imitation with auxiliary supervision. In scalable alignment for mathematical reasoning, process-supervised reward models are trained only on easy problems and then used to evaluate candidate solutions to hard problems, enabling reranking or RL without human supervision on the hard subset (Sun et al., 2024). In iterative self-improvement, the model generates pseudo-labels on slightly harder inputs and is retrained on the growing dataset; at round $f(x)$1,

$f(x)$2

with $f(x)$3, so that each round moves only slightly beyond previous competence (Lee et al., 3 Feb 2025).

Curriculum reinforcement learning makes the schedule part of the RL objective. E2H Reasoner decomposes a hard task into progressively more difficult subtasks and samples from them according to schedules such as a cosine fade from easy to hard or a Gaussian schedule that more aggressively fades out easy tasks (Parashar et al., 7 Jun 2025). This differs from simple easy-first staging because the method explicitly treats prolonged exposure to easy tasks as a source of overfitting.

Beyond generic reasoning, easy-to-hard structures are also embedded into task-specific architectures. FE2H for multi-hop question answering uses a two-stage selector that retrieves one supporting document and then conditions on it to retrieve the second, and a two-stage reader trained first on SQuAD and then on HotpotQA (Li et al., 2022). CAFE for panoptic scene graph generation aligns predicate difficulty with feature complexity, moving from bbox features to mask features to boundary features (Shi et al., 2024). MetaLint trains on easy Ruff-detectable idioms and evaluates on harder PEP-inspired idioms outside the linter’s direct coverage (Naik et al., 15 Jul 2025).

4. Empirical findings: strong transfer, local transfer, and negative results

A substantial body of evidence reports surprisingly strong easy-to-hard transfer under some conditions. In question answering with pretrained LMs, easy supervision often recovers 70%–100% of the supervision gap on hard test data, and models trained on easy data can perform nearly as well as oracle models finetuned on hard data across ARC, MMLU, and StrategyQA, using in-context learning, linear classifier heads, and QLoRA (Hase et al., 2024). That result is especially salient because it is framed by the scalable oversight problem: easy data may be cheaper and less noisy than hard data, yet still competitive for hard-test accuracy (Hase et al., 2024).

Evaluator-guided supervision yields a different positive result. In mathematical reasoning on MATH, direct easy-to-hard SFT is weaker than full SFT, but reward models trained on easy problems generalize better than generators trained on easy problems, and can then improve hard-task performance through reranking or RL (Sun et al., 2024). A process-supervised 7b RL model reaches 34.0% on MATH500 and a 34b reranking system reaches 52.5% on MATH500 despite using human supervision only on easy problems (Sun et al., 2024). This is a strong form of easy-to-hard generalization because the bridge from easy supervision to hard performance is evaluation rather than generation.

Iterative self-improvement provides another positive regime. Standard transformers trained on easier examples and then iteratively retrained on self-generated solutions can generalize from 1–16 digits to over 100 digits in reverse addition, from 1–10 to over 120 digits in copy and reverse, maintain >98% accuracy up to length 70 in forward addition with filtering, extend multiplication from $f(x)$4 to $f(x)$5 near-perfectly, and generalize maze solving to 30-hop paths when filtering is used (Lee et al., 3 Feb 2025). Controlled weak-to-strong curricula and filtering are essential in these results; without filtering, several tasks collapse (Lee et al., 3 Feb 2025).

Learner-specific curricula also produce gains. In chess and mathematics, Level-Up training on transitional problems consistently outperforms random ordering, hard-to-easy ordering, and curricula based on external heuristics such as length, number of steps, legal moves, or human ratings (Tang et al., 14 Mar 2026). The claim here is not that any easy-first schedule works, but that the right easy-to-hard order is the one defined relative to the learner’s competence.

Against these positive results stand several prominent negative or strongly qualified findings. In deductive-reasoning post-training on synthetic arithmetic and logical benchmarks, difficulty-based sequencing shows no robust advantage over standard uniform random sampling in either accuracy or response length, and this remains true across multiple model families, curriculum schedules, SFT, and RL (Mordig et al., 28 Mar 2026). The paper’s conclusion is not that post-training fails, but that the gains from post-training itself are much larger than any gains from curriculum order (Mordig et al., 28 Mar 2026).

Cross-difficulty transfer is also often sharply local. Fine-tuning on a single IRT difficulty bin yields the best transfer to nearby bins, and performance drops as the train–test difficulty gap widens; with sufficiently large gaps, performance can even fall below the zero-shot baseline (Kordi et al., 26 Nov 2025). Easy2Hard-Bench reports a similar generalization ridge along the diagonal where training difficulty is close to evaluation difficulty, and finds that easy-to-hard transfer is not uniformly strong (Ding et al., 2024). These studies reject the idea that easy data or hard data alone can serve as a universal substitute across the full difficulty range.

The literature on curriculum direction further complicates the picture. A unified study of mathematical reasoning concludes that no curriculum strategy dominates universally: the relative value of forward versus reverse curriculum depends jointly on the difficulty metric, the model’s capability, and task complexity (Jia et al., 21 Oct 2025). In another line of work, meta-learning with learned reweighting finds that using hard validation data performs consistently better than using easy validation data and improves classifier margins, implying that in some settings the hard subset is the more informative supervisory target (Jain et al., 2024). Taken together, these results suggest that easy-to-hard generalization is real but conditional, and that its success depends on how supervision, difficulty, and optimization are coupled.

5. Mechanistic explanations and theoretical accounts

A recurrent explanation for positive results is that evaluation is easier than generation. In scalable alignment for mathematical reasoning, direct easy-to-hard generation is weak, but easy-to-hard evaluation is stronger: process reward models trained on easier tasks can score candidate hard-task solutions well enough to support reranking and RL (Sun et al., 2024). This reframes the problem from “generate hard answers from easy demonstrations” to “learn an evaluator on easy tasks that still recognizes good hard-task reasoning.”

Another explanation centers on competence-aware frontiers. Transitional-problem curricula assume that the most useful training examples are those that sit just beyond the current learner’s ability, rather than the globally easiest examples (Tang et al., 14 Mar 2026). The educational intuition is explicit: tasks near the next competence tier are maximally useful for “leveling up.” This suggests that the practical value of easy-to-hard ordering depends less on abstract difficulty and more on alignment between task difficulty and current policy quality.

Theoretical analyses of self-improvement formalize a positive feedback loop. In a task-centric theory, each iteration fine-tunes on a reward-filtered distribution, and better models accept more samples per iteration, increasing the effective training set and improving subsequent updates (Liu et al., 10 Feb 2026). The framework defines a global acceptance rate

$f(x)$6

and shows that this acceptance-mediated data expansion supports sustained improvement but also explains eventual saturation (Liu et al., 10 Feb 2026). In the binary-reward case, the theory derives a recursive lower bound

$f(x)$7

making the feedback loop explicit (Liu et al., 10 Feb 2026).

Curriculum RL adds a complementary API-based account. E2H Reasoner proves convergence guarantees and finite-sample complexity bounds under a decomposition of a hard task into progressively harder subtasks, and argues that when tasks are appropriately decomposed and conditioned, learning through curriculum stages requires fewer total samples than direct learning on the hard distribution (Parashar et al., 7 Jun 2025). A key practical claim is that easy tasks are useful only initially; they must be faded out through appropriate scheduling to avoid overfitting (Parashar et al., 7 Jun 2025).

Negative results motivate different mechanisms. The deductive-reasoning study argues that if models mostly learn surface-level patterns rather than abstract compositional rules, then example ordering should matter little; the invariance of response length across curricula is presented as evidence that curriculum order does not alter the model’s effective reasoning depth (Mordig et al., 28 Mar 2026). The large-scale IRT study similarly argues that easy training data often lacks the coverage needed for harder decision boundaries, producing transfer that is strongest only near the training difficulty band (Kordi et al., 26 Nov 2025). A plausible synthesis is that ordering helps when the curriculum is competence-aware or verifier-supported, but not when the underlying training signal fails to induce rule learning.

6. Extensions, applications, and unresolved issues

Easy-to-hard generalization has been instantiated in multiple application domains beyond generic reasoning benchmarks. In multi-hop question answering, FE2H uses easy-to-hard progression in both document selection and reader transfer, reporting 71.89 Answer EM, 84.44 Answer F1, 64.98 Supporting EM, 89.14 Supporting F1, 50.04 Joint EM, and 76.54 Joint F1 on HotpotQA distractor, while the two-stage selector reaches 96.32 EM and 98.02 F1 on document selection (Li et al., 2022). Here, “easy” refers both to the first retrieval hop and to single-hop QA pretraining before multi-hop transfer.

In panoptic scene graph generation, CAFE turns easy-to-hard progression into a curriculum over feature complexity. Easy predicates are learned with bbox features, medium predicates add mask features, and hard predicates add boundary features; the curriculum and distillation strategy improve both robust and zero-shot PSG, with Transformer + CAFE reaching 46.1 average in zero-shot PredCls compared with 37.6 for the Transformer baseline (Shi et al., 2024). This is an instance in which the easy-to-hard axis applies to predicates and representational richness rather than to sample count or sequence length.

In machine-generated text detection, an easy-to-hard supervision framework trains an easier supervisor on longer texts, where class boundaries are clearer, and uses structural coupling to improve the harder target detector on shorter or more ambiguous texts (Wu et al., 2 Nov 2025). Across 312 cross-LLM settings, the enhancement beats the baseline in 87% of sentence-level cases and 85% of paragraph-level cases (Wu et al., 2 Nov 2025). In code quality analysis, MetaLint trains on Ruff-detectable idioms and transfers to unseen PEP-inspired idioms, reaching 70.37% detection F-score, 70.43% detection recall, and 26.73% localization on the hard benchmark (Naik et al., 15 Jul 2025). These systems treat “easy” as automatically verifiable or structurally clearer supervision.

The topic also extends beyond LLMs. In CNN training on CIFAR-10 and ImageNet-1k, easy examples are defined by low early-training loss averaged across runs, hard examples by high early-training loss; easy examples are visually similar, hard examples visually diverse, and both sets are largely shared across architectures (Kishida et al., 2019). Hard examples contribute more individually to generalization, but removing too many easy examples still harms accuracy, and a stepwise easiness-based removal strategy yields about 1.1% accuracy gain over random removal at 0.3 ablation ratio on ImageNet-1k (Kishida et al., 2019). This complicates any simplistic view that easy examples are merely redundant.

Several unresolved issues recur across the literature. Difficulty may be structural, human-defined, model-defined, or learner-relative; these notions do not reliably coincide (Kordi et al., 26 Nov 2025, Ding et al., 2024). Many conclusions come from synthetic or highly controlled tasks, medium-sized models, fixed training budgets, or narrow domain coverage, and several papers explicitly do not study adaptive curricula, convergence speed under alternate stopping criteria, or broader multilingual settings (Mordig et al., 28 Mar 2026, Ding et al., 2024). The resulting picture is therefore not a universal doctrine of “start easy, end hard,” but a heterogeneous research area in which transfer can be strong, weak, or absent depending on how difficulty is measured, how supervision is delivered, and whether the training signal targets rule learning, calibration, or search.