Collab-REC: Collaborative Design Paradigms

• Collab-REC is a collaboration-centered design pattern that unifies diverse multi-agent frameworks through explicit role coordination and iterative refinement.
• It spans domains such as recommendation, vision-language grounding, document revision, and multimedia co-creation by leveraging specialized agents and structured aggregation.
• Empirical studies show improvements in accuracy, diversity, and processing efficiency, while also revealing trade-offs in cost and complexity.

Collab-REC is a collaboration-centered design pattern rather than a single canonical system. In recent arXiv literature, the term and closely related formulations denote multi-agent recommendation frameworks, specialist–MLLM collaboration for referring expression comprehension, collaborative REC/REG data generation, joint 3D grounding architectures, graph-based recommendation and assistance for document revision, collaborative deepfake annotation, and browser-native co-creation systems. Across these uses, the unifying idea is explicit coordination among complementary roles, models, or users, typically with grounded aggregation, iterative refinement, or structured conflict handling replacing monolithic inference (Wang et al., 2024, Banerjee et al., 20 Aug 2025, Yang et al., 27 Feb 2025, Zhang et al., 28 Sep 2025, Qian et al., 2024, Ruan et al., 2024, Zhang et al., 24 Jan 2026, Favory et al., 2019).

1. Terminological scope and major instantiations

The literature uses Collab-REC in several technically distinct ways. In recommender systems, MACRec defines an LLM-powered, multi-agent framework that directly tackles recommendation tasks through specialized agents such as Manager, User/Item Analyst, Reflector, Searcher, and Task Interpreter, rather than using agents primarily as user or item simulators (Wang et al., 2024). In tourism recommendation, "Collab-REC" denotes a multi-agent LLM framework with Personalization, Popularity, and Sustainability agents, coordinated by a non-LLM moderator to counteract popularity bias and improve diversity (Banerjee et al., 20 Aug 2025).

In vision-language work, REC denotes referring expression comprehension. FineCops-Ref uses "Collab-REC" for specialist–MLLM collaboration in fine-grained compositional REC, with Slow-Fast Adaptation and Candidate Region Selection as the two principal strategies (Yang et al., 27 Feb 2025). ColLab is not a task model but an automated REC/REG data engine that orchestrates multiple MLLMs and an LLM refiner, then applies Spatial Progressive Augmentation to resolve duplicate descriptions (Zhang et al., 28 Sep 2025). MCLN applies collaborative learning to 3DREC and 3DRES via separate branches linked by Relative Superpoint Aggregation and Adaptive Soft Alignment (Qian et al., 2024).

Beyond recommendation and grounding, Re3 supplies a graph-based formalization of collaborative document revision that supports recommendation and assistance over revisions, reviews, and responses (Ruan et al., 2024). The deepfake system "Collab" operationalizes collaborative annotation through confidence-weighted spatio-temporal IoU aggregation and hierarchical demonstration (Zhang et al., 24 Jan 2026). The Multi Web Audio Sequencer shows a browser-native collaborative sequencing model built around shared state, WebSockets, and fixed-grid editing (Favory et al., 2019).

System	Domain	Collaboration mechanism
MACRec	Recommender systems	Specialized LLM agents coordinated by a Manager
Collab-REC	Tourism recommendation	Three LLM agents plus a non-LLM moderator
FineCops-Ref Collab-REC	REC	Specialist–MLLM routing and candidate selection
ColLab	REC/REG data generation	Multi-MLLM description generation plus LLM fusion
MCLN	3D visual grounding	Separate 3DREC and 3DRES branches with soft alignment
Re3	Document revision	Intertextual graph linking revisions, reviews, responses
Collab	Deepfake annotation	Crowd annotation aggregation and layered evidence display
MWAS	Collaborative music making	Shared sequencer state and server-broadcast actions

This dispersion of usage suggests that Collab-REC functions less as a fixed benchmark label than as a family of explicit collaboration paradigms.

2. Multi-agent recommendation frameworks

MACRec operationalizes collaboration as a role-driven workflow for four recommendation settings: rating prediction, sequential recommendation, conversational recommendation, and explanation generation (Wang et al., 2024). The Manager is the central coordinator and follows a Thought–Action–Observation protocol. It plans task execution, assigns sub-tasks, aggregates replies, and generates final predictions, rankings, conversational replies, or explanations. The Reflector critiques the Manager’s previous output from the second run onward and either halts iteration or proposes targeted refinements. User Analyst and Item Analyst combine an info database with an interaction retriever to summarize profiles, attributes, and histories. The Searcher performs external retrieval, for example from Wikipedia, and the Task Interpreter converts dialogue history into an executable task specification.

MACRec’s workflows vary by task. Rating prediction requires User Analyst and Item Analyst, with Reflector optional. Sequential recommendation requires User Analyst and Reflector, with Item Analyst optional given scale. Explanation generation requires User Analyst, Item Analyst, and Searcher, with Reflector optional. Conversational recommendation requires Task Interpreter and Searcher. The framework explicitly states that aggregation is primarily textual synthesis by the Manager; no numeric ensemble or weighting scheme such as $\hat{y}_{ui} = \sum_a w_a s_a(u,i)$ is specified (Wang et al., 2024).

The evaluation context around MACRec uses standard task metrics rather than training losses. For rating prediction, the paper gives

$$\mathrm{RMSE} = \sqrt{\frac{1}{|D|} \sum_{(u,i)\in D} (r_{ui} - \hat{r}_{ui})^2},$$

with MSE and MAE defined analogously. For sequential recommendation, it lists standard next-item and ranking formulations such as $p(i_{t+1} \mid u, i_{1:t})$, Precision@K, Recall@K, NDCG@K, HitRate@K, and MRR. For explanation generation, BLEU, ROUGE, and METEOR are identified as standard metrics, but the paper reports no quantitative results, no baselines, and no datasets; its emphasis is framework introduction and qualitative web demonstration (Wang et al., 2024).

The tourism Collab-REC paper instantiates a more explicitly scored negotiation protocol. Three LLM-based agents—Personalization, Popularity, and Sustainability—each produce a ranked top-$k$ list from a catalog of 200 European cities, and a non-LLM moderator merges them over multiple rounds (Banerjee et al., 20 Aug 2025). The moderator grounds outputs to a knowledge base, flags hallucinations, computes a recursive city score, normalizes via min-max, and produces a "Collective Offer." The core scoring rule is

$$s(c, t) = s(c, t-1) + \sum_{a \in \mathcal{A}} \Bigl( \frac{1}{\mathrm{rank}(c \text{ in } L_a(t))} \cdot \bigl( \mathrm{Success}_a(c, t) + \mathrm{Reliability}_a(t) - \mathrm{Hallucination}_a(t) \bigr) \Bigr).$$

Rejection can follow a Majority strategy, in which a city is rejected if omitted by at least two agents, or an Aggressive strategy, in which omission by even a single agent is sufficient. Early stopping is triggered if Moderator Success reaches $1$ or improves by $\delta\%$ over round $0$, with $\delta \in \{20\%, 60\%\}$ and a minimum of $5$ rounds enforced.

Quantitatively, the tourism framework reports Moderator Success of approximately $$\mathrm{RMSE} = \sqrt{\frac{1}{|D|} \sum_{(u,i)\in D} (r_{ui} - \hat{r}_{ui})^2},$$0–$$\mathrm{RMSE} = \sqrt{\frac{1}{|D|} \sum_{(u,i)\in D} (r_{ui} - \hat{r}_{ui})^2},$$1 for gpt-o4-mini MAMI with early stopping and approximately $$\mathrm{RMSE} = \sqrt{\frac{1}{|D|} \sum_{(u,i)\in D} (r_{ui} - \hat{r}_{ui})^2},$$2–$$\mathrm{RMSE} = \sqrt{\frac{1}{|D|} \sum_{(u,i)\in D} (r_{ui} - \hat{r}_{ui})^2},$$3 for gemini-2.5-flash MAMI with early stopping, compared with SASI values of approximately $$\mathrm{RMSE} = \sqrt{\frac{1}{|D|} \sum_{(u,i)\in D} (r_{ui} - \hat{r}_{ui})^2},$$4 and approximately $$\mathrm{RMSE} = \sqrt{\frac{1}{|D|} \sum_{(u,i)\in D} (r_{ui} - \hat{r}_{ui})^2},$$5–$$\mathrm{RMSE} = \sqrt{\frac{1}{|D|} \sum_{(u,i)\in D} (r_{ui} - \hat{r}_{ui})^2},$$6 respectively (Banerjee et al., 20 Aug 2025). Diversity improves markedly: Gini falls to approximately $$\mathrm{RMSE} = \sqrt{\frac{1}{|D|} \sum_{(u,i)\in D} (r_{ui} - \hat{r}_{ui})^2},$$7 for gpt-o4-mini M20 and approximately $$\mathrm{RMSE} = \sqrt{\frac{1}{|D|} \sum_{(u,i)\in D} (r_{ui} - \hat{r}_{ui})^2},$$8 for gemini-2.5-flash M20, while normalized entropy rises to approximately $$\mathrm{RMSE} = \sqrt{\frac{1}{|D|} \sum_{(u,i)\in D} (r_{ui} - \hat{r}_{ui})^2},$$9 and approximately $p(i_{t+1} \mid u, i_{1:t})$0. The trade-off is cost: across 45 queries per model, the framework uses approximately $p(i_{t+1} \mid u, i_{1:t})$1M tokens, approximately $p(i_{t+1} \mid u, i_{1:t})$2 API calls, and approximately $p(i_{t+1} \mid u, i_{1:t})$3 seconds per round by round $p(i_{t+1} \mid u, i_{1:t})$4, making MAMI approximately $p(i_{t+1} \mid u, i_{1:t})$5 more expensive than SASI (Banerjee et al., 20 Aug 2025).

3. Collaborative referring expression comprehension and generation

In the referring expression literature, REC is formalized as grounding a linguistic expression to an image region. FineCops-Ref writes this as

$p(i_{t+1} \mid u, i_{1:t})$6

with IoU-based evaluation against $p(i_{t+1} \mid u, i_{1:t})$7 (Yang et al., 27 Feb 2025). The dataset is designed around fine-grained compositional reasoning and absent-target rejection. Its test split contains 9,605 positive expressions, 9,814 negative expressions, and 8,507 negative images; train contains 163,792 positive and 80,451 negative expressions; val contains 18,455 positive and 9,029 negative expressions (Yang et al., 27 Feb 2025).

Two collaborative methods define its Collab-REC formulation. Slow-Fast Adaptation is a training-free router that extracts the target category with GPT-3.5-turbo, runs MM-GDINO-L with confidence threshold $p(i_{t+1} \mid u, i_{1:t})$8, and routes Level 1 cases to a specialist or more complex cases to an MLLM. Candidate Region Selection uses a specialist to produce top-$p(i_{t+1} \mid u, i_{1:t})$9 candidate regions after NMS$k$0, then asks the MLLM to pick a single option or "None" via a multi-choice prompt; instruction tuning uses RefCOCO/+/g as source data, while FineCops-Ref training data is used only for "None" option training (Yang et al., 27 Feb 2025).

The empirical pattern is sharply differentiated by difficulty level. MM-GDINO-L achieves 85.13% on Level 1, 43.54% on Level 2, 42.89% on Level 3, and 68.32% overall, while Qwen2-VL-7B achieves 78.24%, 66.13%, 60.72%, and 73.09% respectively (Yang et al., 27 Feb 2025). SFA improves CogVLM from 73.57% to 75.03% on FineCops-Ref with focus enhancement, and CRS with InternVL2.5-8B reaches 78.58% on FineCops-Ref, 82.80% on Ref-Adv, and 40.34% on Ref-Reasoning. The paper also reports that SFA outperforms CogVLM by +5.6% while using approximately 62% of the FLOPs, and that CRS reduces inference time to approximately 10–30% of the original MLLM because output is reduced from four coordinates to a single-token choice (Yang et al., 27 Feb 2025).

ColLab addresses a different point in the pipeline: automated REC/REG data generation without human supervision (Zhang et al., 28 Sep 2025). It collects 100 images across indoor, outdoor, and workplace environments, applies object detection with confidence threshold $k$1, and retains $k$2 instance boxes. For each cropped instance image $k$3, Qwen2.5-VL-3B, -7B, and -32B each generate a candidate description from the prompt "What are the characteristics of C in I_c?", and DeepSeek-V3 merges them with "Extract descriptions of C based on D = {d_1, …, d_N}." The paper reports mean length 16.32, variance 12.88, and time 3.47s for 3B; mean length 20.68, variance 18.31, and time 3.81s for 7B; and mean length 26.54, variance 23.80, and time 4.71s for 32B. Single-threaded throughput reaches approximately 5,082 items/day (Zhang et al., 28 Sep 2025).

ColLab’s Spatial Progressive Augmentation resolves duplicate descriptions among same-category instances. The image is partitioned into center, transition, and edge regions crossed with top, bottom, left, and right; duplicate groups are identified by $k$4 with count $k$5 and recursively subdivided until each instance occupies a unique subregion (Zhang et al., 28 Sep 2025). The traffic-light example begins with three identical descriptions—"The traffic light in the image is red"—and ends with three unique expressions after spatial augmentation.

MCLN extends collaborative REC into 3D grounding by separating 3DREC and 3DRES into independent branches linked by Relative Superpoint Aggregation and Adaptive Soft Alignment (Qian et al., 2024). On ScanRefer, it reports 57.17% Overall [email protected] and 45.53% Overall [email protected] for 3DREC, versus 54.59% and 42.26% for EDA, and 58.70% Overall [email protected], 50.70% Overall [email protected], and 44.72% mIoU for 3DRES, versus 54.60%, 39.80%, and 39.50% for 3D-STMN (Qian et al., 2024). The architecture’s collaborative component is not agentic but multi-branch: task-specific decoders remain separate, while superpoint-level soft alignment reduces conflict and improves mutual coherence.

Taken together, these papers show three distinct collaboration loci inside REC pipelines: routing between specialist and generalist models, collective description generation for data construction, and cross-task coordination between grounding heads.

4. Collaborative revision recommendation and assistance

Re3 formulates collaborative document revision as a structured graph over four document types: original document $k$6, revised document $k$7, review $k$8, and response $k$9 (Ruan et al., 2024). At granularity $$s(c, t) = s(c, t-1) + \sum_{a \in \mathcal{A}} \Bigl( \frac{1}{\mathrm{rank}(c \text{ in } L_a(t))} \cdot \bigl( \mathrm{Success}_a(c, t) + \mathrm{Reliability}_a(t) - \mathrm{Hallucination}_a(t) \bigr) \Bigr).$$0, aligned text elements form edges

$$s(c, t) = s(c, t-1) + \sum_{a \in \mathcal{A}} \Bigl( \frac{1}{\mathrm{rank}(c \text{ in } L_a(t))} \cdot \bigl( \mathrm{Success}_a(c, t) + \mathrm{Reliability}_a(t) - \mathrm{Hallucination}_a(t) \bigr) \Bigr).$$1

with additions and deletions represented as $$s(c, t) = s(c, t-1) + \sum_{a \in \mathcal{A}} \Bigl( \frac{1}{\mathrm{rank}(c \text{ in } L_a(t))} \cdot \bigl( \mathrm{Success}_a(c, t) + \mathrm{Reliability}_a(t) - \mathrm{Hallucination}_a(t) \bigr) \Bigr).$$2 or $$s(c, t) = s(c, t-1) + \sum_{a \in \mathcal{A}} \Bigl( \frac{1}{\mathrm{rank}(c \text{ in } L_a(t))} \cdot \bigl( \mathrm{Success}_a(c, t) + \mathrm{Reliability}_a(t) - \mathrm{Hallucination}_a(t) \bigr) \Bigr).$$3. Each alignment carries a label triple $$s(c, t) = s(c, t-1) + \sum_{a \in \mathcal{A}} \Bigl( \frac{1}{\mathrm{rank}(c \text{ in } L_a(t))} \cdot \bigl( \mathrm{Success}_a(c, t) + \mathrm{Reliability}_a(t) - \mathrm{Hallucination}_a(t) \bigr) \Bigr).$$4, where edit action includes Add, Delete, Modify, Merge, Split, and Fusion, and edit intent includes Grammar, Clarity, Fact/Evidence, Claim, and Other. Reviews and responses are linked to edits through $$s(c, t) = s(c, t-1) + \sum_{a \in \mathcal{A}} \Bigl( \frac{1}{\mathrm{rank}(c \text{ in } L_a(t))} \cdot \bigl( \mathrm{Success}_a(c, t) + \mathrm{Reliability}_a(t) - \mathrm{Hallucination}_a(t) \bigr) \Bigr).$$5 and $$s(c, t) = s(c, t-1) + \sum_{a \in \mathcal{A}} \Bigl( \frac{1}{\mathrm{rank}(c \text{ in } L_a(t))} \cdot \bigl( \mathrm{Success}_a(c, t) + \mathrm{Reliability}_a(t) - \mathrm{Hallucination}_a(t) \bigr) \Bigr).$$6 respectively (Ruan et al., 2024).

The Re3-Sci instantiation contains 314 document revision pairs with full-scope annotations. It includes 11,648 sentence-level aligned and labeled edits, 5,064 paragraph-level edits, 2,008 section-level edits, 2,676 subsentence-level edits extracted from 1,453 sentence revision pairs, 560 review sentences for 42 documents, 413 review-to-revision linkages, 784 human-written summary sentences across 42 documents, and 1,364 sentence-to-edit links (Ruan et al., 2024). Sentence pre-alignment combines Levenshtein distance, fuzzy string matching, and SBERT, with thresholds $$s(c, t) = s(c, t-1) + \sum_{a \in \mathcal{A}} \Bigl( \frac{1}{\mathrm{rank}(c \text{ in } L_a(t))} \cdot \bigl( \mathrm{Success}_a(c, t) + \mathrm{Reliability}_a(t) - \mathrm{Hallucination}_a(t) \bigr) \Bigr).$$7 and $$s(c, t) = s(c, t-1) + \sum_{a \in \mathcal{A}} \Bigl( \frac{1}{\mathrm{rank}(c \text{ in } L_a(t))} \cdot \bigl( \mathrm{Success}_a(c, t) + \mathrm{Reliability}_a(t) - \mathrm{Hallucination}_a(t) \bigr) \Bigr).$$8, achieving 0.95 accuracy after annotator validation.

The framework supports several recommendation and assistance tasks. Edit intent classification over 8,937 test samples reaches accuracy 0.70 and macro F1 0.69 under the best Llama2-70B prompting configuration; additions and deletions alone improve accuracy to approximately 0.83 (Ruan et al., 2024). Revision alignment is framed as binary classification and reaches 0.97 accuracy. Review request extraction reaches 0.80 accuracy, with recall 0.98 for positives and precision 0.95 for negatives. For revision-to-response summarization, GPT-4 zero-shot obtains 95.96% factuality, 79.09% comprehensiveness, 89.82% specificity, compactness 2.36 edits per summary sentence, and 72.5% organization by action labels, compared with human values of 100%, 98.82%, 95.56%, 1.74, and 100% organization by section (Ruan et al., 2024).

Re3 is directly relevant to a recommendation reading of Collab-REC because it enables review-aware edit recommendation, intent prediction, review-to-edit linking, and response drafting. The data also expose structural regularities useful for prioritization. Edit actions are distributed as Modify 54.54%, Add 28.93%, Delete 12.44%, Split 2.30%, Merge 1.53%, and Fusion 0.26, while intents are distributed as Fact/Evidence 45.02%, Clarity 21.78%, Claim 15.44%, Grammar 14.38%, and Other 2.68 (Ruan et al., 2024). Average Edit Ratio is 18.45%, Semantic Edit Ratio is 11.18%, average Crest Factor is 3.79 at paragraph level and 2.54 at section level, and more than half of explicit review requests are implemented. This suggests that collaborative revision recommendation can be grounded not only in semantic similarity but also in document position, request explicitness, and edit concentration.

5. Human-in-the-loop collaborative annotation and co-creation

The deepfake system "Collab" demonstrates a human-centered Collab-REC variant in which collaboration occurs among users rather than model components (Zhang et al., 24 Jan 2026). Each annotation stores a normalized spatial region, a temporal interval, a label, confidence in $$s(c, t) = s(c, t-1) + \sum_{a \in \mathcal{A}} \Bigl( \frac{1}{\mathrm{rank}(c \text{ in } L_a(t))} \cdot \bigl( \mathrm{Success}_a(c, t) + \mathrm{Reliability}_a(t) - \mathrm{Hallucination}_a(t) \bigr) \Bigr).$$9, and an optional rationale. Aggregation is based on spatio-temporal IoU over axis-aligned cuboids in $1$0–$1$1–$1$2 space:

$1$3

Annotations are sorted by confidence, merged into existing regions if IoU $1$4, and otherwise form new clusters. Region coordinates are fused by confidence-weighted averaging,

$1$5

and label scores incorporate both confidence and user history:

$1$6

Aggregated confidence uses the average of the top $1$7 confidence values (Zhang et al., 24 Jan 2026).

The interface presents hierarchical demonstration: semi-transparent overlays on the player, interval markers on the timeline, hover-based candidate labels, and click-through rationale clusters. Overlays are colored by aggregate confidence and agreement: Green if $1$8 and $1$9, Red if $\delta\%$0 or $\delta\%$1, and Orange otherwise (Zhang et al., 24 Jan 2026). In a seven-day online study with $\delta\%$2, Collab achieves overall F1=0.883, outperforming the non-demonstration condition by $\delta\%$3 and the non-aggregation condition by $\delta\%$4; one-way ANOVA gives $\delta\%$5 (Zhang et al., 24 Jan 2026). It also yields smaller and more precise boxes, with median area 1.94% versus 3.27% and 5.87% in comparison conditions, and lower mean confidence, 86.9% versus 88.1% and 93.44%, which the paper interprets as more calibrated judgment.

The Multi Web Audio Sequencer shows a different human-collaborative design centered on remote music making in the browser (Favory et al., 2019). Audio generation and UI run client-side via the Web Audio API, while a Node.js server holds the shared room state in memory and communicates over WebSockets. All edits are serialized to JSON, sent to the server, and rebroadcast to every client in the room. Most actions are binary toggles whose order does not affect the final state, but track creation and deletion mutate array indices; to prevent divergence, those operations become effective only after server acknowledgment and rebroadcast, explicitly relying on TCP’s in-order delivery guarantees (Favory et al., 2019).

MWAS does not implement explicit latency compensation, transport-level clock synchronization, roles, permissions, undo, or CRDT-style concurrency control (Favory et al., 2019). Instead, it constrains the interaction space: fixed-grid step sequencing, local-only solo and mute, chat, presence indicators, and a simple shared-state model. This design is instructive for Collab-REC more broadly because it shows how collaboration can be stabilized by narrowing the set of conflict-prone actions rather than by adding a heavy coordination layer.

6. Recurring design patterns, evaluation regimes, and open issues

Several recurrent design patterns appear across these otherwise disparate systems. One is specialization: MACRec separates Manager, Reflector, Analysts, Searcher, and Interpreter (Wang et al., 2024); the tourism framework separates Personalization, Popularity, Sustainability, and moderator (Banerjee et al., 20 Aug 2025); FineCops-Ref separates specialist detectors from MLLMs (Yang et al., 27 Feb 2025); MCLN separates 3DREC from 3DRES branches (Qian et al., 2024). A second is grounded aggregation: MACRec delegates retrieval to Searcher and structured databases, the tourism moderator validates cities against a knowledge base, ColLab fuses multiple MLLM descriptions and filters low-frequency details, Re3 links revisions to specific review and response evidence, and the deepfake system aggregates annotations through spatio-temporal overlap and user history (Wang et al., 2024, Banerjee et al., 20 Aug 2025, Zhang et al., 28 Sep 2025, Ruan et al., 2024, Zhang et al., 24 Jan 2026). A third is iterative refinement: Reflector loops in MACRec, multi-round negotiation in tourism Collab-REC, recursive subdivision in SPA, two-stage routing and instruction tuning in FineCops-Ref, and adaptive cross-branch consistency in MCLN all embody controlled revision rather than one-shot prediction (Wang et al., 2024, Banerjee et al., 20 Aug 2025, Yang et al., 27 Feb 2025, Zhang et al., 28 Sep 2025, Qian et al., 2024).

The evaluation regimes are correspondingly heterogeneous. Recommendation settings use RMSE, MSE, MAE, Precision@K, Recall@K, NDCG@K, HitRate@K, MRR, Moderator Success, Gini Index, and Normalized Entropy (Wang et al., 2024, Banerjee et al., 20 Aug 2025). REC work uses Precision@1, [email protected], [email protected], Recall@1, AUROC, mIoU, BLEU, METEOR, CIDEr, and downstream discriminability (Yang et al., 27 Feb 2025, Zhang et al., 28 Sep 2025, Qian et al., 2024). Re3 uses Precision, Recall, F1, and Accuracy for alignment, intent, and request extraction, then human evaluation for revision summarization (Ruan et al., 2024). The deepfake annotation system combines F1, geometry statistics, NASA-TLX, and behavioral measures such as box area and 3D IoU-to-consensus trends (Zhang et al., 24 Jan 2026). This suggests that Collab-REC is methodologically unified by coordination structure more than by a common benchmark protocol.

The limitations are equally domain-specific. MACRec reports no datasets, baselines, or quantitative results and does not discuss cost or latency (Wang et al., 2024). ColLab reports no downstream REC/REG accuracy metrics because its contribution is dataset construction rather than model training (Zhang et al., 28 Sep 2025). FineCops-Ref remains bounded by detector quality, routing accuracy, and proposal recall (Yang et al., 27 Feb 2025). MCLN does not report parameter counts, FLOPs, memory footprint, or per-scene runtime (Qian et al., 2024). Re3 is English-only and constrained by CC-BY-NC source data (Ruan et al., 2024). The deepfake system notes risks from coordinated misinformation, annotator bias, and cold start (Zhang et al., 24 Jan 2026). The tourism framework is computationally expensive and limited by a fixed 200-city knowledge base (Banerjee et al., 20 Aug 2025). MWAS leaves concurrency control, persistence, and transport synchronization largely unresolved (Favory et al., 2019).

A plausible implication is that future Collab-REC systems will continue to combine three ingredients already visible across the literature: explicit decomposition into complementary roles, grounding against structured external evidence, and iterative or hierarchical mechanisms for arbitration. The exact form—agentic, multi-branch, graph-based, or crowd-driven—depends on whether the underlying task is recommendation, grounding, revision, annotation, or co-creation.