CAL: Contrastive Alignment Overview

• Contrastive Alignment (CAL) is a training strategy that uses contrastive objectives to pull matching entities together while pushing mismatches apart across various modalities.
• CAL techniques are applied in diverse settings such as unsupervised NLP, vision–language tuning, audio pretraining, protein retrieval, and human preference alignment.
• CAL improves correspondence and grounding at multiple levels (token, instance, distribution) but requires careful tuning of weights and targets to maintain fluency and stability.

Contrastive Alignment (CAL) denotes a family of training strategies in which matched entities are pulled together and mismatched entities are pushed apart so that a model learns a more faithful correspondence between modalities, structures, or supervision signals. In the cited literature, the label is reused across several technically distinct settings: unsupervised word alignment, vision–language instruction tuning, token re-weighted autoregressive training, object-centric diffusion, audio–language pretraining, protein sequence–structure retrieval, human-preference alignment, debiasing, and optimal-transport formulations of contrastive learning [(Liu et al., 2014); (Xiao et al., 2024); (Nguyen et al., 3 Jan 2026); (Sachidananda et al., 2022); (Ranganath et al., 6 Mar 2026); (Fang et al., 2024); (Chen et al., 27 Feb 2025)]. This suggests that CAL is not a single canonical algorithm, but a recurrent design pattern for enforcing correspondence under contrastive supervision.

1. Terminological scope and historical spread

Within the provided literature, the term covers multiple non-equivalent mechanisms. The common thread is not a fixed loss function, but the use of contrastive evidence to improve alignment quality at the instance, token, slot, phrase, or distribution level.

Domain	What is aligned	Representative paper
Unsupervised NLP	observed sentence pairs against noisy shuffled pairs	"Contrastive Unsupervised Word Alignment with Non-Local Features" (Liu et al., 2014)
Vision–language AR training	visually correlated text tokens and image evidence	"Seeing the Image: Prioritizing Visual Correlation by Contrastive Alignment" (Xiao et al., 2024)
Object-centric diffusion	semantic slots and the correct image	"Improved Object-Centric Diffusion Learning with Registers and Contrastive Alignment" (Nguyen et al., 3 Jan 2026)
Human alignment	preferred and rejected responses under ranked feedback	"CLHA: A Simple yet Effective Contrastive Learning Framework for Human Alignment" (Fang et al., 2024)
Audio–language pretraining	acoustic tokens and lexical embeddings	"CALM: Contrastive Aligned Audio-Language Multirate and Multimodal Representations" (Sachidananda et al., 2022)
Distribution alignment theory	transport plans and target correspondence structure	"Your contrastive learning problem is secretly a distribution alignment problem" (Chen et al., 27 Feb 2025)

The historical range in the provided sources begins with unsupervised word alignment in 2014, where the objective maximizes a log-ratio between observed sentence pairs and noisy counterparts and approximates posterior expectations with top-$n$ alignments; in practice, $n=1$ sufficed and the method reported test AER $5.01$ on French–English and $20.24$ on Chinese–English (Liu et al., 2014). Later work broadens CAL to multimodal grounding, generative modeling, and alignment of LMs to human preferences.

A nearby but distinct usage appears in LiDAR–camera calibration: "Cal or No Cal?" shifts from regressing calibration parameters to binary classification of calibrated versus miscalibrated sensor states using a contrastive learning approach over latent embeddings, but the provided material does not define this method as a CAL objective in the same sense as the alignment literature (Tahiraj et al., 31 Mar 2025).

2. Recurrent mathematical constructions

A large subset of CAL methods retains the familiar contrastive structure of a similarity score followed by a discrimination loss. In CG-VLM, projected patch features are average-pooled to a global image descriptor $\hat v$, compared to all token embeddings in a caption through an averaged scaled cosine similarity, and optimized with an InfoNCE loss: $$s^{\,i,j} = \frac{1}{M}\sum_{m=1}^M \tau\,\frac{\bigl(\hat v^i\bigr)^\top e^j_m} {\|\hat v^i\|_2\,\|e^j_m\|_2}, \qquad \mathcal L_{\rm align}^{\rm con} = \frac{1}{B}\sum_{i=1}^B \Bigl[ -\log \frac {\exp\!\bigl(s^{\,i,i}\bigr)} {\exp\!\bigl(s^{\,i,i}\bigr)+\sum_{j\neq i}\exp\!\bigl(s^{\,i,j}\bigr)} \Bigr].$$ The same paper combines this with caption-generation loss as $$\mathcal L_{\rm align}^{\rm CG}=\mathcal L_{\rm align}^{\rm gen}+\alpha \mathcal L_{\rm align}^{\rm con}$$, with $\alpha=1.0$ by default (Liu et al., 2023). ProtAlign uses a CLIP-style symmetric InfoNCE over sequence and structure embeddings, with similarity $s_{ij}=(\mathbf P_i\cdot \mathbf S_j)/\tau$ and best reported temperature $\tau=0.07$ (Ranganath et al., 6 Mar 2026).

Other CAL variants depart from pairwise softmax classification and instead reweight or invert a base objective. In the vision–language CAL of "Seeing the Image," the core signal is a tokenwise logit difference between a forward pass with image and a forward pass without image,

$n=1$0

which is clamped, optionally pooled, and then used to reweight autoregressive cross-entropy. Tokens with large $n=1$1 are treated as visually correlated, whereas small or negative $n=1$2 indicates visually irrelevant or contradictory content (Xiao et al., 2024).

Object-centric diffusion introduces a different construction. CODA retains the usual denoising loss

$n=1$3

and adds a negative-slot reconstruction term

$n=1$4

so that the full objective becomes $n=1$5. In the paper’s interpretation, $n=1$6 serves as a tractable surrogate for maximizing mutual information between slots and inputs (Nguyen et al., 3 Jan 2026).

A more abstract formulation appears in generalized contrastive alignment, where CAL is written as a bilevel distribution-alignment problem: $n=1$7 In this view, InfoNCE becomes a one-step row-wise Sinkhorn projection toward an identity target plan, and unbalanced OT extends the framework to noisy or missing views (Chen et al., 27 Feb 2025).

3. Vision–language grounding and compositionality

In autoregressive VLMs, CAL is frequently used to correct the mismatch between uniform token-level supervision and heterogeneous visual relevance. The token-reweighting CAL of "Seeing the Image" is explicitly motivated by the claim that standard autoregressive vision–language training treats each text token equally, thereby over-emphasizing tokens that are less correlated with or even contradictory to the image. Across LLaVA-1.5, LLaVA-NeXT, and Mini-Gemini variants, the paper reports that CAL wins $n=1$8–$n=1$9 out of $5.01$0 sub-benchmarks; for LLaVA-NeXT-13B the reported gains are $5.01$1 ANLS on VQA$5.01$2, $5.01$3 relaxed accuracy on VQA$5.01$4, $5.01$5 CIDEr on COCO Caption, $5.01$6 CIDEr on TextCaps, and $5.01$7 IoU on RefCOCOg val/test. The extra cost is one additional text-only forward per training step, giving approximately $5.01$8 slower instruction tuning and no extra cost at inference time (Xiao et al., 2024).

CG-VLM uses CAL at the alignment stage between a frozen CLIP ViT-L/14 and a frozen LLM, optimizing only a trainable visual adapter and temperature parameter before full visual instruction tuning. The reported effect is improved zero-shot grounding and greater data efficiency: ScienceQA-Image accuracy rises from $5.01$9 to $20.24$0, ScienceQA-Text from $20.24$1 to $20.24$2, and POPE from $20.24$3 to $20.24$4; with only $20.24$5 of the full instruction dataset, CG-VLM recovers $20.24$6 of LLaVA’s full-data performance on ScienceQA-Image (Liu et al., 2023).

A separate line of work extends CAL from tokenwise relevance to hierarchical and compositional grounding. $20.24$7-CLIP treats caption-, sentence-, and phrase-level queries from the same image as positives with distinct weights controlled by $20.24$8, yielding two variants: $20.24$9-CAL-CE and $\hat v$0-CAL-BCE. The abstract reports $\hat v$1 T2I and $\hat v$2 I2T at R@1 on Urban1K and $\hat v$3 on FG-OVD (Hard), while the ablation indicates that for the CE variant, $\hat v$4–$\hat v$5 is a favorable trade-off region (Zohra et al., 14 Dec 2025). PowerCLIP pushes this further by aligning every non-empty subset of $\hat v$6 image regions with phrases from a constituency parse tree, then replacing the naive $\hat v$7 powerset construction with non-linear aggregators that reduce complexity to $\hat v$8. On CC12M pretraining and downstream zero-shot evaluation, it reports classification average Top-1 $\hat v$9 versus CLIP $$s^{\,i,j} = \frac{1}{M}\sum_{m=1}^M \tau\,\frac{\bigl(\hat v^i\bigr)^\top e^j_m} {\|\hat v^i\|_2\,\|e^j_m\|_2}, \qquad \mathcal L_{\rm align}^{\rm con} = \frac{1}{B}\sum_{i=1}^B \Bigl[ -\log \frac {\exp\!\bigl(s^{\,i,i}\bigr)} {\exp\!\bigl(s^{\,i,i}\bigr)+\sum_{j\neq i}\exp\!\bigl(s^{\,i,j}\bigr)} \Bigr].$$0, retrieval average R@1 $$s^{\,i,j} = \frac{1}{M}\sum_{m=1}^M \tau\,\frac{\bigl(\hat v^i\bigr)^\top e^j_m} {\|\hat v^i\|_2\,\|e^j_m\|_2}, \qquad \mathcal L_{\rm align}^{\rm con} = \frac{1}{B}\sum_{i=1}^B \Bigl[ -\log \frac {\exp\!\bigl(s^{\,i,i}\bigr)} {\exp\!\bigl(s^{\,i,i}\bigr)+\sum_{j\neq i}\exp\!\bigl(s^{\,i,j}\bigr)} \Bigr].$$1 versus $$s^{\,i,j} = \frac{1}{M}\sum_{m=1}^M \tau\,\frac{\bigl(\hat v^i\bigr)^\top e^j_m} {\|\hat v^i\|_2\,\|e^j_m\|_2}, \qquad \mathcal L_{\rm align}^{\rm con} = \frac{1}{B}\sum_{i=1}^B \Bigl[ -\log \frac {\exp\!\bigl(s^{\,i,i}\bigr)} {\exp\!\bigl(s^{\,i,i}\bigr)+\sum_{j\neq i}\exp\!\bigl(s^{\,i,j}\bigr)} \Bigr].$$2, and Winoground image retrieval $$s^{\,i,j} = \frac{1}{M}\sum_{m=1}^M \tau\,\frac{\bigl(\hat v^i\bigr)^\top e^j_m} {\|\hat v^i\|_2\,\|e^j_m\|_2}, \qquad \mathcal L_{\rm align}^{\rm con} = \frac{1}{B}\sum_{i=1}^B \Bigl[ -\log \frac {\exp\!\bigl(s^{\,i,i}\bigr)} {\exp\!\bigl(s^{\,i,i}\bigr)+\sum_{j\neq i}\exp\!\bigl(s^{\,i,j}\bigr)} \Bigr].$$3 versus $$s^{\,i,j} = \frac{1}{M}\sum_{m=1}^M \tau\,\frac{\bigl(\hat v^i\bigr)^\top e^j_m} {\|\hat v^i\|_2\,\|e^j_m\|_2}, \qquad \mathcal L_{\rm align}^{\rm con} = \frac{1}{B}\sum_{i=1}^B \Bigl[ -\log \frac {\exp\!\bigl(s^{\,i,i}\bigr)} {\exp\!\bigl(s^{\,i,i}\bigr)+\sum_{j\neq i}\exp\!\bigl(s^{\,i,j}\bigr)} \Bigr].$$4 (Kawamura et al., 28 Nov 2025).

4. Generative and object-centric variants

In object-centric learning with diffusion models, CAL is used to tighten the relation between latent object slots and image content. CODA augments Slot Attention over frozen DINOv2 ViT-B/14 features with register slots derived from $$s^{\,i,j} = \frac{1}{M}\sum_{m=1}^M \tau\,\frac{\bigl(\hat v^i\bigr)^\top e^j_m} {\|\hat v^i\|_2\,\|e^j_m\|_2}, \qquad \mathcal L_{\rm align}^{\rm con} = \frac{1}{B}\sum_{i=1}^B \Bigl[ -\log \frac {\exp\!\bigl(s^{\,i,i}\bigr)} {\exp\!\bigl(s^{\,i,i}\bigr)+\sum_{j\neq i}\exp\!\bigl(s^{\,i,j}\bigr)} \Bigr].$$5 CLIP padding tokens passed through the frozen Stable Diffusion text encoder, while freezing the pretrained Stable Diffusion v1.5 U-Net except for the key, value, and output projections in cross-attention. Negative slot sets are created by mixing half of the slots from the current image with half from a different image, and the VOC ablation reports that replacing $$s^{\,i,j} = \frac{1}{M}\sum_{m=1}^M \tau\,\frac{\bigl(\hat v^i\bigr)^\top e^j_m} {\|\hat v^i\|_2\,\|e^j_m\|_2}, \qquad \mathcal L_{\rm align}^{\rm con} = \frac{1}{B}\sum_{i=1}^B \Bigl[ -\log \frac {\exp\!\bigl(s^{\,i,i}\bigr)} {\exp\!\bigl(s^{\,i,i}\bigr)+\sum_{j\neq i}\exp\!\bigl(s^{\,i,j}\bigr)} \Bigr].$$6 of the slots works best among $$s^{\,i,j} = \frac{1}{M}\sum_{m=1}^M \tau\,\frac{\bigl(\hat v^i\bigr)^\top e^j_m} {\|\hat v^i\|_2\,\|e^j_m\|_2}, \qquad \mathcal L_{\rm align}^{\rm con} = \frac{1}{B}\sum_{i=1}^B \Bigl[ -\log \frac {\exp\!\bigl(s^{\,i,i}\bigr)} {\exp\!\bigl(s^{\,i,i}\bigr)+\sum_{j\neq i}\exp\!\bigl(s^{\,i,j}\bigr)} \Bigr].$$7. On VOC, FG-ARI rises from $$s^{\,i,j} = \frac{1}{M}\sum_{m=1}^M \tau\,\frac{\bigl(\hat v^i\bigr)^\top e^j_m} {\|\hat v^i\|_2\,\|e^j_m\|_2}, \qquad \mathcal L_{\rm align}^{\rm con} = \frac{1}{B}\sum_{i=1}^B \Bigl[ -\log \frac {\exp\!\bigl(s^{\,i,i}\bigr)} {\exp\!\bigl(s^{\,i,i}\bigr)+\sum_{j\neq i}\exp\!\bigl(s^{\,i,j}\bigr)} \Bigr].$$8 for $$s^{\,i,j} = \frac{1}{M}\sum_{m=1}^M \tau\,\frac{\bigl(\hat v^i\bigr)^\top e^j_m} {\|\hat v^i\|_2\,\|e^j_m\|_2}, \qquad \mathcal L_{\rm align}^{\rm con} = \frac{1}{B}\sum_{i=1}^B \Bigl[ -\log \frac {\exp\!\bigl(s^{\,i,i}\bigr)} {\exp\!\bigl(s^{\,i,i}\bigr)+\sum_{j\neq i}\exp\!\bigl(s^{\,i,j}\bigr)} \Bigr].$$9CA$$\mathcal L_{\rm align}^{\rm CG}=\mathcal L_{\rm align}^{\rm gen}+\alpha \mathcal L_{\rm align}^{\rm con}$$0Reg without contrastive loss to $$\mathcal L_{\rm align}^{\rm CG}=\mathcal L_{\rm align}^{\rm gen}+\alpha \mathcal L_{\rm align}^{\rm con}$$1 for the full model; on COCO, FG-ARI rises from $$\mathcal L_{\rm align}^{\rm CG}=\mathcal L_{\rm align}^{\rm gen}+\alpha \mathcal L_{\rm align}^{\rm con}$$2 to $$\mathcal L_{\rm align}^{\rm CG}=\mathcal L_{\rm align}^{\rm gen}+\alpha \mathcal L_{\rm align}^{\rm con}$$3. The paper also reports that unfreezing the U-Net when applying $$\mathcal L_{\rm align}^{\rm CG}=\mathcal L_{\rm align}^{\rm gen}+\alpha \mathcal L_{\rm align}^{\rm con}$$4 makes training unstable and collapses segmentation and generation metrics to approximately FG-ARI $$\mathcal L_{\rm align}^{\rm CG}=\mathcal L_{\rm align}^{\rm gen}+\alpha \mathcal L_{\rm align}^{\rm con}$$5 (Nguyen et al., 3 Jan 2026).

PuLID employs a different sense of contrastive alignment in text-to-image identity customization. Rather than using InfoNCE, it aligns two denoising trajectories run from identical noise and prompt: one with prompt only and one with prompt plus ID features. The semantic term compares prompt-conditioned feature responses,

$$\mathcal L_{\rm align}^{\rm CG}=\mathcal L_{\rm align}^{\rm gen}+\alpha \mathcal L_{\rm align}^{\rm con}$$6

and the layout term penalizes direct feature drift,

$$\mathcal L_{\rm align}^{\rm CG}=\mathcal L_{\rm align}^{\rm gen}+\alpha \mathcal L_{\rm align}^{\rm con}$$7

The combined alignment loss uses $$\mathcal L_{\rm align}^{\rm CG}=\mathcal L_{\rm align}^{\rm gen}+\alpha \mathcal L_{\rm align}^{\rm con}$$8 and $$\mathcal L_{\rm align}^{\rm CG}=\mathcal L_{\rm align}^{\rm gen}+\alpha \mathcal L_{\rm align}^{\rm con}$$9. The paper states that without $\alpha=1.0$0, the ID adapter collapses layout and prompt-editing ability, whereas with $\alpha=1.0$1 it preserves style, layout, and editability. In the reported staged ablation, ID-loss alone raises cosine similarity on DivID-120 from $\alpha=1.0$2 to $\alpha=1.0$3, while adding CAL lowers it to approximately $\alpha=1.0$4 but restores prompt editing qualitatively (Guo et al., 2024).

Condition Contrastive Alignment (CCA) adapts the same contrastive principle to autoregressive visual generation as a replacement for classifier-free guidance. It defines a residual $\alpha=1.0$5 through a logistic discrimination problem over matched and shuffled condition pairs, then parameterizes that residual as the log-probability gap between a fine-tuned model $\alpha=1.0$6 and a frozen pretrained model $\alpha=1.0$7. Fine-tuning for one epoch, described as approximately $\alpha=1.0$8 of the original pretraining compute, is reported to bring guidance-free sampling close to or on par with guided sampling while cutting sampling cost roughly in half. On ImageNet $\alpha=1.0$9, LlamaGen-3B improves from FID $s_{ij}=(\mathbf P_i\cdot \mathbf S_j)/\tau$0 to $s_{ij}=(\mathbf P_i\cdot \mathbf S_j)/\tau$1 and IS $s_{ij}=(\mathbf P_i\cdot \mathbf S_j)/\tau$2 to $s_{ij}=(\mathbf P_i\cdot \mathbf S_j)/\tau$3, while VAR-d30 improves from FID $s_{ij}=(\mathbf P_i\cdot \mathbf S_j)/\tau$4 to $s_{ij}=(\mathbf P_i\cdot \mathbf S_j)/\tau$5 and IS $s_{ij}=(\mathbf P_i\cdot \mathbf S_j)/\tau$6 to $s_{ij}=(\mathbf P_i\cdot \mathbf S_j)/\tau$7 (Chen et al., 2024).

5. Audio, protein, and other cross-modal extensions

CALM applies contrastive alignment to audio and lexical inputs through a three-stage pipeline: a Spectral Transformer over spectrogram patches, contrastive acoustic–language pretraining (CALP), and a multimodal transformer trained with masked language modeling and masked audio modeling. CALP aligns each acoustic token both to the next acoustic token and to the utterance-level transcript embedding through a weighted NT-Xent objective with $s_{ij}=(\mathbf P_i\cdot \mathbf S_j)/\tau$8 and $s_{ij}=(\mathbf P_i\cdot \mathbf S_j)/\tau$9. The reported training cost for CALP is $\tau=0.07$0 hours on $\tau=0.07$1V100 GPUs. On CMU-MOSEI, CALM$\tau=0.07$2 reports $\tau=0.07$3 weighted accuracy versus a previous best of approximately $\tau=0.07$4; on MSP-Podcasts it reports $\tau=0.07$5 WA and $\tau=0.07$6 $\tau=0.07$7 versus approximately $\tau=0.07$8 WA and $\tau=0.07$9 $n=1$00 (Sachidananda et al., 2022).

Multi-grained contrastive language–audio pretraining extends this idea with a shared codebook $n=1$01, Sparsemax-based global summaries, a locality-aware encoder block, and a hard-negative guided loss. The paper states that the shared codebook bridges the gap between frame and word features by forcing both modalities to use common bases, while the locality-aware block preserves frame-level patterns by removing the final self-attention aggregation step. On zero-shot evaluation, the reported gains over CLAP include AudioCaps R@1 from $n=1$02 to $n=1$03 for text-to-audio and from $n=1$04 to $n=1$05 for audio-to-text, VGGSound accuracy from $n=1$06 to $n=1$07, DESED PSDS$n=1$08 from $n=1$09 to $n=1$10, and TAG PSDS$n=1$11 from $n=1$12 to $n=1$13 (Li et al., 2024).

ProtAlign moves CAL into protein representation learning by aligning ESM2 sequence embeddings with Protein-MPNN structure embeddings in a shared $n=1$14 space using a single multi-head self-attention projection with learned query tokens. The main objective is a symmetric InfoNCE loss over all sequence–structure pairs in a batch of size $n=1$15, trained on PDBBind 2020 with $n=1$16 train pairs, $n=1$17 validation pairs, and the CASF-2016 core $n=1$18-pair test set. With CLIP loss and $n=1$19, the reported retrieval scores are Recall@1 $n=1$20 and Recall@5 $n=1$21, compared with SigLIP’s $n=1$22 and $n=1$23. The paper further reports that post-training similarity matrices show strong diagonal dominance and that t-SNE clusters combine both modalities in neighborhoods often grouping homologous proteins (Ranganath et al., 6 Mar 2026).

6. Human preference alignment, theory, and open issues

In LLM alignment, CAL has been used as a direct alternative to PPO-style RLHF. CLHA begins from a pretrained LLaMA-7B, scores ranked responses with an auxiliary reward model $n=1$24, filters or down-weights pairs with low reward separation, and combines a pairwise contrastive term with adaptive supervised fine-tuning: $n=1$25 Training is reported for two epochs with learning rate $n=1$26. On the Helpful and Harmless benchmark, CLHA$n=1$27 reports BLEU $n=1$28 and Reward $n=1$29, compared with PRO$n=1$30 at BLEU $n=1$31 and Reward $n=1$32; in human evaluation over $n=1$33 samples, CLHA is preferred in $n=1$34 of comparisons, PRO in $n=1$35, and $n=1$36 are ties (Fang et al., 2024).

A related use appears in debiasing under the heading of the “alignment tax,” defined as the degradation in factual accuracy, knowledge retention, or generation coherence that often accompanies debiasing interventions. The CAL framework in this setting constructs faithful positives by backtranslation through German, French, and Spanish with beam width $n=1$37 and temperature $n=1$38, and biased or unfaithful negatives through adversarial toxic generation, inverse beam search, and entity manipulation. It applies a token-level projection head, pools over named entities, and optimizes

$n=1$39

with default $n=1$40, while scaling the contrastive loss by $n=1$41 whenever a toxicity classifier score exceeds $n=1$42. On Reddit TL;DR, the reported deltas are $n=1$43 toxicity and $n=1$44 faithfulness for GPT-2, $n=1$45 and $n=1$46 for Phi2, and $n=1$47 and $n=1$48 for Llama2-7B; the paper states that CAL is the first method in its comparison to improve both metrics simultaneously across all scales (Korkmaz et al., 25 May 2025).

Theoretical work broadens the meaning of CAL still further. Generalized contrastive alignment reframes contrastive learning itself as an entropic optimal-transport problem, showing that InfoNCE is a one-step projection toward an identity transport plan and that multistep or unbalanced variants improve alignment and uniformity while accommodating noisy views. The paper reports that $n=1$49GCA variants of InfoNCE and RINCE improve accuracy by roughly $n=1$50–$n=1$51, and that GCA-UOT can outperform baselines by up to $n=1$52 on CIFAR-10/100, SVHN, and ImageNet-100 under standard and extreme augmentations (Chen et al., 27 Feb 2025). A complementary theoretical notion of contrastive alignment studies the relation between self-supervised CL and negatives-only supervised contrastive learning (NSCL): the main result is that similarity matrices remain close, yielding high-probability CKA and RSA guarantees, even though parameter-space coupling can diverge exponentially with training time. Empirically, the paper reports CL–NSCL CKA/RSA around $n=1$53–$n=1$54 for the first $n=1$55k epochs on CIFAR-10/100, Mini-ImageNet, and Tiny-ImageNet, while alignment strengthens with more classes and higher temperatures (Luthra et al., 9 Oct 2025).

Several limitations recur across these otherwise disparate CAL formulations. Reported sensitivities include empirically chosen clamp bounds $n=1$56 in token-reweighting CAL, dependence on reward-model quality in CLHA, instability when the diffusion backbone is unfrozen in CODA, and additional computation from exhaustive or approximate fine-grained matching in powerset alignment (Xiao et al., 2024, Fang et al., 2024, Nguyen et al., 3 Jan 2026, Kawamura et al., 28 Nov 2025). A plausible implication is that the central challenge in CAL is not defining positives and negatives in the abstract, but engineering the alignment target, weighting scheme, and optimization constraints so that contrast improves correspondence without damaging fluency, diversity, reconstruction fidelity, or robustness.