Inquizzitor: Inquisitive Agent for Legal Dialogue

• Inquizzitor is a dual-agent hierarchical RL system designed for inquisitive legal dialogue in U.S. Supreme Court oral arguments.
• It employs an appraisal agent and a hierarchical dialogue agent to generate probing, context-aware questions that mimic judicial interrogation tactics.
• Reward design balances goal relevance, novelty, and succinctness to ensure effective fact-finding and strategic dialogue control.

Inquizzitor is a dual-agent hierarchical reinforcement learning system introduced as an Inquisitive Conversational Agent (ICA) for U.S. Supreme Court oral arguments. It is designed for settings in which the agent must actively interrogate, probe, and steer an interaction to extract information in service of its own objective rather than merely respond to user requests. In the reported implementation, Inquizzitor imitates judicial questioning patterns such as probing for clarification, testing hypotheticals, challenging assumptions, demanding concise answers, and pressing on inconsistencies, and is presented as a first step toward broader high-stakes, domain-specific applications (Lin et al., 13 May 2026).

1. Motivation and problem setting

The system is motivated by a limitation of most dialogue systems: they are user-driven, assuming that the user controls the agenda and that the system’s role is to fulfill requests. The legal setting of U.S. Supreme Court oral arguments is treated as a contrasting regime. There, the justice controls turn-taking and topic shifts, the interaction is non-cooperative and strategically adversarial, and the objective is not assistance but fact-finding, contradiction spotting, doctrinal testing, and clarification (Lin et al., 13 May 2026).

This framing makes inquisitive dialogue distinct from collaborative and negotiation dialogue. The paper emphasizes that a useful agent in this setting must learn when to keep probing, what kind of question to ask, and when to move on. It also notes that oral-argument transcripts are often long and complex, frequently exceeding 5,000 tokens, so the problem is not only single-turn question generation but long-horizon strategic control of dialogue. A plausible implication is that the central difficulty is policy formation over dialogue trajectories, not only domain language modeling.

2. Inquisitive Conversational Agents and the legal role of Inquizzitor

An Inquisitive Conversational Agent is defined as an agent that actively probes for critical information, pursues its own objective rather than the user’s, evaluates whether responses are sufficient, evasive, or incomplete, and strategically asks follow-up questions to uncover information needed to reach a conclusion (Lin et al., 13 May 2026).

Inquizzitor instantiates this idea in the domain of U.S. Supreme Court oral arguments. Its task is to simulate the justice side of the interaction, generating probing, context-aware questions that help uncover information relevant to the case conclusion. The intended behavior is justice-like rather than assistant-like: the system should ask, evaluate, press, refine, and then move on. The paper therefore treats the agent not as a reactive utterance generator but as an information-seeking policy.

The questioning repertoire is explicitly aligned with judicial practice. The system is designed to probe for clarification, test hypotheticals, challenge assumptions, demand concise answers, and press on inconsistencies. In this sense, the legal domain is not incidental; it supplies both the adversarial structure and the evaluative notion of success.

3. Dual hierarchical reinforcement learning architecture

The architecture consists of two cooperating RL agents, each with its own policy: an Appraisal Agent and a Hierarchical Dialogue Agent (Lin et al., 13 May 2026).

The Appraisal Agent evaluates the attorney’s most recent response. Its function is to decide whether the response has satisfied the justice or whether more probing is warranted. The paper discretizes appraisal types from Supreme Court transcripts, including: sense ambiguity, find deviation, find redundancy, spot weakness, identify flaws, identify chances, keep challenging, dive deeper, and otherwise. The appraisal policy is written as

$$p(s) = \arg\max_p Q_{\text{Appraisal}}(s, p; \theta_{\text{App}}).$$

The appraisal output is treated as an internal state variable and concatenated with the dialogue state:

$S_{\text{aug}} = \text{concat}(s_t, p_t).$

The Hierarchical Dialogue Agent decides what to ask and how to ask it. Legal questioning is represented as a three-level taxonomy. Level 1 selects a coarse dialogue act such as Questioning, Hypothesis Testing, or Declaration. Level 2 selects a subtype such as Clarification, Probing, or Comparison. Level 3 selects the final fine-grained act or utterance realization. This hierarchy separates strategic choice from surface realization.

The overall interaction is formulated as an MDP,

$M = (S, A, R, \gamma),$

and the transition tuple is augmented with appraisal:

$D \sim (s_t, p_t, a_t, r_t, s_{t+1}).$

Within the operational loop, the appraisal agent first judges the previous attorney answer, the dialogue agent then selects a strategic move conditioned on the augmented state, and the chosen action hierarchy is passed to an LLM prompt template to generate the final justice utterance.

4. Reward design, hierarchy modeling, and optimization

The reward is based on how effectively the justice’s utterance elicits useful attorney responses. It combines three components: goal relevance, novelty, and succinctness (Lin et al., 13 May 2026).

The goal-relevance reward measures whether the attorney’s reply contains legally relevant information aligned with the case conclusion. The paper uses Llama-3-8B as a semantic similarity evaluator:

$$R^{\text{goal}}_{t+1}(s_t, u_t) = \max_i \text{sim}(C[i], u_{t+1}).$$

The novelty reward encourages eliciting information not previously discussed, using an adapted Expectation-Adjusted Distinct (EAD) metric:

$$R^{\text{novel}}_{t+1}(s_t, u_t) = \text{EAD}(\cdot).$$

The succinctness reward favors brief, direct attorney answers:

$$R^{\text{clarity}}_{t+1}(s_t, u_t) = -\log(|u_{t+1}|).$$

The implementation weights are reported as 0.2 for relevance, 0.7 for novelty, and 0.1 for succinctness, making novelty the dominant component.

The action hierarchy is embedded in Poincaré hyperbolic space, with objective

$$C = \sum_{(u,v)\in D} \log \frac{e^{-d(u,v)}}{\sum_{v' \in N(u)} e^{-d(u,v')}}.$$

This is used to preserve the tree-structured organization of dialogue acts.

Both agents are trained with Double DQN. For the appraisal agent, the target is

$$Y_{\text{App}} = r + \gamma Q_{\text{App}}(s', \arg\max_{p'} Q_{\text{App}}(s', p'; \theta_{\text{App}}); \theta^-_{\text{App}})$$

with loss

$$C_{\text{App}} = \mathbb{E}_{(s,p,r,s')\sim D}\left( Q_{\text{App}}(s,p;\theta_{\text{App}}) - Y_{\text{App}} \right)^2.$$

For the dialogue agent, the target is analogous:

$S_{\text{aug}} = \text{concat}(s_t, p_t).$0

The hierarchy is regularized through a consistency condition,

$S_{\text{aug}} = \text{concat}(s_t, p_t).$1

and loss

$S_{\text{aug}} = \text{concat}(s_t, p_t).$2

Because training is offline, the system also uses a conservative term to reduce overestimation:

$S_{\text{aug}} = \text{concat}(s_t, p_t).$3

$S_{\text{aug}} = \text{concat}(s_t, p_t).$4

The final objectives are

$S_{\text{aug}} = \text{concat}(s_t, p_t).$5

$S_{\text{aug}} = \text{concat}(s_t, p_t).$6

with coefficients $S_{\text{aug}} = \text{concat}(s_t, p_t).$7, $S_{\text{aug}} = \text{concat}(s_t, p_t).$8, and $S_{\text{aug}} = \text{concat}(s_t, p_t).$9 controlling regularization strength.

5. Dataset, task formulation, and evaluation protocol

The system is trained and evaluated on a public dataset of U.S. Supreme Court oral argument transcripts from Oyez spanning 1955–2023 (Lin et al., 13 May 2026). Each case includes the case name, argument date, speaker IDs, background of the case, argued question, full dialogue transcript, and final conclusion. The reported table summary includes about 9,121.5 cases, an average of 198.6 turns per case, and an average of 45.9 words per utterance. The data is organized across domains including regulatory, civil rights, criminal, intellectual property, commerce, labor, immigration, environment, and others.

The task is to generate the justice-side questioning sequence so as to uncover information relevant to the case conclusion. Training and evaluation are offline, and data splits are based on the year of argument.

Evaluation uses both LLM-based scoring and human judgment. Four single-utterance metrics are reported on a 1–5 scale:

• Conformity Score (CS): judicial style and formality
• Progression Score (PS): whether the exchange moves forward
• Outcome Relevance Score (OS): consistency with legal objective
• Probing Effectiveness Score (PES): how well the utterance elicits new information

Two multi-turn metrics are also used. The Coverage Score measures how many topics from the original transcript are covered in the simulated debate:

$M = (S, A, R, \gamma),$0

The Marginal Relevance (MR) Score, based on MMR, balances staying on-topic with introducing new information:

$M = (S, A, R, \gamma),$1

with $M = (S, A, R, \gamma),$2.

The baselines are Vanilla Llama3, SFT Llama3, SaulLM-7B, Hudeček et al., VaRMI, and ArCHer.

6. Reported performance, ablations, and limitations

The proposed method achieves the best reported performance across the main metrics. The Table 1 results are: CS 4.01, PS 3.98, OS 4.89, PES 4.47, and Overall 4.34 (Lin et al., 13 May 2026). The paper states that this exceeds the strongest baseline overall, VaRMI, at 4.15.

The gain is described as especially large on PES, which the paper interprets as evidence that the appraisal mechanism improves the agent’s ability to ask effective probing questions. The method also leads in coverage and marginal relevance across dialogue rounds. The paper further notes that SaulLM-7B, despite being trained on legal data, is outperformed, and interprets this as evidence that the task depends on dialogue strategy rather than legal language modeling alone.

The ablation study removes four components: the appraisal agent, succinct reward, novelty reward, and goal-relevance reward. The full model has Overall 4.34. The ablations score 4.27 without the appraisal agent, 4.31 without succinct reward, 4.29 without novelty reward, and 4.28 without goal-relevance reward. The reported pattern is that every removed component lowers at least one metric, with the paper highlighting the role of appraisal in probing, novelty in uncovering fresh information, relevance in keeping the exchange tied to the legal objective, and succinctness in supporting direct judicial-style questioning.

Human evaluation reports the same qualitative pattern. The full model achieves the best overall human score, 4.53, outperforming SaulLM-7B and all ablations.

The paper is explicit about several limitations. The system depends heavily on the capability of the underlying LLM used for utterance generation. The reward design and action taxonomy are dataset/domain-specific and would need redesign for other settings. Offline RL quality depends on the dataset policy, so low-quality trajectories may reduce the effectiveness of regularization. The simulated justice responses are generated via prompting rather than direct end-to-end grounding in a real courtroom environment.

7. Position within the broader “inquisitive” research landscape

Inquizzitor belongs to a broader family of work that treats questioning, inquiry, and information-seeking as first-class computational objectives rather than as side effects of answer generation. In that landscape, its defining property is that inquiry is cast as dialogue policy learning under an explicit institutional role: the system acts like a justice whose objective is to extract legally relevant information (Lin et al., 13 May 2026).

A distinct but related direction is the evaluation of scientific inquiry loops. InquiTree formalizes scientific inquiry as interactive Research Trees, representing hypothesis formulation, study design, result interpretation, and belief updating as a dependency-aware process. Its results identify “Erosion of Marginal Capabilities” and “cognitive tunneling” in long-horizon interactions, suggesting that maintaining critical evaluation over sustained inquiry is difficult for current agents (Cui, 8 Jun 2026). This suggests that Inquizzitor’s appraisal-and-policy decomposition is part of a wider effort to make inquiry controllable and measurable rather than leaving it implicit in generic prompting.

Another related line studies the generation of inquisitive questions. “What makes a question inquisitive?” argues that inquisitiveness is a semantic and discourse property tied to types such as explanation, elaboration, background, definition, instantiation, and forward-looking information, rather than to WH-word choice alone (Gao et al., 2022). Inquizzitor differs in operating at the level of multi-turn strategic questioning rather than single-turn question generation, but both lines share the view that productive inquiry depends on the type of missing information.

The name also overlaps with a very different RL setting. In broader discussion, the Inquisitive Reinforcement Learner (Inq) has been referred to as “Inquizzitor”; that agent explores in proportion to expected KL-based information gain and is shown to be strongly asymptotically optimal in all computable probabilistic environments under a bounded-horizon assumption (Cohen et al., 2019). The shared terminology reflects a common emphasis on information-seeking, but the two systems address different problem classes: general-environment exploration in one case, and legal dialogue management in the other.