ReCollab: Collaborative Retrieval & Coordination

• ReCollab is a collaborative AI framework that integrates decomposition, retrieval, and coordination to enhance multi-agent teamwork and question-answering.
• Its design spans ad-hoc teammate modeling and collaborative QA, leveraging specialized modules such as moderators, expert panels, and retrieval systems.
• Empirical results demonstrate improved classification accuracy and episodic returns, underscoring ReCollab’s effectiveness across diverse benchmarks.

Searching arXiv for the cited ReCollab-related papers to ground the article in current records. ReCollab denotes, in the literature provided, a set of collaborative AI formulations organized around decomposition, retrieval, and coordination. Its most specific usage is "Retrieval-augmented Cooperative LLMs for Ad-hoc Teammate Modeling," where a controlled agent infers the behavior type of an unseen teammate from a short probe trajectory and then switches to a corresponding best-response policy (Wallace et al., 5 Dec 2025). Closely related work also uses the label to describe retrieval collaboration between a white-box small LLM and a black-box LLM for multi-hop question answering (Xu et al., 7 Apr 2025), and as a system wrapper over an AC-RAG-style Detector/Resolver loop for domain-specific retrieval-augmented generation (Zhang et al., 18 Sep 2025). A still earlier precursor, CollabQA, formalizes a moderator-mediated collaborative question-answering task over disjoint private knowledge graphs and is explicitly presented as a foundation for "scaled, real-world collaborative QA systems like ReCollab" (Hu et al., 2022).

1. Terminological scope and research lineage

The term is not attached to a single canonical architecture across the cited papers. Instead, it names several closely related collaboration patterns in which multiple components contribute complementary competencies under partial observability, limited context, or disjoint knowledge.

Usage	Core setting	Principal components
ReCollab (Wallace et al., 5 Dec 2025)	Ad-hoc teamwork in Overcooked	LLM type inference, rubric, retrieved exemplar trajectories, policy routing
ReCollab principles in Collab-RAG (Xu et al., 7 Apr 2025)	Multi-hop QA with RAG	White-box SLM decomposer, black-box LLM reader/synthesizer, differentiable retriever
ReCollab API over AC-RAG (Zhang et al., 18 Sep 2025)	Vertical-domain RAG	Detector, Resolver, moderator, iterative retrieval-refinement
ReCollab-style foundation in CollabQA (Hu et al., 2022)	Collaborative QA over private KGs	Moderator $P_0$, expert panelists $P_1,\dots,P_n$, turn-based sub-questioning

This multiplicity is itself significant. A plausible implication is that ReCollab functions less as a standardized benchmark label than as a recurring systems idea: structured collaboration among specialized modules, typically with explicit mediation, retrieval, or both.

2. Formal foundations in collaborative question answering

The CollabQA formulation considers agents $P_0,P_1,\dots,P_n$, where $P_0$ is both the "receptionist" interfacing with the external user and the moderator coordinating experts $P_1,\dots,P_n$ (Hu et al., 2022). Each expert $P_i$ holds a private knowledge graph $\mathrm{KG}_i=(\mathcal V_i,\mathcal E_i)$, while the union $\mathrm{KG}=\bigcup_{i=1}^n \mathrm{KG}_i$ contains all facts. A complex question $Q$ is defined by an unknown ground-truth reasoning path

$$\pi(Q)=[v_0 \xrightarrow{r_1} v_1 \xrightarrow{r_2}\cdots \xrightarrow{r_L} v_L],$$

with $P_1,\dots,P_n$0 and $P_1,\dots,P_n$1. The task requires that $P_1,\dots,P_n$2 "cuts across" multiple $P_1,\dots,P_n$3, so that no single expert can answer $P_1,\dots,P_n$4 without collaboration. Success occurs if and only if the moderator produces the correct terminal entity $P_1,\dots,P_n$5, or a natural-language paraphrase thereof, within $P_1,\dots,P_n$6 turns.

The synthetic dataset builds three disjoint graphs $P_1,\dots,P_n$7 for "Person," "Company," and "City" domains. Each graph contains typed, many-to-one or one-to-one edges so that multi-hop paths produce unique endpoints. Question generation samples a start node, performs a constrained depth-first search of length $P_1,\dots,P_n$8, records the reasoning path $P_1,\dots,P_n$9, instantiates a full question $P_0,P_1,\dots,P_n$0 from templates, and decomposes $P_0,P_1,\dots,P_n$1 into $P_0,P_1,\dots,P_n$2 one-hop sub-questions. The dataset reserves 66.8K training, 8.35K development, and 8.35K test examples, with 49 templates for full $P_0,P_1,\dots,P_n$3 and 28 for one-hop sub-questions (Hu et al., 2022).

The expert model is a KGQA module with three parts. The graph encoder is a simplified Relational GCN with layer update

$P_0,P_1,\dots,P_n$4

followed by a BiLSTM question encoder producing $P_0,P_1,\dots,P_n$5, and a node selector with attention score

$P_0,P_1,\dots,P_n$6

with answer $P_0,P_1,\dots,P_n$7. Sub-question pre-training uses standard cross-entropy over nodes. The moderator encodes dialogue history with a BiLSTM into state $P_0,P_1,\dots,P_n$8, acts over sub-question templates plus a special "terminate" template, and follows

$P_0,P_1,\dots,P_n$9

Training uses policy gradients with entropy regularization, under either a sparse reward $P_0$0 for a correct final answer within $P_0$1 and $P_0$2 otherwise, or an enhanced reward that injects structural prior when multiple or non-one non-UNK responses occur (Hu et al., 2022).

The protocol is explicitly turn-based. At $P_0$3, $P_0$4 broadcasts $P_0$5; experts return either an entity or UNK; the moderator updates state, selects the next sub-question template, fills the one non-UNK entity into it, and repeats until termination. Sub-answers are merged by template substitution into the full reasoning path. Evaluation uses Exact-Match Answer, Exact-Match Path, Collaborative Success Rate, and Communication Overhead. The paper reports that the problem can be challenging without introducing prior of the collaboration structure, unless experts are perfect and uniform, and proposes extensions including role enrichment, directed polls to subsets of experts, reasoning DAGs or trees, natural-language generation modules, and Multi-Agent RL (Hu et al., 2022).

3. ReCollab as retrieval-augmented ad-hoc teammate modeling

In the ad-hoc teamwork setting, the controlled agent must infer the behavior "type" of an unseen teammate $P_0$6 from a short probe trajectory $P_0$7 and then switch to the corresponding best-response policy $P_0$8 (Wallace et al., 5 Dec 2025). The base framework, CoLLAB ("Cooperative LLM-based Agent Belief"), converts low-level trajectory statistics $P_0$9 into a natural-language description $P_1,\dots,P_n$0, presents $P_1,\dots,P_n$1 together with a hand-crafted rubric $P_1,\dots,P_n$2 to an LLM $P_1,\dots,P_n$3, and obtains a predicted type $P_1,\dots,P_n$4. Policy routing is immediate: once $P_1,\dots,P_n$5 is chosen, the controlled agent selects $P_1,\dots,P_n$6 for the rest of the episode.

ReCollab extends this setup by adding Retrieval-Augmented Generation. At inference time, it retrieves $P_1,\dots,P_n$7 exemplar summaries $P_1,\dots,P_n$8 from an offline database $P_1,\dots,P_n$9 of labeled probe trajectories of all types, and concatenates the observed description $P_i$0, the rubric $P_i$1, and the $P_i$2 most similar exemplar descriptions $P_i$3 in the prompt. The stated motivation is that LLM-only prompts can confuse types whose rubric statistics overlap, including "plate-focused" and "mixed." Retrieval grounds the LLM's inference in concrete prior behavior and is reported to improve robustness (Wallace et al., 5 Dec 2025).

Feature extraction is based on the first $P_i$4 steps,

$P_i$5

from which the framework computes raw statistics such as dwell_window, count_pot_interact, dwell_plate_area, cum_reward, and count_blocked. Feature selection retains the top $P_i$6 features according to mutual information

$P_i$7

For each type $P_i$8 and selected feature $P_i$9, the rubric stores

$\mathrm{KG}_i=(\mathcal V_i,\mathcal E_i)$0

which are rendered into natural-language summaries such as "Type A spends on average $\mathrm{KG}_i=(\mathcal V_i,\mathcal E_i)$1 minutes ($\mathrm{KG}_i=(\mathcal V_i,\mathcal E_i)$2) near the window" (Wallace et al., 5 Dec 2025).

The retrieval mechanism uses an embedding model $\mathrm{KG}_i=(\mathcal V_i,\mathcal E_i)$3 to map a feature summary $\mathrm{KG}_i=(\mathcal V_i,\mathcal E_i)$4 to a vector $\mathrm{KG}_i=(\mathcal V_i,\mathcal E_i)$5. For an exemplar $\mathrm{KG}_i=(\mathcal V_i,\mathcal E_i)$6 with embedding $\mathrm{KG}_i=(\mathcal V_i,\mathcal E_i)$7, the score is

$\mathrm{KG}_i=(\mathcal V_i,\mathcal E_i)$8

typically cosine similarity,

$\mathrm{KG}_i=(\mathcal V_i,\mathcal E_i)$9

The prompt then includes the observed behavior, the behavior rubric for types $\mathrm{KG}=\bigcup_{i=1}^n \mathrm{KG}_i$0, similar past behaviors, and the question asking which teammate type best matches the observation and requesting citation of rubric and examples. The paper further expresses the type posterior as

$\mathrm{KG}=\bigcup_{i=1}^n \mathrm{KG}_i$1

with $\mathrm{KG}=\bigcup_{i=1}^n \mathrm{KG}_i$2 over retrieved exemplars (Wallace et al., 5 Dec 2025).

4. Prompting, indexing, and adaptive policy routing

The offline database for ReCollab is constructed by collecting, for each $\mathrm{KG}=\bigcup_{i=1}^n \mathrm{KG}_i$3 and each kitchen layout, $\mathrm{KG}=\bigcup_{i=1}^n \mathrm{KG}_i$4 probe trajectories $\mathrm{KG}=\bigcup_{i=1}^n \mathrm{KG}_i$5 under the best-response default policy, then computing $\mathrm{KG}=\bigcup_{i=1}^n \mathrm{KG}_i$6 and $\mathrm{KG}=\bigcup_{i=1}^n \mathrm{KG}_i$7 for all examples (Wallace et al., 5 Dec 2025). A text embedding model, exemplified by text-embedding-3-large, computes keys $\mathrm{KG}=\bigcup_{i=1}^n \mathrm{KG}_i$8, and the triples $\mathrm{KG}=\bigcup_{i=1}^n \mathrm{KG}_i$9 are stored in a vector-indexed database, exemplified by FAISS. At inference, the system computes $Q$0, queries FAISS for the top $Q$1 most similar embeddings by cosine similarity, and retrieves the corresponding labeled summaries as $Q$2.

Prompting is fully specified. The template begins with a summary of teammate behavior over the first $Q$3 steps, followed by rubric summaries for each possible type and, in the ReCollab variant, examples of past probe summaries labeled by type. The requested output is the most likely type together with reasoning. The paper provides a concrete example in which dwell_window, count_pot_interact, and dwell_plate_area are matched against "Pot-focused" and "Plate-focused" rubric statistics, with the LLM answering "Pot-focused" by citing both the prototype statistics and a retrieved exemplar (Wallace et al., 5 Dec 2025).

The adaptive policy mechanism is deliberately conservative. At $Q$4, after the probe ends, CoLLAB or ReCollab predicts $Q$5 with a confidence score, and the controlled agent switches policy to $Q$6 for $Q$7. To avoid oscillation, the switch is executed only once. The probe length $Q$8 is a hyperparameter; the specification states that larger $Q$9 yields more accurate classification but slower adaptation (Wallace et al., 5 Dec 2025).

The experimental environment is the cooperative Overcooked benchmark in JaxMARL, with three layouts: Cramped Room, Asymmetric Advantage, and Coordination Ring. Five partner types are induced via reward shaping: default, pot-focused, plate-focused, serve-focused, and mixed. For each type, a best-response agent $$\pi(Q)=[v_0 \xrightarrow{r_1} v_1 \xrightarrow{r_2}\cdots \xrightarrow{r_L} v_L],$$0 is trained via PPO. Evaluation tracks classification accuracy and episodic return over 400 steps, and compares against Oracle, Static, Random switching, Logistic Regression on $$\pi(Q)=[v_0 \xrightarrow{r_1} v_1 \xrightarrow{r_2}\cdots \xrightarrow{r_L} v_L],$$1, and PLASTIC (Bayesian), with 5 random seeds per layout and mean$$\pi(Q)=[v_0 \xrightarrow{r_1} v_1 \xrightarrow{r_2}\cdots \xrightarrow{r_L} v_L],$$2std reporting (Wallace et al., 5 Dec 2025).

5. Empirical characteristics in Overcooked

The empirical results emphasize both classification quality and downstream control performance (Wallace et al., 5 Dec 2025). ReCollab improves over CoLLAB in all three benchmark layouts and is reported to lie on or very near the Pareto frontier when plotting accuracy-return pairs.

Layout	Accuracy: CoLLAB $$\pi(Q)=[v_0 \xrightarrow{r_1} v_1 \xrightarrow{r_2}\cdots \xrightarrow{r_L} v_L],$$3 ReCollab	Return: CoLLAB $$\pi(Q)=[v_0 \xrightarrow{r_1} v_1 \xrightarrow{r_2}\cdots \xrightarrow{r_L} v_L],$$4 ReCollab
Cramped Room	$$\pi(Q)=[v_0 \xrightarrow{r_1} v_1 \xrightarrow{r_2}\cdots \xrightarrow{r_L} v_L],$$5	$$\pi(Q)=[v_0 \xrightarrow{r_1} v_1 \xrightarrow{r_2}\cdots \xrightarrow{r_L} v_L],$$6
Asym. Advantage	$$\pi(Q)=[v_0 \xrightarrow{r_1} v_1 \xrightarrow{r_2}\cdots \xrightarrow{r_L} v_L],$$7	$$\pi(Q)=[v_0 \xrightarrow{r_1} v_1 \xrightarrow{r_2}\cdots \xrightarrow{r_L} v_L],$$8
Coord. Ring	$$\pi(Q)=[v_0 \xrightarrow{r_1} v_1 \xrightarrow{r_2}\cdots \xrightarrow{r_L} v_L],$$9	$P_1,\dots,P_n$00

The baseline comparisons are heterogeneous. Logistic Regression obtains classification accuracies of $P_1,\dots,P_n$01 in Cramped Room, $P_1,\dots,P_n$02 in Asym. Advantage, and $P_1,\dots,P_n$03 in Coordination Ring, while PLASTIC obtains $P_1,\dots,P_n$04, $P_1,\dots,P_n$05, and $P_1,\dots,P_n$06, respectively. On the control side, Oracle episodic return is $P_1,\dots,P_n$07 in Cramped Room, $P_1,\dots,P_n$08 in Asym. Advantage, and $P_1,\dots,P_n$09 in Coordination Ring, whereas Static yields $P_1,\dots,P_n$10, $P_1,\dots,P_n$11, and $P_1,\dots,P_n$12 (Wallace et al., 5 Dec 2025).

The reported ablations identify a concrete operating regime. Probe length $P_1,\dots,P_n$13 optimally trades off speed versus accuracy at $P_1,\dots,P_n$14, and the number of exemplars $P_1,\dots,P_n$15 exhibits diminishing returns beyond $P_1,\dots,P_n$16. The paper's interpretation is that retrieval-grounded LLM inference achieves a balanced optimal trade-off between fast, accurate classification and high episodic returns (Wallace et al., 5 Dec 2025).

6. Related retrieval-collaboration variants and interpretive issues

In multi-hop question answering, Collab-RAG presents a two-agent RAG pipeline consisting of a white-box SLM $P_1,\dots,P_n$17 as query decomposer, a fixed black-box LLM $P_1,\dots,P_n$18 as context reader and final synthesizer, and a differentiable retriever $P_1,\dots,P_n$19, Dragon-Plus by default (Xu et al., 7 Apr 2025). For a question $P_1,\dots,P_n$20, the SLM emits an ordered list of sub-questions $P_1,\dots,P_n$21; the retriever returns top-$P_1,\dots,P_n$22 passages $P_1,\dots,P_n$23; the LLM reader generates intermediate answers $P_1,\dots,P_n$24; and the final answer is synthesized from all $P_1,\dots,P_n$25. Training proceeds in two stages: warm-up supervised fine-tuning using a decomposition loss $P_1,\dots,P_n$26, then iterative Direct Preference Optimization using black-box feedback $P_1,\dots,P_n$27. Across HotpotQA, 2WikiMQA, MuSiQue, StrategyQA, and Bamboogle, Collab-RAG reports $P_1,\dots,P_n$28 to $P_1,\dots,P_n$29 absolute EM gains over the best baselines, and a fine-tuned 3B SLM decomposer outperforms a frozen 32B LLM-based decomposer on average. The paper explicitly describes this feedback loop between white-box decomposition and black-box reading as "Retrieval Collaboration ('ReCollab')" (Xu et al., 7 Apr 2025).

A separate line packages AC-RAG as ReCollab through three components: a Detector Agent $P_1,\dots,P_n$30, a Resolver Agent $P_1,\dots,P_n$31, and a stateless moderator enforcing a "Dissect-Retrieve-Reflect" cycle (Zhang et al., 18 Sep 2025). The detector performs pre-check and post-check classification and generates sub-questions $P_1,\dots,P_n$32; the resolver drafts rough answers $P_1,\dots,P_n$33, summarizes retrieved documents $P_1,\dots,P_n$34, and produces the final answer $P_1,\dots,P_n$35. The joint training objective combines a detection loss $P_1,\dots,P_n$36, a resolution loss $P_1,\dots,P_n$37, and an adversarial regularizer $P_1,\dots,P_n$38 in a saddle-point objective $P_1,\dots,P_n$39. The default loop budget is $P_1,\dots,P_n$40, with a pre-check threshold $P_1,\dots,P_n$41 and post-check threshold $P_1,\dots,P_n$42. On medical QA, ReCollab-8B reports $P_1,\dots,P_n$43 on MMLU-Med, $P_1,\dots,P_n$44 on PubMedQA, $P_1,\dots,P_n$45 on MedMCQA, $P_1,\dots,P_n$46 on MedQA, and $P_1,\dots,P_n$47 average, corresponding to $P_1,\dots,P_n$48 points over RAG baselines and $P_1,\dots,P_n$49 points over FT-only Llama-3-8B; cross-domain tests on LegalBench and Huawei DevOps report consistent $P_1,\dots,P_n$50 to $P_1,\dots,P_n$51 gains over standard RAG (Zhang et al., 18 Sep 2025).

The main misconception to avoid is that ReCollab names a single frozen architecture. The cited literature does not support that interpretation. Instead, it presents several systems with different task definitions, agent roles, and optimization procedures: moderator-coordinated expert QA over private knowledge graphs, exemplar-grounded LLM teammate typing in Overcooked, white-box/black-box decomposition-and-reading for multi-hop QA, and adversarial Detector/Resolver collaboration in vertical-domain RAG (Hu et al., 2022). This suggests that the unifying idea is collaborative control over retrieval and reasoning under incomplete information, while the concrete instantiation remains task-specific.