Adaptive Collaboration of Argumentative LLMs (ACAL)

• The paper introduces ACAL, a neuro-symbolic framework that transforms unstructured LLM legal outputs into formal, editable argument graphs for verifiable reasoning.
• Adaptive multi-agent collaboration selects specialized legal agents and employs arena-based clash resolution to adjudicate conflicting support and attack arguments.
• Empirical evaluations demonstrate notable performance improvements on legal benchmarks while enabling human contestation through an auditable reasoning workflow.

Searching arXiv for the cited ACAL and related adaptive argumentative LLM papers to ground the article in current literature. Adaptive Collaboration of Argumentative LLMs (ACAL) is a neuro-symbolic framework that integrates adaptive multi-agent collaboration with an Arena-based Quantitative Bipolar Argumentation Framework (A-QBAF) for explainable and contestable legal reasoning. In the formulation introduced for legal reasoning, ACAL dynamically deploys expert agent teams to construct arguments, employs a clash resolution mechanism to adjudicate conflicting claims, and utilizes uncertainty-aware escalation for borderline cases. Its defining move is to transform LLM-generated legal reasoning from loose text into a formal argumentation graph whose final judgment can be recomputed when the graph is modified, rather than leaving the decision embedded in an unstructured debate transcript or a free-form Chain-of-Thought (Cao et al., 21 Feb 2026).

1. Conceptual basis and problem setting

ACAL is motivated by the claim that legal reasoning is not only a predictive task. It must also be justified, verifiable, contestable, and traceable to evidence and legal rules. On this view, a high-performing legal LLM system is insufficient if it cannot expose why a conclusion was reached, identify which arguments support or undermine that conclusion, and permit a user to challenge or revise parts of the reasoning in a principled way (Cao et al., 21 Feb 2026).

The framework is presented against the limitations of three common paradigms. Chain-of-Thought (CoT) provides step-by-step reasoning, but that reasoning is described as free-form text that is hard to verify logically, not directly editable by users, and not mathematically linked to the final decision. Retrieval-Augmented Generation (RAG) improves grounding by retrieving relevant texts, yet the decision logic remains implicit: citations do not by themselves yield a structured argument graph or a contestable inference procedure. Multi-Agent Debate (MAD) introduces diversity of perspectives, but debate transcripts remain unstructured, performance gains are inconsistent, and there is no principled mechanism for propagation, scoring, or contestability (Cao et al., 21 Feb 2026).

This orientation places ACAL within a broader shift from answer generation to structured adjudication. Related work on “Argumentative LLMs” likewise treats the LLM as a generator of structured pros and cons, while the final decision is computed by formal semantics rather than by the model’s raw output stream (Freedman et al., 2024). ACAL extends that line by making collaboration adaptive, by adding arena-style adjudication for close clashes, and by exposing a Human-in-the-Loop contestation workflow (Cao et al., 21 Feb 2026).

2. Adaptive multi-agent workflow

The end-to-end ACAL workflow is explicitly staged. It proceeds by identifying the legal task and case context, selecting specialized legal agents adaptively, retrieving legal evidence via hybrid RAG, generating supporting and attacking arguments, assigning intrinsic strength scores, identifying support/attack/neutral relations between arguments, resolving clashes in close cases, propagating scores through an A-QBAF, deciding Yes/No from the claim score, escalating borderline cases to a Final Judge agent, and then allowing users to inspect and contest the graph before recomputing the verdict (Cao et al., 21 Feb 2026).

The adaptive component begins with a pool of legal agents

$\mathcal{A} = \{a_1, a_2, \ldots, a_n\},$

where each agent $a_i$ is characterized by a tuple $\{r_i, E_i, P_i, S_i\}$ consisting of role, expertise areas, focus priorities, and argument style. The paper lists 10 legal roles grouped into four categories: Adjudication (Judge, Law Clerk / Judicial Clerk), Litigation Advocacy (Private Practice Lawyer, Prosecutor, Public Defender), Advisory Transactional (Corporate Counsel, Compliance Officer, IP Attorney), and Research Support (Legal Analyst, Paralegal) (Cao et al., 21 Feb 2026).

Agent selection is polarity-sensitive. Given a legal task $\mathcal{T}$, context $c$, and claim $\phi$, ACAL selects

$$\mathcal{A}^{+} = \text{Select}(\mathcal{A}, \mathcal{T}, c, \text{support}); \quad \mathcal{A}^{-} = \text{Select}(\mathcal{A}, \mathcal{T}, c, \text{attack}).$$

Here, $\mathcal{A}^{+}$ is the team for constructing supporting arguments and $\mathcal{A}^{-}$ is the team for constructing attacking arguments. The selection function uses an LLM to match expert profiles to the legal case, so the system does not deploy all agents uniformly on every instance (Cao et al., 21 Feb 2026).

Grounding is supplied by a hybrid retrieval module that combines vector search over a legal database for statutes with external web search for recent case law. The top-$k$ relevant passages are aggregated into the evidentiary context $a_i$0. Argument generation for an agent $a_i$1 is then defined as

$a_i$2

where $a_i$3. Each agent typically generates 2–5 arguments, depending on evidence and case complexity (Cao et al., 21 Feb 2026).

3. Formal argumentative structure and arena-based adjudication

ACAL’s symbolic core is an Arena-based Quantitative Bipolar Argumentation Framework,

$a_i$4

where $a_i$5, $a_i$6 is the attack relation, $a_i$7 is the support relation, and $a_i$8 is the base strength function. The central claim is initialized neutrally with $a_i$9, and each generated argument is attached to $\{r_i, E_i, P_i, S_i\}$0 as either support or attack (Cao et al., 21 Feb 2026).

Each generated argument $\{r_i, E_i, P_i, S_i\}$1 receives an intrinsic score $\{r_i, E_i, P_i, S_i\}$2 from an LLM-based strength estimate. The scoring rubric ranges from 0.1–0.2 for incorrect legal analysis or misidentification of key elements up to 0.9–1.0 for exceptional precision with authoritative citations and flawless logic. The intermediate bands distinguish partially correct analysis, sound analysis with minor gaps, and strong analysis with specific facts and correct legal reasoning (Cao et al., 21 Feb 2026).

Inter-argument relation identification is handled either heuristically or semantically. The heuristic mode treats same stance as mutual support and opposite stance as mutual attack, but the paper rejects this as too simplistic for legal reasoning because legally opposing arguments may still be unrelated. In the LLM-based semantic mode, each pair $\{r_i, E_i, P_i, S_i\}$3 is classified as attack, support, or neutral with a confidence score. If the confidence is below 0.6, the relation is demoted to neutral. Relations are made bidirectional, and pairwise analysis is batched with size $\{r_i, E_i, P_i, S_i\}$4 and default $\{r_i, E_i, P_i, S_i\}$5 to reduce API calls (Cao et al., 21 Feb 2026).

A distinctive component is Clash Resolution via Arena Debating Round. When a support argument and an attack argument have similar scores,

$\{r_i, E_i, P_i, S_i\}$6

with default $\{r_i, E_i, P_i, S_i\}$7, the two arguments are sent to an arena debate in which an LLM acting as a legal expert determines which is stronger on the basis of facts and law. Scores are then updated by

$\{r_i, E_i, P_i, S_i\}$8

where $\{r_i, E_i, P_i, S_i\}$9 is the base adjustment magnitude and $\mathcal{T}$0 is the win rate across clashes. The paper reports the best $\mathcal{T}$1 as 0.15 (Cao et al., 21 Feb 2026).

Final strengths are propagated with Quadratic Energy (QE) semantics. For argument $\mathcal{T}$2,

$\mathcal{T}$3

and the equilibrium score satisfies

$\mathcal{T}$4

The claim score then determines the answer by

$\mathcal{T}$5

with default threshold $\mathcal{T}$6 (Cao et al., 21 Feb 2026).

4. Uncertainty-aware escalation and contestability

ACAL does not treat every propagated score as equally stable. The framework introduces uncertainty-aware escalation (UAE) for borderline cases. If the claim score falls in the interval

$\mathcal{T}$7

the system bypasses the standard decision rule and invokes a Final Judge agent. The Final Judge independently reviews the evidence, re-evaluates the legal standards, resolves key conflicts, and outputs a binding decision. The paper motivates this mechanism as a way to address “same-score saturation” and to prevent arbitrary verdicts near the neutral boundary (Cao et al., 21 Feb 2026).

The second major feature is the Human-in-the-Loop (HITL) contestation workflow. ACAL exposes the argument text and stance, the originating agent role, evidence passages, intrinsic score $\mathcal{T}$8, propagated score $\mathcal{T}$9, the local neighborhood in the graph, and a “why it matters” trace for the claim score. These are the objects through which the graph becomes auditable rather than merely inspectable in prose form (Cao et al., 21 Feb 2026).

The paper identifies five contestation types: factual issues, legal rule issues, precedent issues, missing exceptions/defenses, and procedural fairness concerns. Users may accept or reject arguments, edit arguments, add missing arguments, adjust argument strength, and adjust support/attack relations when mis-specified. Accepted edits produce an updated structure

$c$0

after which the same propagation is rerun to obtain $c$1 and $c$2. The system also records who changed what and how the change affected $c$3, providing an audit log for traceability and accountability (Cao et al., 21 Feb 2026).

A common misconception is that contestability is equivalent to attaching an explanation after prediction. ACAL rejects that interpretation. In the framework, contestation is an intervention on the reasoning graph that can materially change the result when the graph is recomputed. This is structurally closer to the claim-verification setting of ArgLLMs, where a user may add arguments, remove arguments, or modify argument strengths before the final decision is produced by formal semantics (Freedman et al., 2024).

5. Empirical evaluation and ablation findings

The reported evaluation uses LegalBench and focuses on two tasks. Hearsay predicts whether a statement is hearsay under FRE 801(c), with labels hearsay and not_hearsay. Courts uses the Learned Hands taxonomy with labels yes and no. Two backbone models are tested: Gemini-2.5-Flash-Lite and Gemini-2.5-Flash. Baselines are SP (few-shot standard prompting), CoT, RAG, and MAD. For RAG, legal texts are embedded with OpenAI text-embedding-3-large and top-$c$4 chunks are retrieved. The reported metrics are Accuracy (Acc), Precision (Prec), Recall (Rec), and Macro-F1 (F1) (Cao et al., 21 Feb 2026).

On Gemini-2.5-Flash-Lite, the Learned Hands Courts results for ACAL are Acc 70.8, Prec 71.2, Rec 70.8, and F1 70.7; the paper states that this is the best overall among compared methods. On Hearsay, ACAL achieves Acc 74.5, Prec 77.1, Rec 76.3, and F1 74.4, described as best or tied-best on key metrics and best on precision/recall/F1. On Gemini-2.5-Flash, ACAL reaches Acc 75.5, Prec 76.4, Rec 75.5, and F1 75.3 on Learned Hands Courts, and Acc 76.7, Prec 77.1, Rec 77.6, and F1 76.7 on Hearsay; the paper characterizes the latter as the best recall and F1 among the reported methods (Cao et al., 21 Feb 2026).

The ablation study is particularly important for understanding the framework’s internal mechanisms. On Gemini-2.5-Flash-Lite for Hearsay, a vanilla-ish setting gives 64.9% accuracy. Adding CR only yields 72.8% accuracy, a +7.9 points gain. Using UAE alone gives 62.8% accuracy, which the paper interprets as evidence that uncertainty handling is unreliable if the argument scores themselves are not well calibrated. The full ACAL configuration with both CR and UAE gives 74.5% accuracy and 74.4% F1. Performance rises as $c$5 increases from 0.05 to 0.15, then drops when $c$6 becomes too large, suggesting that over-aggressive score updates cause volatility (Cao et al., 21 Feb 2026).

These results support the paper’s central claim that formal conflict adjudication matters more than simply adding multi-agent interaction. The empirical pattern suggests that ACAL’s gain is not reducible to debate alone; rather, it depends on coupling adaptive expert selection with explicit score calibration and symbolic propagation (Cao et al., 21 Feb 2026).

6. Relation to adjacent research and broader significance

ACAL sits at the intersection of several strands of recent LLM research. One strand uses formal computational argumentation without adaptiveness. MQArgEng constructs a Dung-style abstract argumentation framework $c$7, computes the grounded extension or preferred extension with ASPARTIX, and feeds acceptable arguments back into the LLM; on MT-Bench it reports MQInstruct 5.96 → MQArgEng 6.09, a +0.13 absolute points or +2.18% relative improvement (Castagna et al., 2024). Another strand, ArgLLMs, builds quantitative bipolar frameworks from support and attack arguments and resolves them with DF-QuAD or QEM semantics for explainable and contestable claim verification (Freedman et al., 2024). ACAL inherits the formal-argumentation logic of these systems but adds adaptive expert deployment, arena-based clash resolution, and uncertainty-aware escalation (Cao et al., 21 Feb 2026).

A second strand emphasizes adaptive collaboration without a formal editable reasoning graph. In medicine, MDAgents routes cases by a moderator-assigned complexity label $c$8 to a Solo PCP agent, MDT, or ICT, reporting the highest accuracy in 7 out of 10 benchmarks and up to 4.2% improvement over both solo and group/static settings (Kim et al., 2024). In general reasoning, DiMo selects between Divergent Thinking Mode and Logical Thinking Mode, uses four rounds of debate, and reports its largest gains on math under an Exact Match (EM) evaluation regime (He et al., 18 Oct 2025). In document understanding, an adaptive orchestration framework combines dynamic task routing, bidirectional feedback, and parallel agent evaluation, with an evaluator selecting

$c$9

and reporting improvements in factual coverage, coherence, and efficiency over static and partially adaptive baselines (Xia et al., 22 Jul 2025). These systems share ACAL’s commitment to task-aware collaboration, but not its insistence on a formally contestable argument graph.

A third strand studies the learning of collaboration itself. ACC-Collab trains an actor-agent and critic-agent with guided-debate trajectories and Direct Preference Optimization (DPO), rather than assuming collaboration will emerge from off-the-shelf models (Estornell et al., 2024). HILA goes further by learning a metacognitive policy over EVAL, CREATE, and DEFER, combining Group Relative Policy Optimization (GRPO) with continual learning from expert feedback in a Dual-Loop Policy Optimization scheme (Yang et al., 9 Mar 2026). A plausible implication is that future ACAL systems could combine formal argumentation graphs with learned collaboration policies, rather than relying on fixed prompting and hand-specified escalation rules.

In that broader landscape, ACAL is best understood as a specific legal instantiation of a wider research program: adaptive collaboration should not merely increase the number of agents or produce longer deliberation traces. It should determine which expert roles to activate, formalize how supporting and attacking claims interact, resolve close conflicts through explicit adjudication, and preserve a reasoning object that users can audit and change. Within the current literature, that combination of adaptiveness, formal argumentation, and contestability is the framework’s distinctive contribution (Cao et al., 21 Feb 2026).