Judgmental Rationales Generator (JRG)

• JRGs are systems that produce explicit, judgment-conditioned rationales tied to model outputs for enhanced interpretability.
• They employ diverse architectures such as iterative refinement, rationale-first decision making, and selective token extraction.
• Evaluation integrates outcome accuracy with process metrics like F-Score and RC to ensure models are right for the right reasons.

Searching arXiv for the cited papers and closely related JRG formulations. arXiv search: "judgmental rationales generator" Judgmental Rationales Generator (JRG) denotes a class of systems that generate rationales explicitly tied to a model judgment—such as a situation label, preference verdict, legal outcome, moral assessment, or visual answer—and that are trained or evaluated on the quality of the reasoning process rather than on outcome labels alone. Across the literature surveyed here, the term is used in multiple, partly overlapping senses: as the LLM-based component that constructs positive and negative glimpse and gaze rationales in open-vocabulary grounded situation recognition (Cai et al., 19 Jul 2025); as a Generative Reward Model (GenRM) specialized to produce rationales before emitting a preference decision (Lai et al., 6 Feb 2026); and as a broader generator–predictor architecture in which extracted text spans, counterfactual edits, or free-form explanations serve as the rationale for a prediction (Lei et al., 2016). The unifying idea is that a judgment should be accompanied by reasons that are sufficiently specific, grounded, and usable for training, auditing, or downstream control.

1. Definitions and conceptual scope

In the recent literature, “judgmental rationales” are not generic explanations. In the open-vocabulary grounded situation recognition setting, they are free-form textual explanations describing the scene that have been evaluated and refined under an LLM “single-answer grading” regime, with explicit distinctions between glimpse rationales, gaze rationales, and positive versus negative variants (Cai et al., 19 Jul 2025). In the reward-modeling literature, a JRG is defined more narrowly as a GenRM specialized to produce judgmental rationales that faithfully track the valid decision basis, so that a rationale is generated first and a discrete verdict is then parsed from it (Lai et al., 6 Feb 2026).

A second major lineage defines rationale more structurally. In “Rationalizing Neural Predictions,” a rationale is a discrete selection vector $z \in \{0,1\}^n$ that extracts short, coherent text fragments from the input, and the predictor is trained on $x \odot z$ rather than on the full input (Lei et al., 2016). This formulation treats rationales as sufficient evidence subsets rather than as natural-language prose. A related normative strand ties rationale generation to explicit rule systems: in norm-governed agents, justifications are generated from object-oriented temporal logic rules and counterfactual trajectory comparisons, allowing the system to answer factual and “why” questions about decisions and rule violations (Kasenberg et al., 2019).

These usages share a family resemblance rather than a single canonical definition. The common property is that the rationale is judgment-conditioned and is expected to preserve or expose the decisive basis of the output. This distinguishes JRGs from generic captions, unconstrained chain-of-thought, or post-hoc attention visualizations. In the Ov-GSR formulation, for example, generic captions are said to lack role conditioning, judgmental calibration, and the positive/negative polarity needed by the alignment module (Cai et al., 19 Jul 2025).

2. Architectural patterns and generation workflows

One prominent pattern is iterative judged refinement. In the MIPD framework for open-vocabulary grounded situation recognition, JRG receives an image $I$, a base-set situation label $s = \{v, F_v\}$ during training, and pseudo rationales from a frozen teacher MLLM. The judge rates each rationale in $[1,10]$ for relevance, accuracy, and detail with respect to $s$; if the score is below the threshold $N = 8$, the rationale is refined and rescored until the threshold is reached, producing positive glimpse and gaze rationales, after which near-miss negative rationales are generated by modifying activity, entities, and attributes with semantically similar but incorrect choices (Cai et al., 19 Jul 2025).

A second pattern is rationale-first judgment generation. In R-Align, the basic object is a preference tuple $(x, y_a, y_b, l)$, and the GenRM/JRG first samples a rationale

$o \sim \mathcal{R}(\cdot \mid x, y_a, y_b),$

then parses the verdict as $\hat l = f(o)$ (Lai et al., 6 Feb 2026). This ordering is also central to self-rationalizing judges for fine-grained scoring: the model produces rationale and score jointly, but the score is conditioned on the rationale, and iterative DPO training is performed on self-generated judgment pairs (Trivedi et al., 2024).

A third pattern optimizes internal “thinking traces.” Think-J trains a generative judge with an “analysis $x \odot z$0 verdict” schema, initializes it on LIMJ707 with clipped DeepSeek-R1 traces, and then improves the traces by offline critic-guided SFT+DPO or online GRPO-based RL (Huang et al., 20 May 2025). Here the rationale is neither a post-hoc explanation nor a token mask; it is the primary reasoning trajectory that precedes the final verdict line.

Selective-rationalization systems adopt yet another workflow. A selector generates a binary mask or top-$x \odot z$1 subset of tokens, and a predictor is trained only on the selected rationale sequence. Noise Injection places a perturbation module between generator and predictor, randomly replacing rationale tokens during training so that implausible but loss-minimizing “communication schemes” become unstable (Storek et al., 2023). MGR keeps the same generator–predictor principle but replaces a single generator with multiple generators trained jointly against one shared predictor, with the stated aim of addressing both spurious correlation and degeneration (Liu et al., 2023).

3. Training objectives and alignment signals

JRG training objectives vary by domain, but they consistently add explicit structure beyond plain outcome supervision. In rationale extraction for text prediction, the canonical joint objective is

$x \odot z$2

where $x \odot z$3 encourages shortness and $x \odot z$4 encourages coherence by penalizing rationale boundaries (Lei et al., 2016). In multimodal distillation for Ov-GSR, the training target is

$x \odot z$5

where $x \odot z$6 explicitly penalizes similarity to negative rationales and the aligned teacher streams are distilled into the student (Cai et al., 19 Jul 2025).

In preference modeling, a central issue is that label correctness can mask reasoning failure. R-Align formalizes this with a MetaRM decision $x \odot z$7 and defines the rationale-aware reward

$x \odot z$8

This reward is optimized with PPO and differs from outcome-only RLVR, which rewards any correct label regardless of rationale quality (Lai et al., 6 Feb 2026). The later hybrid-signal formulation in “Outcome Accuracy is Not Enough” uses

$x \odot z$9

so that correct reasoning is a necessary condition for receiving reward, and computes rationale reward by Average Precision over an ordered list of atomic reasons matched one-to-one against human atomic reasons (Wang et al., 4 Feb 2026).

Other work enriches the supervision signal at the data level rather than only in the optimizer. RDPO adds an auxiliary rationale likelihood term to DPO,

$I$0

where $I$1 is a machine-generated rationale attached to a preference pair (Just et al., 2024). This design treats rationales as additional training signal for preference discrimination rather than as an auxiliary artifact.

4. Evaluation methodologies and reported findings

Evaluation protocols for JRGs now extend well beyond end-task accuracy. They typically distinguish outcome quality, rationale fidelity, plausibility, human preference, and downstream control effectiveness. In reward-modeling work, the key shift is from label accuracy alone to metrics that explicitly measure whether a model is “right for the right reasons” (Lai et al., 6 Feb 2026, Wang et al., 4 Feb 2026).

Setting	Metric or benchmark	Reported result
R-Align	Pearson correlation with downstream RLHF	F-Score: 0.947 / 0.924 / 0.963 vs L-Acc: 0.366 / 0.382 / 0.220 on HelpSteer3 / RewardBench2 / PPE-Preference
Outcome Accuracy is Not Enough	RM-Bench / JudgeBench	87.1% on RM-Bench and 82% on JudgeBench
Ov-GSR MIPD	Ov-SWiG test Top-1	Baseline 36.78/29.44, direct distillation 38.76/31.54, full MIPD 41.96/34.11 for verb/value
Think-J	RewardBench Overall	90.5 for Think-J-Qwen-2.5-32B
Noise Injection	ERASER Movie Review	BERT-A2R+NI Accuracy 85.7 and F1 38.6
MGR	BeerAdvocate Palate (20%)	F1 57.3 vs Inter-RAT 36.4

R-Align introduces three process-sensitive metrics: Label Accuracy,

$I$2

Spurious Correctness,

$I$3

and F-Score,

$I$4

The reported result is that F-Score is strongly predictive of downstream RLHF quality, whereas L-Acc is weak and saturated (Lai et al., 6 Feb 2026). The later RC framework defines

$I$5

with strict one-to-one matching between model and human atomic reasons, and reports that RC discriminates frontier models better than outcome accuracy, with evaluator-insensitivity statistics of $I$6 and RMSE $I$7 across evaluators (Wang et al., 4 Feb 2026).

Outside reward modeling, rationale plausibility and sufficiency remain central. Noise Injection reports improvements on ERASER Movie Review, MultiRC, FEVER, and a new USR Movie Review benchmark, explicitly separating plausibility metrics such as F1 and IOU-F1 from faithfulness metrics such as comprehensiveness and sufficiency (Storek et al., 2023). Think-J evaluates free-form judgment traces on RewardBench, RMBench, and Auto-J agreement, and reports that formatting reward is critical, that clipped R1 traces improve stability, and that GRPO and Reinforce++ outperform PPO for rationale optimization (Huang et al., 20 May 2025).

5. Domain-specific instantiations

In multimodal situation recognition, JRG is tightly integrated with representation learning. The Ov-GSR version generates four rationale types—$I$8, $I$9, $s = \{v, F_v\}$0, and $s = \{v, F_v\}$1—and these are aligned with scene-aware and instance-perception prompts attached to frozen teacher features. The resulting aligned teacher streams are then distilled into a compact CLIP-based student, and the reported gains extend across seen, rare, and unseen situations as well as zero-shot unseen interactions on HICO-DET (Cai et al., 19 Jul 2025).

In law, JRG-like systems emphasize grounding in institutional authority. NyayaRAG frames legal judgment prediction under Indian common law as decision plus explanation generation conditioned on factual case descriptions, statutory provisions, and retrieved prior cases, with prompts structured in IRAC form and citation verification used as a hallucination mitigation mechanism (Nigam et al., 1 Aug 2025). The Japanese Tort-case Dataset formalizes a different legal version: the rationale is the subset of plaintiff and defendant alleged claims that the court accepted, yielding a joint problem of tort prediction and rationale extraction, and the strongest reported baseline is a multi-task hierarchical Transformer with 0.674 claim-level rationale extraction accuracy and 0.683 tort-prediction accuracy on test (Yamada et al., 2023).

Norm-governed agents provide a rule-based contrast. “Generating Justifications for Norm-Related Agent Decisions” derives natural-language justifications from rules expressed in a fragment of object-oriented temporal logic, compares actual and counterfactual trajectories under lexicographically prioritized violation costs, and realizes explanations in English via rule-based clause generation. The reported human evaluation found main effects of condition on intelligibility, mental model, and trust, with the experimental system outperforming both content and surface baselines on intelligibility and mental model, and outperforming the surface baseline on trust (Kasenberg et al., 2019).

Vision work has also developed non-textual rationale variants. “Generative Visual Rationales” defines the rationale for an image prediction as the difference between an original decoding and a counterfactual decoding obtained by optimizing a latent representation to reduce the target prediction while keeping the decoded image close to the original. On chest radiography, the reported linear regressor on latent codes achieves AUC $s = \{v, F_v\}$2 at the clinically used BNP threshold of 100 ng/L (Seah et al., 2018). In VQA, rationale generation is posed as a conditional text-generation task tied to the model’s predicted answer; joint ViLBERT+GPT-2 training improves both rationale metrics and VQA accuracy, with human raters preferring the jointly trained rationales 80% of the time by majority vote (Ayyubi et al., 2020).

6. Limitations, misconceptions, and open questions

A recurrent misconception is that outcome accuracy alone is an adequate proxy for rationale quality. Multiple papers reject this view explicitly. R-Align shows that high Spurious Correctness is associated with policy degeneration under RLHF, and “Outcome Accuracy is Not Enough” argues that GenRMs and LLM-as-a-Judge systems can exhibit deceptive alignment by producing correct judgments for incorrect reasons (Lai et al., 6 Feb 2026, Wang et al., 4 Feb 2026). In both cases, the core objection is not merely philosophical: the reasoning defect is reported to matter for downstream optimization.

A second misconception is that any explanation-like text is a judgmental rationale. The multimodal literature distinguishes situation-conditioned rationales from generic captions; the selective-rationalization literature distinguishes sufficient evidence subsets from free-form prose; and the persuasiveness literature shows that a rationale can be compelling without being faithful. In pairwise argument ranking, persuasiveness did not significantly correlate with ranking accuracy, and contrastive reasoning and length were more influential than novelty in explaining persuasiveness (2406.13905). This suggests that rhetorical quality and epistemic faithfulness should not be conflated.

Several failure modes recur across domains. Think-J reports reward hacking under margin-based absolute scoring and notes that formatting rewards are crucial (Huang et al., 20 May 2025). PB&J states that no ground-truth rationales are available in standard persona datasets, so generated rationales remain plausible rather than verified reflections of user reasoning (Joshi et al., 25 Apr 2025). Defeasible moral reasoning work identifies cultural bias, everyday-action coverage limits, and the need for gated release, warning that outputs are descriptive and not prescriptive moral advice (Rao et al., 2023). Legal datasets likewise emphasize assistant use rather than replacement of judges or attorneys (Yamada et al., 2023).

The broader trajectory of the field suggests three durable research directions. First, evaluation is shifting from label correctness to process-sensitive metrics such as F-Score, S-Corr, and RC. Second, training increasingly combines outcome supervision with rationale-aware rewards, filtering, or data augmentation rather than treating rationales as purely explanatory by-products. Third, JRGs are becoming domain-specific: legal systems ground in statutes and accepted claims, multimodal systems ground in scene structure and instance cues, and selective rationalizers ground in explicit evidence subsets. A plausible implication is that “JRG” will remain less a single architecture than a design principle: the judgment mechanism and the rationale mechanism are coupled tightly enough that the rationale can shape learning, auditing, and deployment behavior rather than merely narrate the result.