Hard Sample Generation and Learning

Updated 5 July 2026

HSGL is a design pattern that defines instance difficulty relative to model failures and integrates a hardness oracle with tailored generation mechanisms.
It utilizes methods such as neural policy networks, adversarial synthesis, and contrastive learning to produce challenging and semantically valid samples.
Empirical results demonstrate significant performance improvements in graph algorithms, re-identification, tool-use agents, and deep learning through optimized hard-sample training.

Searching arXiv for papers and representative uses of “Hard Sample Generation and Learning”. Hard Sample Generation and Learning (HSGL) denotes a family of methods that explicitly define what makes an instance difficult for a target solver or learner, synthesize or select such instances, and incorporate them into optimization or evaluation. In the literature summarized here, the term covers probabilistic generation of hard graph instances for classical algorithms, multimodal construction of hard positives and negatives for clothes-changing person re-identification, hard trace and query synthesis for tool-use agents, anchor-based correction of hard samples under instance-dependent label noise, synthetic hard negative construction for contrastive learning, adversarial hard-example generation for deep hashing, and a related theoretical formulation of learning-to-sample for Ising models (Sato et al., 2019, Liu et al., 15 Jul 2025, Hao et al., 4 Jan 2026, Huang et al., 2024, Zhu et al., 2021, Dong et al., 2023, Jin et al., 2019, Risteski et al., 23 May 2026).

1. Core definitions and problem formulations

Across these works, “hardness” is not a single universal quantity. It is defined relative to the failure mode of a particular algorithm, model, or inference task. In graph algorithms, hardness is the measurable cost of running a target algorithm $L$ on a graph $G$ , instantiated as the number of recursive calls, the number of SAT decisions, or wall-clock time. In clothes-changing person Re-ID, hard samples are semantically specified pairs: hard positives are same-identity pairs with different clothing or viewpoint, and hard negatives are different-identity pairs with the same clothing. In instance-dependent noisy-label learning, hard samples are not equated with noisy ones; the method explicitly distinguishes clean versus noisy and easy versus hard. In metric and contrastive learning, hardness is primarily geometric, centered on positives near the class boundary or negatives close to the anchor in embedding space. In tool-use agents, hardness is tied to implicit dependencies, executable multi-step traces, and verifiable complex Chain-of-Thought. In the Ising-model setting, the task is formalized as learning a computationally efficient sampler whose output distribution is within total-variation distance $\epsilon$ of the target (Sato et al., 2019, Liu et al., 15 Jul 2025, Huang et al., 2024, Zhu et al., 2021, Dong et al., 2023, Hao et al., 4 Jan 2026, Risteski et al., 23 May 2026).

Domain	Hardness notion	Representative formulation
Graph algorithms	High solver cost on $G$ for algorithm $L$	recursive calls, SAT decisions, wall-clock time
CC-ReID	Semantically confusing positive/negative pairs	$\mathrm{IS\_HP}_{ij}$ , $\mathrm{IS\_HN}_{ij}$
Noisy-label learning	Hard but potentially clean samples near boundaries	anchor hallucination + hard label correction
Metric / contrastive learning	Samples close in embedding space but semantically distinct	hard positives, hard negatives, triplets
Tool-use agents	Complex executable traces with implicit dependencies	dynamic API graph, hard traces, hard queries
Ising learning-to-sample	Computational difficulty of generating fresh approximate samples	$d_{\mathrm{TV}}$ -bounded sampler

A central consequence of this diversity is that HSGL is best viewed as a design pattern rather than a single algorithm. Each instantiation specifies: a hardness oracle or surrogate, a generation or selection mechanism, and a downstream learner or solver that is adapted to the resulting hard distribution.

2. Mechanisms for generating hard samples

The earliest formulation in this set is HiSampler for graph algorithms. Graphs with fixed $n$ are encoded by the upper-triangular adjacency vector $A \in \{0,1\}^{n(n-1)/2}$ . A fully connected neural policy network takes noise $G$ 0, produces Bernoulli edge probabilities $G$ 1, and samples each edge independently. In the reported experiments, the multilayer perceptron uses dimensions $G$ 2, ReLU hidden activations, and a sigmoid output layer. This construction yields a learned generative distribution over graphs without hand-engineered features, and HiSampler-PER augments the basic sampler with a priority queue of the top- $G$ 3 hardest instances for replay (Sato et al., 2019).

In clothes-changing person Re-ID, HSGL uses Dual-Granularity Hard Sample Generation (DGHSG). The coarse-grained branch, CHPSG, segments a clothing mask $G$ 4 from a source image and applies a text-conditional diffusion model with the prompt “generate a new cloth for the person,” producing

$G$ 5

The fine-grained branch, SHSG, filters for frontal high-quality images using MediaPipe-based keypoint checks and image-quality criteria, extracts garment tiles with TryOffAnyone, and synthesizes new images through IDM-VTON. The distinction between coarse and fine granularity is explicit: the former prioritizes scale and variety, while the latter uses actual garment images, textual garment descriptions, and identity images for tighter semantic control (Liu et al., 15 Jul 2025).

HardGen for tool-use agents constructs a dynamic API graph $G$ 6 over failure-prone tools. From 2,095 available tools, the pipeline identifies a failure tool set of size 1,204 and uses legality-constrained dependency traversal to synthesize hard traces. Each trace is then abstracted by a Tool Maker agent into a modular advanced tool with a JSON schema, and a Hard-Query agent formulates a user request that requires the abstract tool while omitting intermediate steps. This couples hardness generation to hidden compositional structure rather than to superficial perturbation (Hao et al., 4 Jan 2026).

Feature-space HSGL methods use several distinct synthesis operators. Under instance-dependent noisy labels, anchor hallucination samples easy features $G$ 7 and $G$ 8 from different classes, concatenates them, and maps them through a hallucinator $G$ 9 to produce a synthetic anchor $\epsilon$ 0 near a decision boundary. In two-stage hard-sample generation for deep metric learning, Stage 1 stretches an anchor-positive pair by piecewise linear manipulation and then refines it with a conditional GAN to obtain a hard positive; Stage 2 generates a hard negative through an adaptive reverse metric constraint. In SSCL for contrastive learning, hard negatives are synthesized directly in feature space by random linear interpolation between the top- $\epsilon$ 1 hardest negatives. In SSAH for deep hashing, hard samples are generated by a dedicated adversarial network consisting of a Multi-Angle Rotation Network and a Multi-Scale Mask Network, yielding occlusions and rotated deformations that compete against robust hash-code learning (Huang et al., 2024, Zhu et al., 2021, Dong et al., 2023, Jin et al., 2019).

3. Learning objectives and optimization strategies

HiSampler optimizes the expected hardness of sampled graphs. With $\epsilon$ 2, the objective is

$\epsilon$ 3

and the method uses the REINFORCE estimator

$\epsilon$ 4

A single Monte Carlo sample induces a policy-gradient update, and the reported optimization uses Adam with learning rate $\epsilon$ 5. No additional regularizer is used (Sato et al., 2019).

In HSGL for CC-ReID, learning is hardness-aware rather than purely generative. The Hard Sample Analyzer computes batchwise boolean matrices $\epsilon$ 6 and $\epsilon$ 7, and Hard Sample Distance Adjustment rescales distances for true hard pairs through

$\epsilon$ 8

The overall loss combines cross-entropy with an aggregated triplet term $\epsilon$ 9, with $G$ 0 typically in the range $G$ 1– $G$ 2 and $G$ 3 for non-metric baselines (Liu et al., 15 Jul 2025).

The noisy-label framework couples a hallucination objective with semi-supervised training. The hallucinator is trained by

$G$ 4

where $G$ 5 encourages an anchor to lie between easy samples from two classes while $G$ 6 preserves the pseudo-label of the source easy sample. Corrected hard samples are then merged with easy samples into a labeled set, while uncorrected hard samples are treated as unlabeled. Training alternates between a classification phase that optimizes a MixMatch-style loss and a hallucinator phase that optimizes $G$ 7 (Huang et al., 2024).

Two-stage hard-sample generation for metric learning integrates generative and discriminative losses throughout. Stage 1 uses label-consistency, conditional adversarial, and reconstruction losses for the hard-positive generator. Stage 2 introduces the Adaptive Reverse Triplet loss

$G$ 8

with an increasing margin $G$ 9, and the final embedding loss combines original triplet, generated-triplet, and category terms with adaptive weighting (Zhu et al., 2021).

SSCL changes the contrastive denominator itself. Hard negatives are synthesized by mixing hard negatives, weighted by similarity to the anchor, and then debiased to compensate for false negatives using an estimate of the false-negative rate $L$ 0. SSAH, by contrast, uses alternating optimization between a hard-sample generator and a hashing network, with adversarial, semantic, consistency, and quantization terms, and a self-paced margin schedule $L$ 1 that gradually increases hardness (Dong et al., 2023, Jin et al., 2019).

HardGen closes the loop between generation and verification. A Reasoner agent produces Chain-of-Thought and function calls, a Verifier compares them to ground-truth calls, and corrective hints are iteratively appended to the prompt, with $L$ 2 in practice. In the RL variant, the reward is binary over the entire trajectory, and the objective is

$L$ 3

with KL coefficient $L$ 4 (Hao et al., 4 Jan 2026).

4. Empirical performance across domains

Reported results show that HSGL can improve either the hardness of generated instances or the downstream robustness of learned representations, but the effect size depends strongly on how generation and learning are coupled (Sato et al., 2019, Liu et al., 15 Jul 2025, Hao et al., 4 Jan 2026, Huang et al., 2024, Zhu et al., 2021, Dong et al., 2023, Jin et al., 2019).

Domain / paper	Setting	Reported result
Graph algorithms	DSATUR, $L$ 5, $L$ 6	Random ER 407 calls; Generic GA 2,464; HiSampler-vanilla $L$ 7; HiSampler-PER $L$ 8
CC-ReID	PRCC	HSGL + FIRe²: 68.9% Rank-1 / 63.4% mAP; baseline IFD: 65.3% / 61.7%
Tool-use agents	BFCLv3 overall	HardGen-4B-RL 79.14%; Claude-Opus-4.5 78.92%; Gemini-3-Pro 78.17%; GPT-5.2 60.12%
Noisy-label learning	CIFAR-10, classification-based noise 40%	Ours 92.47; TSCSI 84.18; SEAL 82.98
Contrastive learning	TinyImageNet linear eval	SimCLR 25.22% $L$ 9 SimCLR-SSCL 37.31%
Metric learning	Cars196, R@1	OURS 82.1; SS 69.7; HDML 61.0
Deep hashing	CIFAR-10/NUS-WIDE; CUB/Stanford Dogs	mAP gain of 3–4% over SSGAH; 4–6% absolute boost on fine-grained datasets

For graph algorithms, the paper reports that all seven algorithms across four problems improve by “a few to several orders of magnitude” over random and generic evolutionary baselines, and that HiSampler-PER strictly outperforms the vanilla version in almost every case. HiSampler also outperforms rule-based algorithms on the 3-coloring problem, including the Cheeseman–Hogg, Vlasie, and Mizuno–Nishihara constructions (Sato et al., 2019).

In CC-ReID, the strongest ablation result is the joint use of SHSG and HSAL: SHSG+HSAL yields $\mathrm{IS\_HP}_{ij}$ 0 and $\mathrm{IS\_HP}_{ij}$ 1 Rank-1 on CAL and AIM, while the full pipeline gives $\mathrm{IS\_HP}_{ij}$ 2 Rank-1 and $\mathrm{IS\_HP}_{ij}$ 3 mAP on FIRe² for PRCC. HSAL also changes optimization dynamics: on PRCC the model reaches baseline accuracy by epoch 5–10 rather than epoch 50–60, which the paper describes as a $\mathrm{IS\_HP}_{ij}$ 4– $\mathrm{IS\_HP}_{ij}$ 5 speedup in converging to the same performance (Liu et al., 15 Jul 2025).

HardGen reports strong gains on the most difficult multi-turn setting. On BFCLv3, HardGen-4B-RL scores 63.13% on multi-turn tasks versus 28.13% for GPT-5.2, and the overall score of 79.14% exceeds the listed scores for GPT-5.2, Gemini-3-Pro, and Claude-Opus-4.5. The same summary reports generalization to BFCLv4, API-Bank, ACEBench, and other base models (Hao et al., 4 Jan 2026).

The noisy-label method improves top-1 accuracy across synthetic and real-world instance-dependent-noise settings, though Clothing1M is a partial exception: the reported 74.62 for the proposed method exceeds DivideMix at 74.40 but is below TSCSI at 75.40. In contrastive learning, the full SSCL design outperforms variants that use only synthesis, only sampling, or only synthesis plus debiasing, indicating that the reported gains are not produced by synthetic negatives alone (Huang et al., 2024, Dong et al., 2023).

5. Structural insights and recurring patterns

HiSampler is one of the clearest demonstrations that HSGL can recover interpretable structure rather than a single adversarial artifact. For DSATUR, a hard instance contains a hidden 4-clique attached via a long path to the rest of the graph; the algorithm spends billions of recursive calls coloring the remainder before discovering that the final 4-clique is uncolorable. Frequent subgraph mining on 1,000 HiSampler samples identifies 4-cliques with support $\mathrm{IS\_HP}_{ij}$ 6 and bottleneck-path motifs with support $\mathrm{IS\_HP}_{ij}$ 7. A diversity check further shows that even graphs with Jaccard similarity at most 0.7 to the single hardest instance retain average hardness of approximately $\mathrm{IS\_HP}_{ij}$ 8 and maximum hardness of approximately $\mathrm{IS\_HP}_{ij}$ 9, indicating that the learned object is a distribution rather than a memorized instance (Sato et al., 2019).

Several later works impose a similar separation between semantic validity and difficulty. In the noisy-label setting, high loss is not treated as evidence of label corruption; hard clean samples are explicitly recovered by hallucinated anchors and majority-vote relabeling. In CC-ReID, hard positives and hard negatives are defined from identity, clothing, and viewpoint metadata rather than inferred solely from feature distance. In SSAH, hard generation is constrained by semantic and consistency objectives so that adversarial masks and rotations remain compatible with retrieval supervision (Huang et al., 2024, Liu et al., 15 Jul 2025, Jin et al., 2019).

This suggests a common HSGL template: first formalize difficulty with respect to the target task, then generate or select candidates under semantic or structural constraints, and finally train with an objective that preserves those constraints while amplifying informative error signals. HardGen makes this pattern especially explicit by requiring that hard queries remain executable and that their reasoning be verifiable by a separate agent. The same theme appears, in a different mathematical language, in the Ising-model results: generating fresh samples is not reducible to merely identifying model parameters, and hardness can be inherent to the sampling problem itself (Hao et al., 4 Jan 2026, Risteski et al., 23 May 2026).

6. Limitations, misconceptions, and open directions

A persistent misconception is that adding harder samples is uniformly beneficial. The reported ablations contradict that view. In CC-ReID, CHPSG alone yields minor or even negative gains for methods that depend on true clothing labels, SHSG alone can hurt if used without HSAL, and HSAL alone produces only $\mathrm{IS\_HN}_{ij}$ 0 to $\mathrm{IS\_HN}_{ij}$ 1 gains. In SSCL, synthetic negatives alone provide marginal or negative gains, while the best results require the combination of synthesis, hardness-aware weighting, and debiasing. These results indicate that hard-sample generation and hard-sample learning are not interchangeable modules; their coupling is often the main source of improvement (Liu et al., 15 Jul 2025, Dong et al., 2023).

There are also domain-specific structural limitations. HiSampler does not enforce graph isomorphism invariance, depends on a fixed $\mathrm{IS\_HN}_{ij}$ 2, and may be limited by MLP capacity as $\mathrm{IS\_HN}_{ij}$ 3 grows because the output dimension scales as approximately $\mathrm{IS\_HN}_{ij}$ 4. The noisy-label method relies on a sufficiently large, class-balanced easy set and may underperform on heavily imbalanced datasets. SSCL remains sensitive to the estimate of the false-negative rate $\mathrm{IS\_HN}_{ij}$ 5, and overmixing can degrade performance. These are not merely engineering caveats; they reveal that HSGL depends critically on the quality of its hardness surrogate and on the validity constraints imposed during synthesis (Sato et al., 2019, Huang et al., 2024, Dong et al., 2023).

Theoretical work on Ising models sharpens these concerns into a hardness boundary. For a family of bounded-width Ising models with spectral width just above $\mathrm{IS\_HN}_{ij}$ 6, learning-to-sample is computationally hard under the Computational Diffie–Hellman assumption even when the learner receives both polynomially many i.i.d. samples and exact access to $\mathrm{IS\_HN}_{ij}$ 7. The same work establishes a memorization–hallucination dichotomy: any efficient learner must either output configurations that match transformed training data under a map $\mathrm{IS\_HN}_{ij}$ 8 or assign substantial probability to configurations of negligible mass under the target distribution. This result separates parameter learning from sampler learning and shows that, in some settings, hard-sample generation cannot be solved by better optimization alone (Risteski et al., 23 May 2026).

The extension directions proposed in these papers are correspondingly diverse: richer graph generators such as GNNs, GANs, or autoregressive models; permutation-equivariant architectures; new hardness criteria such as approximation ratio, maximum delay, or memory usage; adaptation to detection, segmentation, domain adaptation, and few-shot learning; learned mixing coefficients for contrastive negatives; and cross-modal hard-sample generation. Taken together, these proposals indicate that HSGL has evolved from a narrow mechanism for mining difficult examples into a general methodology for shaping training and evaluation distributions around task-relevant failure modes (Sato et al., 2019, Huang et al., 2024, Dong et al., 2023).