Bilateral Cross-Prompt Coordination

• Bilateral Cross-Prompt Coordination is a strategy where two prompt-bearing entities are jointly optimized to enable bidirectional transfer and robust task performance.
• It employs mechanisms like bidirectional projection, co-attention, and shared objectives to integrate insights from modalities such as vision, language, and essay scoring.
• This approach addresses distribution shifts and alignment challenges across applications including vision–language models, LLM migration, and multi-modal verification.

Searching arXiv for the cited papers and closely related terms to ground the article in current papers. arXiv search query: "bilateral cross-prompt coordination prompt transfer cross-modal prompting" arXiv search query: "ChordPrompt CLIP multi-domain incremental learning (Wang et al., 24 Jun 2025)" Bilateral Cross-Prompt Coordination denotes a class of prompt-learning and prompt-transfer strategies in which two prompt-bearing entities are optimized, exchanged, or conditioned jointly rather than in isolation. The paired entities vary by problem setting: visual and textual prompts inside a frozen vision–LLM, source- and target-model prompts during LLM migration, system and user prompts in multi-component prompting, prompt-conditioned representations across writing prompts in automated essay scoring, or bilingual prompt behavior across languages. Across these literatures, the common objective is to preserve or improve task performance under distribution shift while maintaining some form of mutual dependence between the two sides, such as bidirectional conditioning, symmetric transfer, coordinated retrieval, or shared optimization (Wang et al., 24 Jun 2025, Wang et al., 1 Dec 2025, Zhang et al., 21 Jul 2025).

1. Conceptual scope and defining criteria

In the strongest sense, bilateral cross-prompt coordination means that each side affects the other during prompt construction, prompt selection, or prompt-conditioned inference. In ChordPrompt, visual prompts are projected into text space and textual prompts are projected into vision space at every encoder layer, so text prompts become visually aware and visual prompts become linguistically aware (Wang et al., 24 Jun 2025). In DCP, bilaterality is realized by symmetric co-attention between prompt tokens in the vision and language branches, with residual updates applied progressively across depths (Liu et al., 2023). In MM-Prompt, bilaterality appears twice: prompt selection in each modality depends on a cross-modal query formed from the opposite modality, and prompt recovery then reconstructs visual and textual prompts through iterative bidirectional interactions (Li et al., 26 May 2025).

A broader usage appears in LLM systems. PromptBridge defines bilateral cross-prompt coordination as preserving prompt effectiveness in both directions between two models, so a prompt engineered for model $S$ can be translated to model $T$, and vice versa, without per-task re-optimization (Wang et al., 1 Dec 2025). P3 uses the term operationally rather than explicitly: system prompt $s$ and user prompt $u$ are optimized together because unilateral optimization leaves unresolved affinity between role instructions and query-level complements. Its joint objective is

$$\max_{s,u} \;\; \mathbb{E}_{(q,t)\sim D}\big[M(s,u;q,t)\big] - \lambda\,\Omega(s,u),$$

which makes bilaterality a property of coordinated optimization rather than prompt exchange (Zhang et al., 21 Jul 2025).

A narrower and more contested usage appears in cross-prompt essay scoring. MAPLE coordinates across prompts through meta-learning and prompt/rubric conditioning, but explicitly does not implement bilateral, pairwise, bi-directional coordination between prompt pairs (Albatarni et al., 19 Apr 2026). PAES is even more restrictive: it is prompt-agnostic, learns prompt-invariant features, and contains no explicit bilateral prompt alignment at all (Ridley et al., 2020). These cases are important because they delimit the concept: not every cross-prompt method is bilateral.

2. Canonical mechanisms

Three mechanisms recur across the literature. The first is bidirectional projection or attention. ChordPrompt uses two learnable linear aligners,

$$\hat{T}_i^l = A_{V2T} V_i^l, \qquad \hat{V}_i^l = A_{T2V} T_i^l,$$

and injects the cross-projected prompts through the Value pathway only, leaving Query and Key untouched so that CLIP’s pre-trained attention patterns are preserved (Wang et al., 24 Jun 2025). DCP’s CMPA instead performs full prompt-to-prompt multi-head attention in both directions,

$$P_v^{(\ell+1)} = P_v^{(\ell)} + \mathrm{MHAttn}(P_v^{(\ell)},P_t^{(\ell)},P_t^{(\ell)}),\qquad P_t^{(\ell+1)} = P_t^{(\ell)} + \mathrm{MHAttn}(P_t^{(\ell)},P_v^{(\ell)},P_v^{(\ell)}),$$

with shared parameters across depths (Liu et al., 2023).

The second is bilateral retrieval or routing. ChordPrompt stores a textual prototype key $K_i$ for each domain, together with that domain’s prompts and aligners, and at inference computes a query prototype $P_x$, selects $i^*=\arg\max_i S(P_x,K_i)$, and falls back to vanilla CLIP if the best cosine similarity is below $T$0 (Wang et al., 24 Jun 2025). MM-Prompt similarly ranks prompt-pool keys using cross-enriched queries, so visual prompt choice depends on text-derived information and textual prompt choice depends on visual information (Li et al., 26 May 2025). PromptBridge externalizes routing into a natural-language mapping summary $T$1 or $T$2, which functions as a transfer rule rather than a latent controller (Wang et al., 1 Dec 2025).

The third is shared objectives over paired prompt variables. In PromptBridge, model drifting is formalized as

$T$3

and bilateral coordination learns both $T$4 and $T$5 from calibrated prompt pairs (Wang et al., 1 Dec 2025). In P3, joint optimization is gradient-free and mutation-like, but the score is always computed on the pair $T$6, not on either component alone (Zhang et al., 21 Jul 2025). In XPE’s DUAL variant, the encoder-produced prompt $T$7 and directly trained soft prompt $T$8 are concatenated and optimized under a single multilingual cross-entropy objective with a frozen backbone, so coordination emerges through a shared downstream loss rather than an explicit alignment penalty (Mikaberidze et al., 14 Aug 2025).

3. Cross-modal and vision–language instantiations

The most explicit architectural realizations occur in frozen CLIP-like models. ChordPrompt targets multi-domain task incremental learning with a frozen CLIP ViT-B/16 image encoder and a frozen Transformer text encoder. It attaches learnable prompt tokens to both branches layerwise, stores per-domain aligners and prompt sets in a prompt pool, and uses prototype-based domain selection with zero-shot fallback. On MTIL Order I, it reports Transfer $T$9, Avg $s$0, and Last $s$1, compared with $s$2 for ZSCL and $s$3 for DDAS, while training about $s$4M parameters versus $s$5M for DDAS adapters and $s$6M for full fine-tuning baselines (Wang et al., 24 Jun 2025).

Earlier CLIP prompt-learning work established the same design tendency under different names. DCP’s deeply coupled cross-modal prompt learning uses prompt-only co-attention repeated across $s$7 depths, with prompt length $s$8 in few-shot experiments. Averaged over 11 datasets, DCP improves mean accuracy over MaPLe by $s$9 percentage points at $u$0 shots, respectively (Liu et al., 2023). PMPO instead binds multiple learnable text prompts to depth-partitioned visual prompts generated through learned linear projections; with $u$1 prompts, prompt length $u$2, and ViT-B/16, it achieves a harmonic mean $u$3 on base-to-new generalization across 11 datasets, a $u$4 absolute gain over CoOp’s $u$5 (Tian et al., 2023).

The same bilateral pattern extends beyond image classification. MM-Prompt addresses continual VQA, where isolated visual and textual prompt selection causes modality imbalance over tasks. It combines cross-modal prompt query and cross-modal prompt recovery, uses General prompts at layers $u$6 and Expert prompts at $u$7, and optimizes

$u$8

On VQA v2, it reports $u$9 and $$\max_{s,u} \;\; \mathbb{E}_{(q,t)\sim D}\big[M(s,u;q,t)\big] - \lambda\,\Omega(s,u),$$0 for DI, outperforming MaPLe’s $$\max_{s,u} \;\; \mathbb{E}_{(q,t)\sim D}\big[M(s,u;q,t)\big] - \lambda\,\Omega(s,u),$$1 and $$\max_{s,u} \;\; \mathbb{E}_{(q,t)\sim D}\big[M(s,u;q,t)\big] - \lambda\,\Omega(s,u),$$2, and on NExT-QA it reports $$\max_{s,u} \;\; \mathbb{E}_{(q,t)\sim D}\big[M(s,u;q,t)\big] - \lambda\,\Omega(s,u),$$3 with $$\max_{s,u} \;\; \mathbb{E}_{(q,t)\sim D}\big[M(s,u;q,t)\big] - \lambda\,\Omega(s,u),$$4 for QI (Li et al., 26 May 2025).

Two related lines generalize the concept to asymmetric or debiased multimodal inference. Bi-CMPStereo treats event-to-frame and frame-to-event adaptation as prompt generation into a target canonical space; CDEA acts as the source-to-target prompt generator, SCC enforces canonical reconstruction, and bilateral fusion concatenates cost volumes from the event-target and frame-target branches. On DSEC All, Bi-CMPStereo reports MAE $$\max_{s,u} \;\; \mathbb{E}_{(q,t)\sim D}\big[M(s,u;q,t)\big] - \lambda\,\Omega(s,u),$$5, 1PE $$\max_{s,u} \;\; \mathbb{E}_{(q,t)\sim D}\big[M(s,u;q,t)\big] - \lambda\,\Omega(s,u),$$6, 2PE $$\max_{s,u} \;\; \mathbb{E}_{(q,t)\sim D}\big[M(s,u;q,t)\big] - \lambda\,\Omega(s,u),$$7, and RMSE $$\max_{s,u} \;\; \mathbb{E}_{(q,t)\sim D}\big[M(s,u;q,t)\big] - \lambda\,\Omega(s,u),$$8 (Xu et al., 16 Apr 2026). BiPrompt shifts the focus from transfer to test-time debiasing, coupling attention-guided visual erasure with Balanced Prompt Normalization on the text side. On Waterbirds it reports AVG $$\max_{s,u} \;\; \mathbb{E}_{(q,t)\sim D}\big[M(s,u;q,t)\big] - \lambda\,\Omega(s,u),$$9 and worst-group $$\hat{T}_i^l = A_{V2T} V_i^l, \qquad \hat{V}_i^l = A_{T2V} T_i^l,$$0; across Waterbirds, CamelDeer, and SpiderCrab it reports AVG $$\hat{T}_i^l = A_{V2T} V_i^l, \qquad \hat{V}_i^l = A_{T2V} T_i^l,$$1 and worst-group $$\hat{T}_i^l = A_{V2T} V_i^l, \qquad \hat{V}_i^l = A_{T2V} T_i^l,$$2 (Gupta et al., 5 Jan 2026).

4. Cross-model, multi-component, and multilingual LLM coordination

In LLM migration, bilateral coordination is primarily a transfer problem. PromptBridge assumes access to a small alignment suite, runs MAP-RPE separately on source and target models to obtain prompt optima $$\hat{T}_i^l = A_{V2T} V_i^l, \qquad \hat{V}_i^l = A_{T2V} T_i^l,$$3 and $$\hat{T}_i^l = A_{V2T} V_i^l, \qquad \hat{V}_i^l = A_{T2V} T_i^l,$$4, summarizes their systematic differences into natural-language transfer effects $$\hat{T}_i^l = A_{V2T} V_i^l, \qquad \hat{V}_i^l = A_{T2V} T_i^l,$$5, and learns the reverse direction by role reversal. Its test-time mapping is written as

$$\hat{T}_i^l = A_{V2T} V_i^l, \qquad \hat{V}_i^l = A_{T2V} T_i^l,$$6

Using GPT-4o as source, it reports $$\hat{T}_i^l = A_{V2T} V_i^l, \qquad \hat{V}_i^l = A_{T2V} T_i^l,$$7 on HumanEval and $$\hat{T}_i^l = A_{V2T} V_i^l, \qquad \hat{V}_i^l = A_{T2V} T_i^l,$$8 on CodeContests when transferring to o3, $$\hat{T}_i^l = A_{V2T} V_i^l, \qquad \hat{V}_i^l = A_{T2V} T_i^l,$$9 and $$P_v^{(\ell+1)} = P_v^{(\ell)} + \mathrm{MHAttn}(P_v^{(\ell)},P_t^{(\ell)},P_t^{(\ell)}),\qquad P_t^{(\ell+1)} = P_t^{(\ell)} + \mathrm{MHAttn}(P_t^{(\ell)},P_v^{(\ell)},P_v^{(\ell)}),$$0 when transferring to o4-mini, and $$P_v^{(\ell+1)} = P_v^{(\ell)} + \mathrm{MHAttn}(P_v^{(\ell)},P_t^{(\ell)},P_t^{(\ell)}),\qquad P_t^{(\ell+1)} = P_t^{(\ell)} + \mathrm{MHAttn}(P_t^{(\ell)},P_v^{(\ell)},P_v^{(\ell)}),$$1 and $$P_v^{(\ell+1)} = P_v^{(\ell)} + \mathrm{MHAttn}(P_v^{(\ell)},P_t^{(\ell)},P_t^{(\ell)}),\qquad P_t^{(\ell+1)} = P_t^{(\ell)} + \mathrm{MHAttn}(P_t^{(\ell)},P_v^{(\ell)},P_v^{(\ell)}),$$2 when transferring to Llama-3.1-70B-Instruct, with corresponding gains in agentic settings such as SWE-bench Verified and Terminal-Bench (Wang et al., 1 Dec 2025).

P3 addresses a different bilateral pair: system and user prompts. It searches over system prompts and short user complements, uses an LLM-as-judge for a numeric score in $$P_v^{(\ell+1)} = P_v^{(\ell)} + \mathrm{MHAttn}(P_v^{(\ell)},P_t^{(\ell)},P_t^{(\ell)}),\qquad P_t^{(\ell+1)} = P_t^{(\ell)} + \mathrm{MHAttn}(P_t^{(\ell)},P_v^{(\ell)},P_v^{(\ell)}),$$3, and then promotes online prompting either through a small complement generator or retrieval-based ICL. The empirical evidence directly targets the “affinity issue” between the two prompt components: for GPT-3.5-turbo on general QA, P3 averages $$P_v^{(\ell+1)} = P_v^{(\ell)} + \mathrm{MHAttn}(P_v^{(\ell)},P_t^{(\ell)},P_t^{(\ell)}),\qquad P_t^{(\ell+1)} = P_t^{(\ell)} + \mathrm{MHAttn}(P_t^{(\ell)},P_v^{(\ell)},P_v^{(\ell)}),$$4, versus $$P_v^{(\ell+1)} = P_v^{(\ell)} + \mathrm{MHAttn}(P_v^{(\ell)},P_t^{(\ell)},P_t^{(\ell)}),\qquad P_t^{(\ell+1)} = P_t^{(\ell)} + \mathrm{MHAttn}(P_t^{(\ell)},P_v^{(\ell)},P_v^{(\ell)}),$$5 for P3 without system optimization, while PAS+system reaches $$P_v^{(\ell+1)} = P_v^{(\ell)} + \mathrm{MHAttn}(P_v^{(\ell)},P_t^{(\ell)},P_t^{(\ell)}),\qquad P_t^{(\ell+1)} = P_t^{(\ell)} + \mathrm{MHAttn}(P_t^{(\ell)},P_v^{(\ell)},P_v^{(\ell)}),$$6, still below coordinated P3 (Zhang et al., 21 Jul 2025). On reasoning tasks, it reports $$P_v^{(\ell+1)} = P_v^{(\ell)} + \mathrm{MHAttn}(P_v^{(\ell)},P_t^{(\ell)},P_t^{(\ell)}),\qquad P_t^{(\ell+1)} = P_t^{(\ell)} + \mathrm{MHAttn}(P_t^{(\ell)},P_v^{(\ell)},P_v^{(\ell)}),$$7 on GSM8K for GPT-3.5-turbo, above PAS at $$P_v^{(\ell+1)} = P_v^{(\ell)} + \mathrm{MHAttn}(P_v^{(\ell)},P_t^{(\ell)},P_t^{(\ell)}),\qquad P_t^{(\ell+1)} = P_t^{(\ell)} + \mathrm{MHAttn}(P_t^{(\ell)},P_v^{(\ell)},P_v^{(\ell)}),$$8 and P3-ICL at $$P_v^{(\ell+1)} = P_v^{(\ell)} + \mathrm{MHAttn}(P_v^{(\ell)},P_t^{(\ell)},P_t^{(\ell)}),\qquad P_t^{(\ell+1)} = P_t^{(\ell)} + \mathrm{MHAttn}(P_t^{(\ell)},P_v^{(\ell)},P_v^{(\ell)}),$$9, and $K_i$0 on GPQA, above PAS at $K_i$1 (Zhang et al., 21 Jul 2025).

Multilingual work broadens bilaterality from prompt pairs to language pairs under a shared system prompt. Cross-Lingual Prompt Steerability formalizes a four-metric objective over languages—$K_i$2, $K_i$3, Consistency, and $K_i$4—and combines them into

$K_i$5

Its regression analysis over 1,000 random prompts finds positive associations for CoT, emotion, and scenario components, and negative associations for style, cross-language synthesis, role, and behavioral components. After optimization, Qwen2.5-7B-Instruct improves mean $K_i$6 from $K_i$7 to $K_i$8 and Consistency from $K_i$9 to $P_x$0, while output length variance drops from $P_x$1 to $P_x$2 (Zhang et al., 2 Dec 2025).

Cross-Prompt Encoder occupies an intermediate position between multilingual prompt transfer and coordinated prompt composition. XPE learns a reusable prompt encoder $P_x$3 over a pseudo prompt $P_x$4, producing $P_x$5, and DUAL concatenates this with a directly trained soft prompt: $P_x$6 With a frozen XLM-R Large backbone and prompt length $P_x$7, it reports on SIB-200 that, for Low-Performing targets trained on Seen sources, SPT reaches $P_x$8, DUAL$P_x$9 reaches $i^*=\arg\max_i S(P_x,K_i)$0, and XPE reaches $i^*=\arg\max_i S(P_x,K_i)$1; for Seen /wo Joshi5 targets trained on Joshi5, DUAL$i^*=\arg\max_i S(P_x,K_i)$2 reaches $i^*=\arg\max_i S(P_x,K_i)$3, above SPT at $i^*=\arg\max_i S(P_x,K_i)$4 and XPE at $i^*=\arg\max_i S(P_x,K_i)$5 (Mikaberidze et al., 14 Aug 2025).

5. Cross-prompt essay scoring: explicit, implicit, and absent bilaterality

Automated essay scoring provides the clearest contrast between genuine bilateral coordination and prompt-agnostic transfer. ProTACT explicitly coordinates essays with prompts and traits with traits. It encodes prompt-aware essay representations through essay–prompt cross-attention and adds an unsupervised topic-coherence feature extracted via LDA, then regularizes multiple trait predictors with a trait-similarity loss,

$i^*=\arg\max_i S(P_x,K_i)$6

using $i^*=\arg\max_i S(P_x,K_i)$7 and $i^*=\arg\max_i S(P_x,K_i)$8 in experiments. On ASAP/ASAP++, it reports average QWK $i^*=\arg\max_i S(P_x,K_i)$9, versus $T$00 for its CTS reimplementation and $T$01 for PAES, with particularly strong gains on low-resource Prompt 7 (Do et al., 2023).

MAPLE achieves cross-prompt coordination only implicitly. It builds cross-prompt episodes for prototypical meta-learning, conditions essay representations on prompt and rubric features through gating,

$T$02

and scores queries by proximity to support-set prototypes. It reports state-of-the-art performance on ELLIPSE and LAILA, exceeding strong baselines by $T$03 and $T$04 QWK points, respectively, while setting new highs on ASAP traits with unified ranges (Albatarni et al., 19 Apr 2026). However, the paper is explicit that MAPLE does not implement bilateral pairwise coordination between prompt pairs; its coordination arises from shared $T$05, prompt/rubric gating, and multi-prompt prototypes rather than a bi-directional coupling objective (Albatarni et al., 19 Apr 2026).

PAES represents the opposite endpoint. It is prompt-agnostic by design, excludes lexical word embeddings, relies on POS-based hierarchical encoding plus handcrafted non-prompt-specific features, and minimizes only pooled MSE over source prompts. On ASAP cross-prompt evaluation it achieves average QWK $T$06, tying TDNN’s $T$07 without using labeled or unlabeled target-prompt data during training, but it contains no bilateral prompt coordination mechanism (Ridley et al., 2020). This contrast is methodologically important: cross-prompt generalization can be achieved either by explicit bilateral alignment, by implicit prompt-conditioned meta-learning, or by suppressing prompt dependence altogether.

6. Evaluation regimes, limitations, and open directions

The literature evaluates bilateral cross-prompt coordination through the metrics of its host task rather than a shared benchmark family. Continual vision–language work uses Transfer, Avg, and Last in MTIL (Wang et al., 24 Jun 2025). LLM migration uses Pass@1, resolved rate, planning metrics, and relative gains under model drift (Wang et al., 1 Dec 2025). P3 reports preference-style averages on Arena-hard and Alpaca-Eval together with GSM8K and GPQA accuracy (Zhang et al., 21 Jul 2025). Essay scoring work is dominated by QWK (Albatarni et al., 19 Apr 2026, Do et al., 2023, Ridley et al., 2020). Stereo work uses MAE, 1PE, 2PE, 3PE, and RMSE (Xu et al., 16 Apr 2026). Test-time debiasing reports average and worst-group accuracy (Gupta et al., 5 Jan 2026). This suggests that bilaterality is presently an architectural and optimization principle, not a standardized evaluation task.

Recurring limitations are equally domain-specific. ChordPrompt is vulnerable to domain-selection errors when prototype keys are ambiguous, to degraded prototypes when class names are uninformative, and to the expressivity limits of linear aligners $T$08 and $T$09 (Wang et al., 24 Jun 2025). PromptBridge requires access to both models during calibration, and its textual mappings can degrade on niche tasks or rapidly changing APIs (Wang et al., 1 Dec 2025). BiPrompt can fail when Grad-CAM localizes spurious regions or when $T$10 and $T$11 are set aggressively enough to cause prompt collapse (Gupta et al., 5 Jan 2026). PMPO incurs higher memory and compute because multiple prompts remain active across visual depths, and its performance depends on the depth partitioning scheme (Tian et al., 2023). MAPLE remains limited in heterogeneous scoring settings such as ASAP because missing score levels imply missing prototypes (Albatarni et al., 19 Apr 2026).

Several forward directions recur explicitly in the papers. PromptBridge suggests parametric learning over prompt deltas with objectives such as

$T$12

as future work (Wang et al., 1 Dec 2025). MAPLE proposes pairwise prototype alignment, cross-prompt contrastive losses, and cycle consistency as hypothetical additions that would convert implicit multi-prompt coordination into explicit bilateral coordination (Albatarni et al., 19 Apr 2026). ChordPrompt identifies stronger adapters, such as gated MLPs or attention-based fusion, as plausible replacements for linear aligners (Wang et al., 24 Jun 2025). Cross-Lingual Prompt Steerability points to multilingual prompt optimization as a route to lower language-switching and more structured reasoning across languages, but also shows that stable cross-language gains remain difficult, especially in lower-resource cases such as Hindi (Zhang et al., 2 Dec 2025).

Taken together, these works establish bilateral cross-prompt coordination as a general design principle for prompt-based adaptation under shift. Its explicit forms use two-way projections, co-attention, or bidirectional transfer mappings; its weaker forms use shared objectives over paired prompt components; and its boundary cases show that cross-prompt generalization does not by itself imply bilaterality. A plausible synthesis is that the central technical question is not whether prompts are learned, but whether the two sides of a prompt relation are allowed to shape one another during selection, optimization, or inference.