PsyJev：通过系统1判断进行心理分析

# PsyJev: Psychological Analysis via System-1 Judgment
Source: [https://arxiv.org/html/2610.36689](https://arxiv.org/html/2610.36689)
Fangyuan LiuAffiliation:State Key Laboratory of Cognitive Intelligence,School of Computer Science and Technology,University of Science and Technology of China,Hefei,Chinaemail:[liufangyuan@mail\.ustc\.edu\.cn](mailto:[email protected])Sirui ZhaoAffiliation:State Key Laboratory of Cognitive Intelligence,School of Computer Science and Technology,University of Science and Technology of China,Hefei,Chinaemail:[siruit@ustc\.edu\.cn](mailto:[email protected]),Xiaohao WangAffiliation:State Key Laboratory of Cognitive Intelligence,School of Computer Science and Technology,University of Science and Technology of China,Hefei,Chinaemail:[wangxh236@mail\.ustc\.edu\.cn](mailto:[email protected]),Guo ChengAffiliation:Computer Science,Beijing University of Technology,Beijing,Chinaemail:[chengguo@emails\.bjut\.edu\.cn](mailto:[email protected]),Peng ZhangAffiliation:School of Information Science and Technology,University of Science and Technology of China,Hefei,Chinaemail:[zppeng@mail\.ustc\.edu\.cn](mailto:[email protected])andEnhong ChenAffiliation:State Key Laboratory of Cognitive Intelligence,School of Computer Science and Technology,University of Science and Technology of China,Hefei,Chinaemail:[cheneh@ustc\.edu\.cn](mailto:[email protected])

© none

###### Abstract\.

Self\-disclosures on the Web support diverse psychological analyses, from recognizing expressed experiences to identifying symptom\-relevant information\. We study a shared, question\-conditioned decision interface for this heterogeneous problem in content analysis\. We organize PsyJudgeBench, comprising 13 annotated datasets and 82 questions across six psychological areas, while preserving source\-specific criteria and masks for unobserved labels\. We introduce PsyJev, a System\-1 model that directly produces distributions over answers specified in each request\. Joint supervised training establishes the shared judgment function, followed by Judgment\-Anchored RLCD \(JA\-RLCD\), which adapts logit\-space reward optimization to heterogeneous psychological judgments through balanced supervision, selective retention of correct SFT predictions, and question\-group reward optimization\. The Qwen3\.5\-based 2B, 4B, and 9B models achieve 64\.33, 66\.21, and 67\.56 average test F1, respectively\. The 9B model exceeds the strongest external comparison by 2\.54 points\. Controlled 2B comparisons examine the training stages and anchored refinement\. Further analyses examine question\-conditioned judgment, inference cost, and probability quality, complemented by qualitative comparisons across models\.

###### Keywords:

Psychological Analysis, Social Web, System\-1 Models, Large Language Models, Psychological Analysis Benchmark

## 1\.Introduction

Social media, online forums, and Web communities continuously produce spontaneous accounts of personal experiences, emotional difficulties, relationships, and everyday stress\. These self\-disclosures provide a rich source of Web content for psychological analysis, supporting tasks ranging from recognizing expressed experiences to identifying symptom\-relevant information and questionnaire answerability\. However, analyzing the same content for different purposes requires distinct judgments\. A post may mention a symptom without expressing it, or contain symptom\-relevant information without providing enough detail to answer a questionnaire item\. Such distinctions matter when organizing and analyzing Web content across different psychological tasks, where similar textual expressions may need to be interpreted under different decision criteria\. This motivates a shared framework for addressing a challenge in psychological analysis of Web content: distinguishing question\-specific judgments across heterogeneous datasets drawn from the social Web while preserving their source\-specific decision criteria\.

Research on web\-based psychological analysis has produced specialized datasets and models for diverse targets, from symptom relevance and questionnaire answerability to stressors and interpersonal risk factors\([Hengle et al\., 2024](https://arxiv.org/html/2610.36689#bib.bib1);[Pérez et al\., 2023](https://arxiv.org/html/2610.36689#bib.bib2);[Gupta et al\., 2022](https://arxiv.org/html/2610.36689#bib.bib7);[Mauriello et al\., 2021](https://arxiv.org/html/2610.36689#bib.bib8);[Garg et al\., 2023c](https://arxiv.org/html/2610.36689#bib.bib5)\)\. Recent work uses multi\-task learning and instruction tuning to bring these tasks into shared frameworks\. For instance, MentaLLaMA and the IMHI benchmark unify mental health analysis by generating joint predictions and text explanations\([Yang et al\., 2024](https://arxiv.org/html/2610.36689#bib.bib9)\)\. The constituent tasks retain distinct label semantics, answer spaces, and decision criteria\. A shared prediction interface therefore needs to preserve these distinctions while supporting categorical evaluation\.

Building such a framework poses three main challenges\. First,*heterogeneous judgment semantics*: different analytical questions may require distinct criteria and answer sets, even when applied to the same text\. This calls for predictions conditioned on the question rather than the text alone\. Second,*fragmented supervision*: existing datasets differ in annotation targets, label schemes, and coverage, complicating their integration into a shared judgment format while retaining source\-specific criteria\. Third,*uneven supervision*: differences in dataset sizes and label frequencies can favor well\-represented tasks and frequent classes during joint training\. Addressing these challenges calls for a shared question\-conditioned judgment framework and a training strategy that accounts for heterogeneous, imbalanced supervision\.

To address these challenges, we introduce PsyJudgeBench, a benchmark for question\-conditioned psychological analysis\. It organizes 13 existing datasets into 82 registered questions across six psychological content areas\. Rather than merging heterogeneous tasks into a single label space, PsyJudgeBench represents each judgment through an analytical question, its legal answer set, and source\-specific decision criteria retained through documented label mappings\. Missing\-label masking and coordinated dataset splits support joint training and task\-specific evaluation under partial supervision\. The resulting benchmark also enables evaluating how models distinguish different judgments about the same text\.

Building on this task formulation, we introduce PsyJev, a System\-1 model that directly predicts distributions over question\-defined answers\. Given an input text, an analytical question, and candidate options, PsyJev produces a categorical answer distribution through a shared readout, without intermediate response generation or post\-hoc category extraction\([Almeida, 2026](https://arxiv.org/html/2610.36689#bib.bib28);[Gao and Wang, 2026](https://arxiv.org/html/2610.36689#bib.bib29)\)\. To account for uneven supervision across sources, we further propose Judgment\-Anchored RLCD \(JA\-RLCD\), adapting logit\-space RLCD\([Laya Contributors, 2026](https://arxiv.org/html/2610.36689#bib.bib12)\)to heterogeneous psychological judgments\. Its judgment anchor combines source/question weighting and class\-balanced supervision with selective retention of correct SFT predictions\. A Brier\-based question\-group reward further refines predictions through logit\-space perturbations and pathwise gradients\. Across 2B, 4B, and 9B model scales, PsyJev outperforms the corresponding same\-backbone answer\-SFT baselines\. Controlled 2B comparisons show that the complete JA\-RLCD formulation outperforms continued cross\-entropy training and anchor\-only refinement\. Further analyses examine question\-conditioned discrimination on co\-annotated texts, inference latency, and probability quality\.

Our main contributions are as follows:

- •A unified psychological judgment benchmark\.We construct PsyJudgeBench, organizing 13 heterogeneous datasets into a shared question\-defined judgment framework that preserves task\-specific criteria and supports unified training and evaluation across questions\.
- •System\-1 modeling with judgment\-anchored reward refinement\.We propose PsyJev for direct probabilistic psychological judgment andJA\-RLCD, a joint refinement formulation that adapts logit\-space RLCD through a class\-balanced judgment anchor, selective SFT\-reference retention, and question\-group reward optimization under uneven supervision\.
- •Systematic empirical evaluation\.We evaluate PsyJev across model scales and heterogeneous psychological tasks, with comparisons against benchmark\-adapted baselines and matched training continuations\. Ablations and further analyses assess the training components, question\-conditioned discrimination, inference efficiency, and probability quality\.

## 2\.Related Work

### 2\.1\.Psychological Analysis

Psychological analysis of social Web content spans diverse tasks, including identifying depression and anxiety indications, interpersonal difficulties, self\-perceptions, and emotional experiences\([Hengle et al\., 2024](https://arxiv.org/html/2610.36689#bib.bib1);[Garg et al\., 2023c](https://arxiv.org/html/2610.36689#bib.bib5);[Garg et al\., 2023a](https://arxiv.org/html/2610.36689#bib.bib6);[Garg et al\., 2023b](https://arxiv.org/html/2610.36689#bib.bib24);[Rahman et al\., 2024](https://arxiv.org/html/2610.36689#bib.bib4)\)\. Other resources address questionnaire answerability, symptom relevance, stressors, cognitive distortions, wellness dimensions, and crisis\-related expressions\([Gupta et al\., 2022](https://arxiv.org/html/2610.36689#bib.bib7);[Pérez et al\., 2023](https://arxiv.org/html/2610.36689#bib.bib2);[Mauriello et al\., 2021](https://arxiv.org/html/2610.36689#bib.bib8);[Garg et al\., 2022](https://arxiv.org/html/2610.36689#bib.bib23);[Sharma et al\., 2023](https://arxiv.org/html/2610.36689#bib.bib26);[Shreevastava and Foltz, 2021](https://arxiv.org/html/2610.36689#bib.bib30);[Garg et al\., 2024](https://arxiv.org/html/2610.36689#bib.bib25);[Byun et al\., 2026](https://arxiv.org/html/2610.36689#bib.bib27)\)\. These tasks involve related psychological content but differ in their analytical targets and annotation criteria\. Existing models range from domain\-adapted encoders such as MentalBERT to instruction\-tuned language models for psychological prediction and explanation\([Ji et al\., 2022](https://arxiv.org/html/2610.36689#bib.bib15);[Xu et al\., 2024](https://arxiv.org/html/2610.36689#bib.bib16);[Zhai et al\., 2025](https://arxiv.org/html/2610.36689#bib.bib17)\)\. Flexible classification approaches, including GLiClass and GLiNER2, further support label\- or schema\-conditioned predictions\([Stepanov et al\., 2025](https://arxiv.org/html/2610.36689#bib.bib10);[Zaratiana et al\., 2025](https://arxiv.org/html/2610.36689#bib.bib18)\)\. MentaLLaMA and its IMHI benchmark extend psychological analysis toward multi\-task instruction learning over social\-media content\([Yang et al\., 2024](https://arxiv.org/html/2610.36689#bib.bib9)\)\. Building on this direction, PsyJudgeBench organizes heterogeneous psychological tasks as question\-defined judgments, while PsyJev learns a shared direct\-prediction function and refines it throughJA\-RLCD\.

### 2\.2\.System\-1 Decision Models

System\-1 decision models produce structured predictions without autoregressive answer generation\. Jev introduced natural\-language interfaces for probabilistic decisions\([Almeida, 2026](https://arxiv.org/html/2610.36689#bib.bib28)\), while Laya developed typed prediction with reinforcement learning for calibrated decisions \(RLCD\)\([Laya Contributors, 2026](https://arxiv.org/html/2610.36689#bib.bib12)\)\. OpenJev further introduced dynamic\-option modeling with candidate\-conditioned representations and option\-set interaction\([OpenJev Contributors, 2026](https://arxiv.org/html/2610.36689#bib.bib32)\)\. Related advances include rationale\-conditioned reward learning and direct probabilistic video understanding\([Gao and Wang, 2026](https://arxiv.org/html/2610.36689#bib.bib29);[Zhu et al\., 2026](https://arxiv.org/html/2610.36689#bib.bib31)\)\. Building on this line of work, PsyJev supports heterogeneous psychological judgments through joint supervision on PsyJudgeBench\. OurJA\-RLCDadapts logit\-space RLCD with a class\-balanced judgment anchor, selective SFT reference retention, and question\-group reward optimization using pathwise gradients\.

## 3\.PsyJudgeBench

PsyJudgeBench organizes 13 annotated datasets into 53,775 source records and 222,191 observed judgments\. Figure[1](https://arxiv.org/html/2610.36689#acmlabel1)summarizes their coverage across six thematic groups: depression/anxiety\-related information, self and relationships, stress and distress, cognition, wellness, and crisis experiences\. Construction involves a question registry, source\-specific annotation conversion, and coordinated data partitions\.

##### Question\-defined tasks\.

PsyJudgeBench represents psychological analysis tasks as question\-defined judgments, each specifying a judgment target, a decision criterion, and a set of valid answers\. This formulation accommodates different analytical requirements while preserving their original annotation criteria\. For example, PRIMATE assesses whether a text provides enough information to answer a PHQ\-9 item, BDI\-Sen evaluates sentence relevance to a BDI\-II symptom\([Gupta et al\., 2022](https://arxiv.org/html/2610.36689#bib.bib7);[Pérez et al\., 2023](https://arxiv.org/html/2610.36689#bib.bib2)\), and MultiWD identifies mentions of specific wellness dimensions\. Each judgment is represented as\(x,q,𝒜q,y\)\(x,q,\\mathcal\{A\}\_\{q\},y\), comprising the source text, question, valid answer set, and observed answer\. The resulting registry contains 82 questions, including 79 binary and three multiclass questions\.

![A concentric benchmark overview groups thirteen datasets into six psychological content areas and displays each dataset's question count and number of source records.](https://arxiv.org/html/2610.36689v1/dataset.png)Figure 1\.PsyJudgeBench coverage\. The inner groups organize 13 source datasets into six psychological content areas\. Q denotes the number of questions associated with each source, and the outer bars and numbers indicate source\-record counts\.A concentric benchmark overview groups thirteen datasets into six psychological content areas and displays each dataset's question count and number of source records\.
##### Annotation harmonization\.

Following these question definitions, we convert heterogeneous source annotations into observed judgments while retaining their original labeling criteria\. Binary and categorical labels are mapped to their corresponding answer sets, while multi\-label and category–time annotations are expanded into separate questions\. For Thinking Traps, we normalize category names and aggregate repeated completions to recover 15\-category label sets\. For CRADLE, seven crisis categories combined with ongoing/past status yield 14 questions\. For TherapistQA, we retain the dominant distortion category and use only the patient’s question, excluding therapist responses and annotated spans\. After conversion, we check for conflicting annotations associated with the same text and judgment target\. We remove two TherapistQA records with identical questions but conflicting dominant labels and mask 22 conflicting judgments across 11 same\-text/same\-target keys\. Unaffected judgments remain available for supervision, while source identifiers and annotation provenance, including gold/silver distinctions, are preserved\.

##### Partial supervision and coordinated splits\.

The harmonized records provide supervision only for questions with observed annotations; unannotated questions are excluded rather than treated as negative examples\. To construct consistent evaluation partitions, we coordinate records sharing identical texts or available source\-group identifiers and group identified near\-duplicates to reduce overlap across splits\. When grouping conflicts with the original partitions, held\-out assignments take precedence\. The resulting benchmark contains 33,511 training, 5,852 validation, and 14,412 test source records\. These counts refer to source records rather than unique texts\. Further details on data provenance, partitioning, and duplicate checks are provided in Appendix[A](https://arxiv.org/html/2610.36689#A1)\.

![The prediction pathway serializes the text, question, criterion, and labeled candidates, encodes them with a Qwen3.5 backbone, and normalizes answer-position vocabulary logits over the legal answer identifiers. The training pathway applies supervised fine-tuning followed by JA-RLCD, which explores answer logits, evaluates Brier-based question-group rewards, and updates the judgment policy with pathwise reward gradients and a judgment anchor.](https://arxiv.org/html/2610.36689v1/model_arc_v2_compress.png)Figure 2\.PsyJev overview\. \(a\) System\-1\-style direct judgment through question\-conditioned input serialization and constrained legal\-answer scoring\. \(b\) Two\-stage training with joint SFT followed by JA\-RLCD, combining judgment anchoring with logit\-space exploration and question\-group reward optimization\.The prediction pathway serializes the text, question, criterion, and labeled candidates, encodes them with a Qwen3\.5 backbone, and normalizes answer\-position vocabulary logits over the legal answer identifiers\. The training pathway applies supervised fine\-tuning followed by JA\-RLCD, which explores answer logits, evaluates Brier\-based question\-group rewards, and updates the judgment policy with pathwise reward gradients and a judgment anchor\.

## 4\.PsyJev

PsyJev is a System\-1\-style model for unified psychological judgment\. Given a text, an analytical question, and its legal answers, PsyJev directly produces a categorical distribution through a shared, question\-conditioned readout, without autoregressive answer generation\. Its training proceeds in two stages: joint supervised fine\-tuning \(SFT\) learns the shared decision function across heterogeneous tasks, followed by Judgment\-Anchored RLCD \(JA\-RLCD\), which combines balanced supervision, selective SFT\-reference retention, and question\-group reward optimization through logit\-space exploration\. Figure[2](https://arxiv.org/html/2610.36689#acmlabel2)illustrates the overall framework\.

### 4\.1\.System\-1\-Style Direct Judgment

PsyJev implements direct psychological judgment through a question\-conditioned, constrained prediction interface\. Rather than generating and parsing free\-form responses, it evaluates the admissible answers using their corresponding language\-model logits and directly produces a normalized probability distribution\.

##### Question\-conditioned representation\.

Given a textxx, an analytical questionqq, and its legal answer set𝒜q=\{a1,…,aKq\}\\mathcal\{A\}\_\{q\}=\\\{a\_\{1\},\\ldots,a\_\{K\_\{q\}\}\\\}, PsyJev constructs a structured input that combines the source text, question, decision criterion, and candidate descriptions\. Each candidate is assigned a distinct answer identifier, allowing different psychological judgments to be expressed through the same input format\. A shared Qwen3\.5 backbone processes the complete judgment request and produces a contextual representation at the answer position\.

##### Constrained answer scoring\.

Instead of generating an answer sequence, PsyJev directly extracts the vocabulary logits corresponding to the admissible answer identifiers\. Lettkt\_\{k\}denote the token assigned to candidateaka\_\{k\}, and letzθ​\(tk∣x,q,𝒜q\)z\_\{\\theta\}\(t\_\{k\}\\mid x,q,\\mathcal\{A\}\_\{q\}\)denote its logit at the answer position\. The resulting judgment distribution is

\(1\)ℓq​k\\displaystyle\\ell\_\{qk\}=zθ​\(tk∣x,q,𝒜q\),\\displaystyle=z\_\{\\theta\}\(t\_\{k\}\\mid x,q,\\mathcal\{A\}\_\{q\}\),pθ​\(k∣x,q,𝒜q\)\\displaystyle p\_\{\\theta\}\(k\\mid x,q,\\mathcal\{A\}\_\{q\}\)=exp⁡\(ℓq​k\)∑j=1Kqexp⁡\(ℓq​j\)\.\\displaystyle=\\frac\{\\exp\(\\ell\_\{qk\}\)\}\{\\sum\_\{j=1\}^\{K\_\{q\}\}\\exp\(\\ell\_\{qj\}\)\}\.
Normalization is restricted to the legal answer set, yielding a categorical distribution whose dimensionality is determined by the current question\.

##### Direct probabilistic judgment\.

The predicted answer is obtained directly from the normalized answer distribution:

\(2\)y^=arg⁡maxk∈\{1,…,Kq\}​pθ​\(k∣x,q,𝒜q\)\.\\hat\{y\}=\\arg\\max\_\{k\\in\\\{1,\\ldots,K\_\{q\}\\\}\}p\_\{\\theta\}\(k\\mid x,q,\\mathcal\{A\}\_\{q\}\)\.
The same scoring procedure supports binary and multiclass psychological judgments without task\-specific classification heads or autoregressive answer generation\. By retaining the complete distribution over admissible answers, PsyJev also provides the probabilistic decision interface used by subsequent JA\-RLCD refinement\. Appendix[B](https://arxiv.org/html/2610.36689#A2)provides the implementation details\.

### 4\.2\.Supervised Fine\-Tuning

Joint SFT establishes the shared judgment function by learning from observed annotations across source datasets\. Let𝒟s\\mathcal\{D\}\_\{s\}denote the records from sourcess, and let𝒪⁡\(x\)\\mathcal\{O\}\(x\)denote the questions annotated for recordxx\. We minimize categorical cross\-entropy:

\(3\)ℒSFT=−1S∑s=1S𝔼x∼𝒟s\[1\|𝒪⁡\(x\)\|∑q∈𝒪⁡\(x\)logpθ\(yq∣x,q,𝒜q\)\]\.\\mathcal\{L\}\_\{\\rm SFT\}=\-\\frac\{1\}\{S\}\\sum\_\{s=1\}^\{S\}\\mathbb\{E\}\_\{x\\sim\\mathcal\{D\}\_\{s\}\}\\left\[\\frac\{1\}\{\|\\mathcal\{O\}\(x\)\|\}\\sum\_\{q\\in\\mathcal\{O\}\(x\)\}\\log p\_\{\\theta\}\(y\_\{q\}\\mid x,q,\\mathcal\{A\}\_\{q\}\)\\right\]\.
The outer average assigns equal weight to each source, while the inner average prevents records with more annotated questions from dominating the objective\. Missing question labels contribute no loss\. The resulting checkpointθSFT\\theta\_\{\\rm SFT\}initializes the subsequent JA\-RLCD refinement\.

### 4\.3\.Judgment\-Anchored RLCD

RLCD explores probabilistic decisions through logit\-space perturbations and reward evaluation\([Laya Contributors, 2026](https://arxiv.org/html/2610.36689#bib.bib12)\)\. Building on this formulation,JA\-RLCDaddresses uneven source, question, and answer frequencies in psychological judgments through a joint refinement objective: a judgment anchor combines class\-balanced supervision and selective retention of correct SFT predictions, while a Brier\-based reward emphasizes higher\-error question groups within each microbatch\. Starting from the SFT checkpoint, we treat the legal\-answer distribution as a judgment policy and refine it through paired Gaussian perturbations and differentiable reward optimization\. For each source\-specific microbatch of observed judgments, reward gradients are computed through pathwise differentiation and combined with the anchor gradients before accumulation and AdamW updates\. Algorithm[1](https://arxiv.org/html/2610.36689#alg1)summarizes the procedure\.

Algorithm 1Judgment\-Anchored RLCD \(JA\-RLCD\)1:Pretrained LLM

θpre\\theta\_\{\\rm pre\}; annotated records

𝒟\\mathcal\{D\}
2:Trained direct\-judgment model

θ\\theta
3:

θSFT←SFT⁡\(θpre,𝒟\)\\theta\_\{\\rm SFT\}\\leftarrow\\operatorname\{SFT\}\(\\theta\_\{\\rm pre\},\\mathcal\{D\}\)using Eq\.[3](https://arxiv.org/html/2610.36689#S4.E3)

4:Initialize

θ←θSFT\\theta\\leftarrow\\theta\_\{\\rm SFT\}; freeze reference

θ0←θSFT\\theta\_\{0\}\\leftarrow\\theta\_\{\\rm SFT\}
5:Compute class costs

cic\_\{i\}and source/question training counts

6:foreach record batch

𝒞⊂𝒟\\mathcal\{C\}\\subset\\mathcal\{D\}do

7:Set

aia\_\{i\}using batch size

B=\|𝒞\|B=\|\\mathcal\{C\}\|⊳\\trianglerightAppendix[B](https://arxiv.org/html/2610.36689#A2)

8:

dacc←0d\_\{\\rm acc\}\\leftarrow 0
9:foreach source\-specific pack

ℬ\\mathcal\{B\}of at most 16 judgmentsdo

10:

ℓ←Readout⁡\(LLMθ,ℬ\)\\ell\\leftarrow\\operatorname\{Readout\}\(\\operatorname\{LLM\}\_\{\\theta\},\\mathcal\{B\}\)⊳\\trianglerightEq\.[1](https://arxiv.org/html/2610.36689#S4.E1)

11:

p0←ReadoutProb⁡\(LLMθ0,ℬ\)p\_\{0\}\\leftarrow\\operatorname\{ReadoutProb\}\(\\operatorname\{LLM\}\_\{\\theta\_\{0\}\},\\mathcal\{B\}\)⊳\\trianglerightFrozen

12:Set

mi←𝕀\[argmaxkp0,i,k=yi\]m\_\{i\}\\leftarrow\\mathbb\{I\}\[\\arg\\max\_\{k\}p\_\{0,i,k\}=y\_\{i\}\]for

i∈ℬi\\in\\mathcal\{B\}
13:Form balanced CE and gated KL anchor

ℒA\\mathcal\{L\}\_\{\\rm A\}⊳\\trianglerightEq\.[4](https://arxiv.org/html/2610.36689#S4.E4)

14:Set

Aℬ←∑i∈ℬaiA\_\{\\mathcal\{B\}\}\\leftarrow\\sum\_\{i\\in\\mathcal\{B\}\}a\_\{i\}and

𝒢←\{qi:i∈ℬ\}\\mathcal\{G\}\\leftarrow\\\{q\_\{i\}:i\\in\\mathcal\{B\}\\\}
15:Explore with two Gaussian directions per judgment; form

z1±,z2±z\_\{1\}^\{\\pm\},z\_\{2\}^\{\\pm\}
16:Compute legal\-answer probabilities

𝒫1±,𝒫2±\\mathcal\{P\}\_\{1\}^\{\\pm\},\\mathcal\{P\}\_\{2\}^\{\\pm\}⊳\\trianglerightEq\.[6](https://arxiv.org/html/2610.36689#S4.E6)

17:Evaluate

ℛτ​\(𝒫mv\)\\mathcal\{R\}\_\{\\tau\}\(\\mathcal\{P\}\_\{m\}^\{v\}\)on

ℬ\\mathcal\{B\}for all four draws

18:

𝒥^R←Aℬ4​∑m=12∑v∈\{\+,−\}ℛτ​\(𝒫mv\)\\widehat\{\\mathcal\{J\}\}\_\{\\rm R\}\\leftarrow\\frac\{A\_\{\\mathcal\{B\}\}\}\{4\}\\sum\_\{m=1\}^\{2\}\\sum\_\{v\\in\\\{\+,\-\\\}\}\\mathcal\{R\}\_\{\\tau\}\(\\mathcal\{P\}\_\{m\}^\{v\}\)
19:

g←∇θℒAg\\leftarrow\\nabla\_\{\\theta\}\\mathcal\{L\}\_\{\\rm A\};

u←∇θ𝒥^Ru\\leftarrow\\nabla\_\{\\theta\}\\widehat\{\\mathcal\{J\}\}\_\{\\rm R\}⊳\\trianglerightPathwise

20:

dℬ←g−ud\_\{\\mathcal\{B\}\}\\leftarrow g\-u⊳\\trianglerightAnchored refinement; Eq\.[7](https://arxiv.org/html/2610.36689#S4.E7)

21:

dacc←dacc\+dℬd\_\{\\rm acc\}\\leftarrow d\_\{\\rm acc\}\+d\_\{\\mathcal\{B\}\}
22:endfor

23:

θ←AdamW⁡\(θ,ClipNorm⁡\(dacc,1\)\)\\theta\\leftarrow\\operatorname\{AdamW\}\(\\theta,\\operatorname\{ClipNorm\}\(d\_\{\\rm acc\},1\)\)
24:endfor

25:return

θ\\theta⊳\\trianglerightUse direct readout for inference

##### Judgment anchor\.

For judgmenti∈ℬi\\in\\mathcal\{B\}, training\-count weightaia\_\{i\}balances source/question observation frequency, and costci=cqi,yic\_\{i\}=c\_\{q\_\{i\},y\_\{i\}\}balances answer classes \(Appendix[B](https://arxiv.org/html/2610.36689#A2)\)\. Letpip\_\{i\}andp0,ip\_\{0,i\}denote the trainable and frozen distributions over𝒜qi\\mathcal\{A\}\_\{q\_\{i\}\}\. The anchor is:

\(4\)ℒA=∑i∈ℬaici\[−logpi\(yi\)\]\+λ∑i∈ℬaimiKL\(p0,i∥pi\),\\mathcal\{L\}\_\{\\rm A\}=\\sum\_\{i\\in\\mathcal\{B\}\}a\_\{i\}c\_\{i\}\[\-\\log p\_\{i\}\(y\_\{i\}\)\]\+\\lambda\\sum\_\{i\\in\\mathcal\{B\}\}a\_\{i\}m\_\{i\}\\operatorname\{KL\}\(p\_\{0,i\}\\\|p\_\{i\}\),wheremi=𝕀\[argmaxkp0,i,k=yi\]m\_\{i\}=\\mathbb\{I\}\[\\arg\\max\_\{k\}p\_\{0,i,k\}=y\_\{i\}\]\. The indicator𝕀⁡\[⋅\]\\mathbb\{I\}\[\\cdot\]equals 1 when its condition holds and 0 otherwise, so reference retention applies only to judgments the frozen model answers correctly\. The label term anchors reward\-driven changes to observed answers, while the reference term encourages retention of the SFT distribution where its prediction agrees with the annotation\.

##### Brier\-based question\-group reward\.

Let𝒢=\{qi:i∈ℬ\}\\mathcal\{G\}=\\\{q\_\{i\}:i\\in\\mathcal\{B\}\\\}be the observed questions in a microbatch, and let𝒫=\(pi\)i∈ℬ\\mathcal\{P\}=\(p\_\{i\}\)\_\{i\\in\\mathcal\{B\}\}collect its legal\-answer distributions\. We define each question’s reward as the negative class\-weighted Brier error\([Gneiting and Raftery, 2007](https://arxiv.org/html/2610.36689#bib.bib14)\), then aggregate the question rewards with a mean and an entropic soft minimum:

\(5\)Rq​\(𝒫\)\\displaystyle R\_\{q\}\(\\mathcal\{P\}\)=−∑i∈ℬ:qi=qci12∥pi−eyi∥22∑i∈ℬ:qi=qci,\\displaystyle=\-\\frac\{\\sum\_\{i\\in\\mathcal\{B\}:q\_\{i\}=q\}c\_\{i\}\\,\\tfrac\{1\}\{2\}\\\|p\_\{i\}\-e\_\{y\_\{i\}\}\\\|\_\{2\}^\{2\}\}\{\\sum\_\{i\\in\\mathcal\{B\}:q\_\{i\}=q\}c\_\{i\}\},ℛτ​\(𝒫\)\\displaystyle\\mathcal\{R\}\_\{\\tau\}\(\\mathcal\{P\}\)=12​R¯−12​τ​log⁡\(1\|𝒢\|​∑q∈𝒢e−τ​Rq​\(𝒫\)\)\.\\displaystyle=\\tfrac\{1\}\{2\}\\overline\{R\}\-\\frac\{1\}\{2\\tau\}\\log\\left\(\\frac\{1\}\{\|\\mathcal\{G\}\|\}\\sum\_\{q\\in\\mathcal\{G\}\}e^\{\-\\tau R\_\{q\}\(\\mathcal\{P\}\)\}\\right\)\.
Hereeyie\_\{y\_\{i\}\}is the one\-hot answer andR¯=\|𝒢\|−1​∑q∈𝒢Rq​\(𝒫\)\\overline\{R\}=\|\\mathcal\{G\}\|^\{\-1\}\\sum\_\{q\\in\\mathcal\{G\}\}R\_\{q\}\(\\mathcal\{P\}\)\. Higher reward indicates lower prediction error\. Withτ=4\\tau=4, the soft\-minimum term places more weight on questions with lower reward within the current microbatch\. This aggregate is the negative of the corresponding question\-group risk\.

##### Logit\-space exploration\.

We drawM=2M=2independent Gaussian directionsϵm,i∼𝒩⁡\(0,IKi\)\\epsilon\_\{m,i\}\\sim\\mathcal\{N\}\(0,I\_\{K\_\{i\}\}\)per judgment, whereKi=\|𝒜qi\|K\_\{i\}=\|\\mathcal\{A\}\_\{q\_\{i\}\}\|\. The projectionPi=I−𝟏𝟏⊤/KiP\_\{i\}=I\-\\mathbf\{1\}\\mathbf\{1\}^\{\\top\}/K\_\{i\}removes the softmax\-invariant common shift;𝟏∈ℝKi\\mathbf\{1\}\\in\\mathbb\{R\}^\{K\_\{i\}\}denotes the all\-ones vector\. WithAℬ=∑i∈ℬaiA\_\{\\mathcal\{B\}\}=\\sum\_\{i\\in\\mathcal\{B\}\}a\_\{i\}, paired perturbations give the sampled, Gaussian\-smoothed reward:

\(6\)zm,i±\\displaystyle z\_\{m,i\}^\{\\pm\}=ℓθ,i±σ​Pi​ϵm,i,\\displaystyle=\\ell\_\{\\theta,i\}\\pm\\sigma P\_\{i\}\\epsilon\_\{m,i\},pm,iv\\displaystyle p\_\{m,i\}^\{v\}=softmax\(zm,iv\),v∈\{\+,−\},\\displaystyle=\\operatorname\{softmax\}\(z\_\{m,i\}^\{v\}\),\\qquad v\\in\\\{\+,\-\\\},𝒥^R\\displaystyle\\widehat\{\\mathcal\{J\}\}\_\{\\rm R\}=Aℬ2​M​∑m=1M∑v∈\{\+,−\}ℛτ​\(𝒫mv\)\.\\displaystyle=\\frac\{A\_\{\\mathcal\{B\}\}\}\{2M\}\\sum\_\{m=1\}^\{M\}\\sum\_\{v\\in\\\{\+,\-\\\}\}\\mathcal\{R\}\_\{\\tau\}\(\\mathcal\{P\}\_\{m\}^\{v\}\)\.
Here𝒫mv=\(pm,iv\)i∈ℬ\\mathcal\{P\}\_\{m\}^\{v\}=\(p\_\{m,i\}^\{v\}\)\_\{i\\in\\mathcal\{B\}\}, and softmax is over each judgment’s legal answers\. All four reward evaluations recompute Eq\.[5](https://arxiv.org/html/2610.36689#S4.E5)on the same microbatch from the same model logits\. The individual anchor weightsaia\_\{i\}balance judgments; their sumAℬA\_\{\\mathcal\{B\}\}scales the entire reward estimate rather than reweighting its question groups\.

##### Pathwise reward gradient\.

Since the question\-group reward is differentiable with respect to the answer distributions, we directly propagate its gradients through the perturbed logits and legal\-answer softmax\. With Gaussian noise sampled independently ofθ\\theta, this yields the pathwise reward gradientu=∇θ𝒥^Ru=\\nabla\_\{\\theta\}\\widehat\{\\mathcal\{J\}\}\_\{\\rm R\}, allowing the sampled rewards to guide refinement without relying on a score\-function estimator\. We further examine the choice of gradient estimator in Section[5\.3\.2](https://arxiv.org/html/2610.36689#S5.SS3.SSS2)\.

##### Anchored policy refinement\.

We maximize reward while maintaining the judgment anchor\. Writingg=∇θℒAg=\\nabla\_\{\\theta\}\\mathcal\{L\}\_\{\\rm A\}, the anchored objective and its descent direction are:

\(7\)𝒥^AP\\displaystyle\\widehat\{\\mathcal\{J\}\}\_\{\\rm AP\}=𝒥^R−ℒA,\\displaystyle=\\widehat\{\\mathcal\{J\}\}\_\{\\rm R\}\-\\mathcal\{L\}\_\{\\rm A\},d\\displaystyle d=−∇θ𝒥^AP=g−u\.\\displaystyle=\-\\nabla\_\{\\theta\}\\widehat\{\\mathcal\{J\}\}\_\{\\rm AP\}=g\-u\.
We accumulateddacross microbatches, clip the accumulated direction to a global norm of 1, and apply AdamW, keeping the SFT reference frozen\. Equivalently, the anchored objective minimizes a supervised loss combining the judgment anchor with Gaussian\-smoothed, question\-group prediction risk\.

## 5\.Experiments

### 5\.1\.Experimental Setup

#### 5\.1\.1\.Benchmark and Metrics

We evaluate PsyJev on PsyJudgeBench using the coordinated splits described in Section[3](https://arxiv.org/html/2610.36689#S3)\. We report positive\-class F1 for binary questions and macro\-F1 for multiclass questions, averaging across observed questions within each dataset\. The overall score equally weights the 13 datasets\. We assess probability quality using normalized negative log\-likelihood \(NLL\), Brier score, and expected calibration error \(ECE\) with 15 equal\-width bins\. NLL is normalized bylog⁡\|𝒜q\|\\log\|\\mathcal\{A\}\_\{q\}\|to account for varying answer\-set sizes\. Efficiency is measured by model computation time per judgment\.

#### 5\.1\.2\.Baselines

We compare with 18 benchmark\-adapted baselines and the zero\-shot Jev API\. The adapted baselines fall into three groups\. Compact classification and decision models include MentalBERT, MentalRoBERTa, ModernBERT\-large, GLiClass, GLiNER2, and Laya\-multilingual\([Ji et al\., 2022](https://arxiv.org/html/2610.36689#bib.bib15);[Warner et al\., 2025](https://arxiv.org/html/2610.36689#bib.bib11);[Stepanov et al\., 2025](https://arxiv.org/html/2610.36689#bib.bib10);[Zaratiana et al\., 2025](https://arxiv.org/html/2610.36689#bib.bib18);[Laya Contributors, 2026](https://arxiv.org/html/2610.36689#bib.bib12)\)\. Laya uses its released architecture and checkpoint; its reported result follows benchmark SFT and continuation with its native RLCD objective\. The 2–4B language models are FLAN\-T5\-XL, Qwen3\.5\-2B/4B, Llama\-3\.2\-3B, Phi\-4\-mini, and Gemma\-3\-4B\([Chung and others, 2024](https://arxiv.org/html/2610.36689#bib.bib19);[Qwen Team, 2026](https://arxiv.org/html/2610.36689#bib.bib22);[Meta, 2024](https://arxiv.org/html/2610.36689#bib.bib36);[Microsoft and others, 2025](https://arxiv.org/html/2610.36689#bib.bib35);[Gemma Team and others, 2025](https://arxiv.org/html/2610.36689#bib.bib34)\)\. The 7–9B models include MentaLLaMA\-7B, Qwen2\.5\-7B, Llama\-3\.1\-8B, Qwen3\-8B, Gemma\-2\-9B, and Qwen3\.5\-9B\([Yang et al\., 2024](https://arxiv.org/html/2610.36689#bib.bib9);[Qwen Team, 2024](https://arxiv.org/html/2610.36689#bib.bib13);[Dubey and others, 2024](https://arxiv.org/html/2610.36689#bib.bib20);[Yang et al\., 2025](https://arxiv.org/html/2610.36689#bib.bib21);[Gemma Team and others, 2024](https://arxiv.org/html/2610.36689#bib.bib33);[Qwen Team, 2026](https://arxiv.org/html/2610.36689#bib.bib22)\)\. All 18 models are adapted on the benchmark training partition\. Classification and decision models predict the observed task labels directly\. Language models learn to produce the corresponding answers and are evaluated by scoring the allowed answer sequences\. We additionally evaluate Typesafe’s Jev \(jev\-1\.14\.0\) through its official API without benchmark fine\-tuning or demonstrations\([Almeida, 2026](https://arxiv.org/html/2610.36689#bib.bib28)\)\. Appendix[B\.3](https://arxiv.org/html/2610.36689#A2.SS3)specifies both adaptation and API protocols\.

#### 5\.1\.3\.Implementation Details

We initialize PsyJev\-2B/4B/9B from the corresponding Qwen3\.5 text backbones and train each model with joint SFT followed by JA\-RLCD, using one epoch per stage\. Source texts are truncated to 512 tokens, with batches of up to 64 texts\. The 2B model updates all parameters at a learning rate of2×10−62\\times 10^\{\-6\}in both stages\. The 4B/9B models use LoRA \(rank 16, scaling 32, zero dropout\), with learning rates of2×10−42\\times 10^\{\-4\}for SFT and2×10−62\\times 10^\{\-6\}for JA\-RLCD\. All models use AdamW with weight decay 0\.01, gradient clipping at 1\.0, and 3% warmup followed by cosine decay\. Backbone computation uses BF16, and the final checkpoint of each stage is retained\. We measure model computation time using a single H100 80GB GPU, one text per request, and a fixed 13\-text training probe shared by all timed models; classification metrics use the complete test partition\. Further details appear in Appendices[B](https://arxiv.org/html/2610.36689#A2)and[B\.4](https://arxiv.org/html/2610.36689#A2.SS4)\.

Table 1\.Performance comparison on PsyJudgeBench \(F1, %\)\. Avg\. denotes the unweighted mean across 13 datasets\. Bold indicates the best result within each benchmark\-adapted model group\. Jev is evaluated zero\-shot\. Latency measures model computation \(ms/judgment\) on the fixed H100 CUDA Graph probe\. Abbreviations: DepEmo \(DepressionEmo\), BDISen \(BDI\-Sen\), LonXpl \(LonXplain\), ThinkTrap \(Thinking Traps\), and TherapQA \(TherapistQA\)\.

### 5\.2\.Performance Evaluation

Table[1](https://arxiv.org/html/2610.36689#S5.T1)shows that PsyJev consistently achieves the highest average F1 within each model\-size group, reaching 64\.33, 66\.21, and 67\.56 at 2B, 4B, and 9B, respectively\. Compared with the corresponding Qwen3\.5 answer models, these results represent gains of 4\.01, 1\.71, and 3\.88 points\. Notably, the 4B and 9B variants also outperform the strongest external baseline, despite the smaller size of the former\. Across datasets, PsyJev shows notable gains on tasks involving cognitive distortions and wellness\-related judgments, although gains vary across question criteria and are not uniform across all datasets\. The general\-purpose Jev API achieves 52\.91 average F1 in zero\-shot evaluation, while the benchmark\-adapted PsyJev variants perform substantially better, providing a zero\-shot reference alongside the benchmark\-adapted comparisons\. Overall, these results demonstrate the effectiveness of PsyJev across heterogeneous psychological judgment tasks and model scales\.

### 5\.3\.Ablation Study

Table[2](https://arxiv.org/html/2610.36689#S5.T2)uses the 2B model to examine training stages, JA\-RLCD components, and matched\-budget prediction formats\. Direct\-scoring continuations share the same SFT initialization and schedule; the answer\-sequence control uses its corresponding SFT model with the same update budget\.

#### 5\.3\.1\.Training Stages

JA\-RLCD achieves 64\.33 average F1, improving upon SFT by 3\.65 points\. It also outperforms continued cross\-entropy \(CE\) training and adapted RLCD by 2\.61 and 5\.46 points, respectively\. Continued CE controls for additional training, while adapted RLCD combines CE supervision with the original RLCD’s noisy\-logit score\-function objective\. These comparisons show that additional training alone does not explain the improvement and demonstrate the effectiveness of the overall JA\-RLCD formulation\. Since adapted RLCD differs in both its reward formulation and gradient estimator, the comparison does not isolate the contribution of either component\.

Table 2\.2B ablations and matched\-budget controls\.Δ\\DeltaF1 is relative to full JA\-RLCD, computed before rounding\.
#### 5\.3\.2\.Component Ablations

Class\-balanced CE alone achieves 63\.29 F1, compared with 63\.70 for anchor\-only training, which also retains selective reference KL\. Full JA\-RLCD reaches 64\.33, improving over balanced CE by 1\.04 points and anchor\-only training by 0\.63 points\.

Replacing the question\-group/tail reward with a per\-judgment Brier reward gives 63\.36 F1 while retaining the anchor, perturbations, and pathwise estimator\. The 0\.97\-point advantage supports the combined group/tail reward formulation\. Removing class costs reduces F1 to 62\.15; removing reference KL or replacing pathwise differentiation with a paired score\-function estimator gives 64\.12 and 63\.91, respectively\. Appendix[B\.3](https://arxiv.org/html/2610.36689#A2.SS3.SSS0.Px3)details these controls\.

#### 5\.3\.3\.Performance Across Questions

We further examine whether JA\-RLCD improves questions that are difficult for the SFT model\. Using only SFT validation performance, we define the subset as the 16 lowest\-F1 questions among 62 questions with sufficient validation annotations\. As shown in Table[3](https://arxiv.org/html/2610.36689#S5.T3), JA\-RLCD improves mean test F1 by 19\.45 points over SFT on these difficult questions, compared with 6\.23 points across all 82 questions\. The gains remain positive relative to continued CE and anchor\-only training, reaching 13\.38 and 3\.02 points on the difficult subset, respectively\. JA\-RLCD improves test F1 on 57\.32% of all questions relative to SFT\. These findings show larger average gains on the validation\-defined difficult subset, while improvements are not uniform across individual questions\. Unlike the dataset\-balanced benchmark average, these analyses weight questions equally\.

Table 3\.Question\-averaged test\-F1 gains of JA\-RLCD across all 82 questions and 16 validation\-selected difficult questions\.

### 5\.4\.Analysis

Beyond overall performance, we analyze PsyJev’s ability to distinguish question\-specific judgments, its inference efficiency, the reliability of its predicted probabilities, and its transfer to an unseen dataset, complemented by qualitative examples\.

#### 5\.4\.1\.Question\-Conditioned Judgment

We examine whether PsyJev can distinguish different psychological judgments about the same text\. Specifically, we evaluate co\-annotated binary question pairs with opposite ground\-truth labels, where successful prediction requires both questions to be answered correctly\. We compare PsyJev\-4B with GLiClass, Gemma\-2\-9B, and the corresponding Qwen3\.5\-4B answer model using identical texts, questions, and semantic answer sets\.

Table 4\.Question\-conditioned judgment on opposite\-label question pairs\. Both correct and same\-answer errors are reported as source\-balanced percentages\.As shown in Table[4](https://arxiv.org/html/2610.36689#S5.T4), PsyJev correctly answers both questions in 70\.06% of the 35,992 pairs from eight sources, outperforming the same\-size answer model \(66\.24%\), Gemma\-2\-9B \(67\.85%\), and GLiClass \(55\.84%\)\. Compared with the answer model, same\-answer errors decrease from 31\.95% to 28\.09%, with a paired text\-group 95% confidence interval of 1\.73–5\.96 percentage points for the improvement in paired correctness\. These results indicate stronger question\-conditioned discrimination on this matched\-pair subset, where the same text requires different answers under different criteria\. Additional outcome breakdowns are provided in Appendix[C\.3](https://arxiv.org/html/2610.36689#A3.SS3)\.

#### 5\.4\.2\.Inference Efficiency

Figure[3](https://arxiv.org/html/2610.36689#acmlabel3)compares model computation time and full\-test F1 for three trained 4B configurations\. Full\-answer and first\-token scoring use the answer\-SFT checkpoint, while PsyJev uses its final JA\-RLCD checkpoint\. On the fixed 13\-text training probe with CUDA Graphs on an H100 80GB, PsyJev requires 12\.01 ms per judgment, a2\.10×2\.10\\timesspeedup over full\-answer scoring \(25\.25 ms\)\. Its full\-test F1 is 2\.38 points higher than the first\-token configuration at comparable model computation time\. Appendix[B\.4](https://arxiv.org/html/2610.36689#A2.SS4)specifies the shared timing protocol\.

Figure 3\.F1–latency comparison of three 4B configurations\. PsyJev gains 2\.38 F1 points over answer\-SFT first\-token scoring at comparable GPU latency, while running 2\.10×\\timesfaster than full\-answer scoring under the shared timing protocol\.Horizontal bars compare PsyJev at 12\.01 milliseconds per judgment, first\-token scoring at 11\.87 milliseconds, and full\-answer scoring at 25\.25 milliseconds\. An arrow indicates the speedup over full\-answer scoring, while an annotation marks the F1 improvement over first\-token scoring\.Table 5\.Classification performance and probability quality on PsyJudgeBench\. F1 and ECE are percentages\. NLL is normalized by the log answer\-set size\. Bold indicates the best result among the displayed models\.Figure 4\.Source\-mean ECE accumulated over the full 0–100% confidence range\. Curves sum contributions from all 15 bins; their 100% endpoints reproduce Table[5](https://arxiv.org/html/2610.36689#S5.T5)\. Lower endpoints indicate better overall calibration\.Six curves show cumulative contributions to source\-mean ECE for PsyJev\-4B, GLiNER2, ModernBERT\-large, FLAN\-T5\-XL, MentalBERT, and MentalRoBERTa\. Both axes start at zero\. The horizontal axis covers all 15 confidence bins from zero to 100 percent; the vertical axis measures cumulative ECE in percentage points\. PsyJev\-4B ends at 4\.58, the lowest overall ECE among the six models, with a relative reduction of 3\.5 to 28\.5 percent over the five baselines\.
#### 5\.4\.3\.Prediction Quality

Beyond classification accuracy, we assess the quality of predicted answer distributions using Brier score, normalized NLL, and ECE\. Table[5](https://arxiv.org/html/2610.36689#S5.T5)compares PsyJev\-4B with representative encoder\-based, description\-driven, and answer\-generation models on the same test judgments, using source\-balanced probability metrics \(Appendix[C\.1](https://arxiv.org/html/2610.36689#A3.SS1)\)\. PsyJev\-4B achieves the highest F1 \(66\.21\) and the lowest Brier score \(0\.2777\) and normalized NLL \(0\.536\) among the models displayed in Table[5](https://arxiv.org/html/2610.36689#S5.T5), outperforming the strongest reference, GLiNER2, on all three metrics\. It also obtains the lowest ECE among these models \(4\.58%\) under the same source\-balanced evaluation\. Figure[4](https://arxiv.org/html/2610.36689#acmlabel4)shows how source\-mean calibration error accumulates across the full confidence range\.

Table 6\.Illustrative judgment corrections by PsyJev\-4B after JA\-RLCD\.×\\timesindicates disagreement with source annotations\. Displayed texts are excerpts of the original test inputs\.
#### 5\.4\.4\.Unseen\-Dataset Transfer

We evaluate transfer to Dreaddit, an unseen stress\-detection dataset with 715 test examples excluded from PsyJudgeBench supervision\([Turcan and McKeown, 2019](https://arxiv.org/html/2610.36689#bib.bib3)\)\. As shown in Table[7](https://arxiv.org/html/2610.36689#S5.T7), PsyJev achieves 77\.76 F1 at 2B and 78\.70 at 4B, improving over its direct\-judgment SFT initialization by 3\.85 and 0\.83 points, respectively\. Both variants also outperform their corresponding answer\-SFT baselines by 3\.44 and 2\.17 points\. The 4B result is competitive with Gemma\-3\-4B \(78\.26\) and Llama\-3\.2\-3B \(78\.00\)\. These results suggest that the benefits of judgment\-anchored refinement can extend beyond the benchmark’s training datasets, with gains varying by model scale\. Appendix[C\.2](https://arxiv.org/html/2610.36689#A3.SS2)reports the complete comparison\.

Table 7\.Dreaddit transfer on all 715 test examples \(F1, %\)\. No Dreaddit supervision is used\.ScaleModel / trainingF1 \(%\)2BQwen3\.5 answer\-SFT74\.32PsyJev SFT73\.91PsyJev JA\-RLCD77\.764BQwen3\.5 answer\-SFT76\.53PsyJev SFT77\.87PsyJev JA\-RLCD78\.703BLlama\-3\.2\-3B78\.004BGemma\-3\-4B78\.26
#### 5\.4\.5\.Qualitative Analysis

Table[6](https://arxiv.org/html/2610.36689#S5.T6)presents three test cases in which JA\-RLCD corrects predictions from its SFT initialization\. The cases cover co\-occurring emotions, competing sources of distress, and closely related cognitive distortion categories\. In Case A, PsyJev\-4B recovers the annotated emotion set without the additional Sadness label predicted by SFT\. In Case B, it identifies employment\-related distress despite concurrent relationship difficulties\. In Case C, it selects the annotated Overgeneralization category, whereas SFT favors Labeling\. The answer model also selects the gold category\. These examples illustrate how JA\-RLCD can improve label selection across different psychological judgment criteria\. Appendix[C\.4](https://arxiv.org/html/2610.36689#A3.SS4)shows the corresponding SFT\-to\-JA\-RLCD probability shifts\.

##### Scope of use\.

These judgments describe annotated textual content and do not establish clinical diagnoses\. Reuse should follow each source dataset’s access and usage terms and safeguard the privacy of sensitive self\-disclosures\.

## 6\.Conclusion

We presented PsyJudgeBench and PsyJev for unified psychological analysis of Web content through question\-conditioned judgments\. PsyJudgeBench organizes heterogeneous tasks while retaining their distinct decision criteria, and PsyJev enables direct probabilistic prediction withJA\-RLCDfor judgment\-anchored refinement\. Across 2B, 4B, and 9B scales, PsyJev achieves average F1 scores of 64\.33, 66\.21, and 67\.56, respectively, outperforming the corresponding answer\-SFT baselines\. Additional analyses examine question\-conditioned discrimination, inference efficiency, and probability quality\. These results demonstrate the value of a shared, question\-conditioned approach to heterogeneous psychological content analysis\.

## References

- Almeida \(2026\)D\. AlmeidaIntroducing system one models & Jev\.Note:TypeSafe AI blogExternal Links:[Link](https://typesafe.ai/blog/introducing-system-one-models-and-jev)Cited by:[§1](https://arxiv.org/html/2610.36689#S1.p5.1),[§2\.2](https://arxiv.org/html/2610.36689#S2.SS2.p1.1),[§5\.1\.2](https://arxiv.org/html/2610.36689#S5.SS1.SSS2.p1.1)\.
- Byunet al\.\(2026\)G\. Byun, R\. Lipschutz, S\. T\. Minton, A\. Powers, and J\. D\. ChoiCRADLE bench: a clinician\-annotated benchmark for multi\-faceted mental health crisis and safety risk detection\.InProceedings of the 19th Conference of the European Chapter of the Association for Computational Linguistics \(Volume 1: Long Papers\),Rabat, Morocco,pp\. 1572–1590\.External Links:[Document](https://dx.doi.org/10.18653/v1/2026.eacl-long.73),[Link](https://aclanthology.org/2026.eacl-long.73/)Cited by:[§2\.1](https://arxiv.org/html/2610.36689#S2.SS1.p1.1)\.
- Chunget al\.\(2024\)H\. W\. Chunget al\.Scaling instruction\-finetuned language models\.Journal of Machine Learning Research25\(70\),pp\. 1–53\.External Links:[Link](https://jmlr.org/papers/v25/23-0870.html)Cited by:[§5\.1\.2](https://arxiv.org/html/2610.36689#S5.SS1.SSS2.p1.1)\.
- Dubeyet al\.\(2024\)A\. Dubeyet al\.The Llama 3 herd of models\.Note:Version 1External Links:2407\.21783,[Link](https://arxiv.org/abs/2407.21783v1)Cited by:[§5\.1\.2](https://arxiv.org/html/2610.36689#S5.SS1.SSS2.p1.1)\.
- Gao and Wang \(2026\)Z\. Gao and P\. WangOpenJev\-RLCD: a working RLCD implementation\.External Links:2609\.38850,[Link](https://arxiv.org/abs/2609.38850)Cited by:[§1](https://arxiv.org/html/2610.36689#S1.p5.1),[§2\.2](https://arxiv.org/html/2610.36689#S2.SS2.p1.1)\.
- Garget al\.\(2023a\)M\. Garg, M\. Gaur, R\. Goswami, and S\. SohnLoST: a mental health dataset of low self\-esteem in Reddit posts\.In2023 IEEE International Conference on Systems, Man, and Cybernetics \(SMC\),Honolulu, Oahu, HI, USA,pp\. 3854–3859\.External Links:[Link](https://doi.org/10.1109/SMC53992.2023.10394671),[Document](https://dx.doi.org/10.1109/SMC53992.2023.10394671)Cited by:[§2\.1](https://arxiv.org/html/2610.36689#S2.SS1.p1.1)\.
- Garget al\.\(2024\)M\. Garg, X\. Liu, M\. S\. V\. P\. J\. Sathvik, S\. Raza, and S\. SohnMultiWD: multi\-label wellness dimensions in social media posts\.Journal of Biomedical Informatics150,pp\. 104586\.External Links:[Document](https://dx.doi.org/10.1016/j.jbi.2024.104586),[Link](https://doi.org/10.1016/j.jbi.2024.104586)Cited by:[§2\.1](https://arxiv.org/html/2610.36689#S2.SS1.p1.1)\.
- Garget al\.\(2022\)M\. Garg, C\. Saxena, S\. Saha, V\. Krishnan, R\. Joshi, and V\. MagoCAMS: an annotated corpus for causal analysis of mental health issues in social media posts\.InProceedings of the Thirteenth Language Resources and Evaluation Conference,Marseille, France,pp\. 6387–6396\.External Links:[Link](https://aclanthology.org/2022.lrec-1.686/)Cited by:[§2\.1](https://arxiv.org/html/2610.36689#S2.SS1.p1.1)\.
- Garget al\.\(2023b\)M\. Garg, C\. Saxena, D\. Samanta, and B\. J\. DorrLonXplain: lonesomeness as a consequence of mental disturbance in reddit posts\.External Links:2305\.18736,[Link](https://arxiv.org/abs/2305.18736)Cited by:[§2\.1](https://arxiv.org/html/2610.36689#S2.SS1.p1.1)\.
- Garget al\.\(2023c\)M\. Garg, A\. Shahbandegan, A\. Chadha, and V\. MagoAn annotated dataset for explainable interpersonal risk factors of mental disturbance in social media posts\.InFindings of the Association for Computational Linguistics: ACL 2023,Toronto, Canada,pp\. 11960–11969\.External Links:[Link](https://aclanthology.org/2023.findings-acl.757/),[Document](https://dx.doi.org/10.18653/v1/2023.findings-acl.757)Cited by:[§1](https://arxiv.org/html/2610.36689#S1.p2.1),[§2\.1](https://arxiv.org/html/2610.36689#S2.SS1.p1.1)\.
- Gemma Teamet al\.\(2024\)Gemma Teamet al\.Gemma 2: improving open language models at a practical size\.External Links:2408\.00118,[Link](https://arxiv.org/abs/2408.00118)Cited by:[§5\.1\.2](https://arxiv.org/html/2610.36689#S5.SS1.SSS2.p1.1)\.
- Gemma Teamet al\.\(2025\)Gemma Teamet al\.Gemma 3 technical report\.External Links:2503\.19786,[Link](https://arxiv.org/abs/2503.19786)Cited by:[§5\.1\.2](https://arxiv.org/html/2610.36689#S5.SS1.SSS2.p1.1)\.
- Gneiting and Raftery \(2007\)T\. Gneiting and A\. E\. RafteryStrictly proper scoring rules, prediction, and estimation\.Journal of the American Statistical Association102\(477\),pp\. 359–378\.External Links:[Document](https://dx.doi.org/10.1198/016214506000001437),[Link](https://doi.org/10.1198/016214506000001437)Cited by:[§4\.3](https://arxiv.org/html/2610.36689#S4.SS3.SSS0.Px2.p1.2)\.
- Guptaet al\.\(2022\)S\. Gupta, A\. Agarwal, M\. Gaur, K\. Roy, V\. Narayanan, P\. Kumaraguru, and A\. ShethLearning to automate follow\-up question generation using process knowledge for depression triage on Reddit posts\.InProceedings of the Eighth Workshop on Computational Linguistics and Clinical Psychology,Seattle, USA,pp\. 137–147\.External Links:[Document](https://dx.doi.org/10.18653/v1/2022.clpsych-1.12),[Link](https://aclanthology.org/2022.clpsych-1.12/)Cited by:[§1](https://arxiv.org/html/2610.36689#S1.p2.1),[§2\.1](https://arxiv.org/html/2610.36689#S2.SS1.p1.1),[§3](https://arxiv.org/html/2610.36689#S3.SS0.SSS0.Px1.p1.1)\.
- Hengleet al\.\(2024\)A\. Hengle, A\. Kulkarni, S\. D\. Patankar, M\. Chandrasekaran, S\. D’silva, J\. S\. Jacob, and R\. GuptaStill not quite there\! evaluating large language models for comorbid mental health diagnosis\.InProceedings of the 2024 Conference on Empirical Methods in Natural Language Processing,Miami, Florida, USA,pp\. 16698–16721\.External Links:[Document](https://dx.doi.org/10.18653/v1/2024.emnlp-main.931),[Link](https://aclanthology.org/2024.emnlp-main.931/)Cited by:[§1](https://arxiv.org/html/2610.36689#S1.p2.1),[§2\.1](https://arxiv.org/html/2610.36689#S2.SS1.p1.1)\.
- Jiet al\.\(2022\)S\. Ji, T\. Zhang, L\. Ansari, J\. Fu, P\. Tiwari, and E\. CambriaMentalBERT: publicly available pretrained language models for mental healthcare\.InProceedings of the Thirteenth Language Resources and Evaluation Conference,Marseille, France,pp\. 7184–7190\.External Links:[Link](https://aclanthology.org/2022.lrec-1.778/)Cited by:[§2\.1](https://arxiv.org/html/2610.36689#S2.SS1.p1.1),[§5\.1\.2](https://arxiv.org/html/2610.36689#S5.SS1.SSS2.p1.1)\.
- Laya Contributors \(2026\)Laya ContributorsLaya\.Note:Commit 682ac6719faa5b1c04b529501feb93e58ee91f3dSoftware repositoryExternal Links:[Link](https://github.com/NandhaKishorM/laya/tree/682ac6719faa5b1c04b529501feb93e58ee91f3d)Cited by:[§B\.3](https://arxiv.org/html/2610.36689#A2.SS3.SSS0.Px3.p1.1),[§1](https://arxiv.org/html/2610.36689#S1.p5.1),[§2\.2](https://arxiv.org/html/2610.36689#S2.SS2.p1.1),[§4\.3](https://arxiv.org/html/2610.36689#S4.SS3.p1.1),[§5\.1\.2](https://arxiv.org/html/2610.36689#S5.SS1.SSS2.p1.1)\.
- Maurielloet al\.\(2021\)M\. L\. Mauriello, E\. T\. Lincoln, G\. Hon, D\. Simon, D\. Jurafsky, and P\. E\. ParedesSAD: a stress annotated dataset for recognizing everyday stressors in SMS\-like conversational systems\.InExtended Abstracts of the 2021 CHI Conference on Human Factors in Computing Systems,New York, NY, USA,pp\. 1–7\.External Links:[Document](https://dx.doi.org/10.1145/3411763.3451799),[Link](https://sensifylab.cis.udel.edu/docs/2021-Mauriello-CHI-StressAnnotatedDataset-Abstract.pdf)Cited by:[§1](https://arxiv.org/html/2610.36689#S1.p2.1),[§2\.1](https://arxiv.org/html/2610.36689#S2.SS1.p1.1)\.
- Meta \(2024\)MetaLlama\-3\.2\-3B\-Instruct model card\.Note:Hugging Face model repositoryExternal Links:[Link](https://huggingface.co/meta-llama/Llama-3.2-3B-Instruct)Cited by:[§5\.1\.2](https://arxiv.org/html/2610.36689#S5.SS1.SSS2.p1.1)\.
- Microsoftet al\.\(2025\)Microsoftet al\.Phi\-4\-Mini technical report: compact yet powerful multimodal language models via Mixture\-of\-LoRAs\.External Links:2503\.01743,[Link](https://arxiv.org/abs/2503.01743)Cited by:[§5\.1\.2](https://arxiv.org/html/2610.36689#S5.SS1.SSS2.p1.1)\.
- OpenJev Contributors \(2026\)OpenJev ContributorsOpenJev: learning open\-vocabulary probabilistic decision models from pretrained language models\.Note:Open\-source implementationExternal Links:[Link](https://github.com/ejhshen/OpenJev)Cited by:[§2\.2](https://arxiv.org/html/2610.36689#S2.SS2.p1.1)\.
- Pérezet al\.\(2023\)A\. Pérez, J\. Parapar, Á\. Barreiro, and S\. López\-LarrosaBDI\-Sen: a sentence dataset for clinical symptoms of depression\.InProceedings of the 46th International ACM SIGIR Conference on Research and Development in Information Retrieval,New York, NY, USA,pp\. 2996–3006\.External Links:[Document](https://dx.doi.org/10.1145/3539618.3591905),[Link](https://doi.org/10.1145/3539618.3591905)Cited by:[§1](https://arxiv.org/html/2610.36689#S1.p2.1),[§2\.1](https://arxiv.org/html/2610.36689#S2.SS1.p1.1),[§3](https://arxiv.org/html/2610.36689#S3.SS0.SSS0.Px1.p1.1)\.
- Qwen Team \(2024\)Qwen TeamQwen2\.5: a party of foundation models\!\.Note:Baseline checkpoint: Qwen/Qwen2\.5\-7B\-InstructExternal Links:[Link](https://qwenlm.github.io/blog/qwen2.5/)Cited by:[§5\.1\.2](https://arxiv.org/html/2610.36689#S5.SS1.SSS2.p1.1)\.
- Qwen Team \(2026\)Qwen TeamQwen3\.5\.Note:Hugging Face model collectionExternal Links:[Link](https://huggingface.co/collections/Qwen/qwen35)Cited by:[§5\.1\.2](https://arxiv.org/html/2610.36689#S5.SS1.SSS2.p1.1)\.
- Rahmanet al\.\(2024\)A\. B\. S\. Rahman, H\. T\. Ta, L\. Najjar, A\. Azadmanesh, and A\. S\. GönülDepressionEmo: a novel dataset for multilabel classification of depression emotions\.Journal of Affective Disorders366,pp\. 445–458\.External Links:[Document](https://dx.doi.org/10.1016/j.jad.2024.08.013),[Link](https://doi.org/10.1016/j.jad.2024.08.013)Cited by:[§2\.1](https://arxiv.org/html/2610.36689#S2.SS1.p1.1)\.
- Sharmaet al\.\(2023\)A\. Sharma, K\. Rushton, I\. W\. Lin, D\. Wadden, K\. G\. Lucas, A\. S\. Miner, T\. Nguyen, and T\. AlthoffCognitive reframing of negative thoughts through human\-language model interaction\.InProceedings of the 61st Annual Meeting of the Association for Computational Linguistics \(Volume 1: Long Papers\),Toronto, Canada,pp\. 9977–10000\.External Links:[Link](https://aclanthology.org/2023.acl-long.555/),[Document](https://dx.doi.org/10.18653/v1/2023.acl-long.555)Cited by:[§2\.1](https://arxiv.org/html/2610.36689#S2.SS1.p1.1)\.
- Shreevastava and Foltz \(2021\)S\. Shreevastava and P\. W\. FoltzDetecting cognitive distortions from patient\-therapist interactions\.InProceedings of the Seventh Workshop on Computational Linguistics and Clinical Psychology: Improving Access,Online,pp\. 151–158\.External Links:[Document](https://dx.doi.org/10.18653/v1/2021.clpsych-1.17),[Link](https://aclanthology.org/2021.clpsych-1.17/)Cited by:[§2\.1](https://arxiv.org/html/2610.36689#S2.SS1.p1.1)\.
- Stepanovet al\.\(2025\)I\. Stepanov, M\. Shtopko, D\. Vodianytskyi, O\. Lukashov, A\. Yavorskyi, and M\. YaroshenkoGLiClass: generalist lightweight model for sequence classification tasks\.External Links:2508\.07662,[Link](https://arxiv.org/abs/2508.07662)Cited by:[§2\.1](https://arxiv.org/html/2610.36689#S2.SS1.p1.1),[§5\.1\.2](https://arxiv.org/html/2610.36689#S5.SS1.SSS2.p1.1)\.
- Turcan and McKeown \(2019\)E\. Turcan and K\. McKeownDreaddit: a reddit dataset for stress analysis in social media\.InProceedings of the Tenth International Workshop on Health Text Mining and Information Analysis \(LOUHI 2019\),Hong Kong,pp\. 97–107\.External Links:[Document](https://dx.doi.org/10.18653/v1/D19-6213),[Link](https://aclanthology.org/D19-6213/)Cited by:[§C\.2](https://arxiv.org/html/2610.36689#A3.SS2.p1.1),[§5\.4\.4](https://arxiv.org/html/2610.36689#S5.SS4.SSS4.p1.1)\.
- Warneret al\.\(2025\)B\. Warner, A\. Chaffin, B\. Clavié, O\. Weller, O\. Hallström, S\. Taghadouini, A\. Gallagher, R\. Biswas, F\. Ladhak, T\. Aarsen, N\. Cooper, G\. Adams, J\. Howard, and I\. PoliSmarter, better, faster, longer: a modern bidirectional encoder for fast, memory efficient, and long context finetuning and inference\.InProceedings of the 63rd Annual Meeting of the Association for Computational Linguistics \(Volume 1: Long Papers\),Vienna, Austria,pp\. 2526–2547\.External Links:[Link](https://aclanthology.org/2025.acl-long.127/),[Document](https://dx.doi.org/10.18653/v1/2025.acl-long.127)Cited by:[§5\.1\.2](https://arxiv.org/html/2610.36689#S5.SS1.SSS2.p1.1)\.
- Xuet al\.\(2024\)X\. Xu, B\. Yao, Y\. Dong, S\. Gabriel, H\. Yu, J\. Hendler, M\. Ghassemi, A\. K\. Dey, and D\. WangMental\-LLM: leveraging large language models for mental health prediction via online text data\.Proceedings of the ACM on Interactive, Mobile, Wearable and Ubiquitous Technologies8\(1\)\.External Links:[Link](https://doi.org/10.1145/3643540),[Document](https://dx.doi.org/10.1145/3643540)Cited by:[§2\.1](https://arxiv.org/html/2610.36689#S2.SS1.p1.1)\.
- Yanget al\.\(2025\)A\. Yang, A\. Li, B\. Yang,et al\.Qwen3 technical report\.External Links:2505\.09388,[Link](https://arxiv.org/abs/2505.09388)Cited by:[§5\.1\.2](https://arxiv.org/html/2610.36689#S5.SS1.SSS2.p1.1)\.
- Yanget al\.\(2024\)K\. Yang, T\. Zhang, Z\. Kuang, Q\. Xie, J\. Huang, and S\. AnaniadouMentaLLaMA: interpretable mental health analysis on social media with large language models\.InProceedings of the ACM Web Conference 2024,New York, NY, USA,pp\. 4489–4500\.External Links:[Document](https://dx.doi.org/10.1145/3589334.3648137),[Link](https://doi.org/10.1145/3589334.3648137)Cited by:[§B\.3](https://arxiv.org/html/2610.36689#A2.SS3.p1.1),[§1](https://arxiv.org/html/2610.36689#S1.p2.1),[§2\.1](https://arxiv.org/html/2610.36689#S2.SS1.p1.1),[§5\.1\.2](https://arxiv.org/html/2610.36689#S5.SS1.SSS2.p1.1)\.
- Zaratianaet al\.\(2025\)U\. Zaratiana, G\. Pasternak, O\. Boyd, G\. Hurn\-Maloney, and A\. LewisGLiNER2: schema\-driven multi\-task learning for structured information extraction\.InProceedings of the 2025 Conference on Empirical Methods in Natural Language Processing: System Demonstrations,Suzhou, China,pp\. 130–140\.External Links:[Link](https://aclanthology.org/2025.emnlp-demos.10/),[Document](https://dx.doi.org/10.18653/v1/2025.emnlp-demos.10)Cited by:[§2\.1](https://arxiv.org/html/2610.36689#S2.SS1.p1.1),[§5\.1\.2](https://arxiv.org/html/2610.36689#S5.SS1.SSS2.p1.1)\.
- Zhaiet al\.\(2025\)W\. Zhai, N\. Bai, Q\. Zhao, J\. Li, F\. Wang, H\. Qi, M\. Jiang, X\. Wang, B\. X\. Yang, and G\. FuMentalGLM series: explainable large language models for mental health analysis on Chinese social media\.InProceedings of the 2025 Conference on Empirical Methods in Natural Language Processing,Suzhou, China,pp\. 13588–13603\.External Links:[Link](https://aclanthology.org/2025.emnlp-main.686/),[Document](https://dx.doi.org/10.18653/v1/2025.emnlp-main.686)Cited by:[§2\.1](https://arxiv.org/html/2610.36689#S2.SS1.p1.1)\.
- Zhuet al\.\(2026\)Y\. Zhu, X\. Zeng, Q\. Si, C\. Yang, Y\. Li, L\. Qin, J\. Zhou, D\. Yao, J\. Zhang, C\. Qin, Z\. Zhang, Y\. Yang, Y\. Xu, T\. Zhang, H\. Guo, J\. Wang, N\. Duan, and L\. WangVideoJev: system one model for video understanding and interaction\.External Links:[Link](https://arxiv.org/abs/2610.18615),2610\.18615Cited by:[§2\.2](https://arxiv.org/html/2610.36689#S2.SS2.p1.1)\.

## Appendix AData Construction and Splits

Table 8\.Source records and coordinated splits\. Q counts registry questions: 79 binary and three multiclass in total\.##### Label construction\.

Missing labels are masked\. BDI\-Sen retains symptom relevance, excluding severity\. TherapistQA uses the patient question and dominant distortion, excluding therapist responses, spans, and secondary labels\. Thinking Traps merges normalized repeated completions into 15\-category label sets; CRADLE expands seven crisis categories across ongoing/past status into 14 questions\. Missing CAMS categories remain unobserved\. Table[9](https://arxiv.org/html/2610.36689#A1.T9)gives four verbatim registry entries\.

Table 9\.Verbatim registry instructions and legal answers for four judgment types in PsyJudgeBench\.
##### Split coordination\.

Exact texts and available source groups are coordinated across all 13 sources, with test taking priority over validation and training\. Normalization applies Unicode NFKC, case folding, and whitespace collapse\. Shared texts retain their distinct source\-defined judgments\. We mask 22 conflicting judgments across 11 identical text–target keys\.

##### TherapistQA refinement\.

After removing two identical questions with conflicting labels, normalized\-question groups receive a seed\-42 hash split, inheriting exact\-match assignments\. Candidates use up to 12 smallest distinct CRC32 four\-token shingle hashes and the first 128 indexed texts per hash\. Components satisfying the Jaccard or containment criteria in Table[10](https://arxiv.org/html/2610.36689#A1.T10)follow the same held\-out priority; the other 12 sources remain unchanged\. ANGST retains silver training/validation and expert\-labeled test annotations; CRADLE’s coordinated test partition contains 600 gold and one silver record\.

##### Final\-split audit\.

The frozen snapshot contains 53,775 records and 41,349 normalized texts\. All 3,600 cross\-source shared\-text groups are confined to one split \(training: 1,301; validation: 494; test: 1,805\), including all 3,518 IRF–LonXplain shared texts\. Table[10](https://arxiv.org/html/2610.36689#A1.T10)checks every partition pair using uncapped prefix filtering and exact four\-word shingles, on both full texts and their first 512 tokens\. No source\-group identifier crosses partitions\. The audit measures lexical instance reuse\.

Table 10\.Cross\-partition overlap on full texts and 512\-token inputs\.

## Appendix BReproducibility and Comparison Protocols

### B\.1\.Input Processing

We useQwen/Qwen3\.5\-2B,Qwen/Qwen3\.5\-4B, andQwen/Qwen3\.5\-9B\(revision prefixes15852e8c,851bf6e8, andc2022362\), each with its native tokenizer\. Input texts are encoded without added special tokens and truncated to 512 tokens, while question criteria and legal answer descriptions are preserved in full and in registry order\. Requests serialize the text underContext:, followed byQuestion:and letter\-prefixed answer options\. We apply the backbone chat template with thinking disabled and append a separately tokenizedAnswer:prefix for constrained readout\. For multilabel sources, each observed binary question is scored separately; a positive label is assigned only whenp⁡\(yes\)\>p⁡\(no\)p\(\\mathrm\{yes\}\)\>p\(\\mathrm\{no\}\), with ties assigned to no\.

### B\.2\.Training and Matched Controls

##### Optimizer and accumulation\.

All reported runs use seed 42\. Each stage visits all 33,511 training records once in 524 optimizer updates, with effective batches of 63 or 64 records, retaining its final checkpoint\. All observed questions are retained in record/question order, grouped by source, and split into consecutive source\-homogeneous microbatches of at most 16 judgments\. Their weighted gradients are summed over each complete record batch before one optimizer step; the microbatch count varies with the observed labels\. AdamW usesβ1=0\.9\\beta\_\{1\}=0\.9,β2=0\.999\\beta\_\{2\}=0\.999,ϵ=10−8\\epsilon=10^\{\-8\}, weight decay 0\.01, global gradient clipping at 1, and 3% warmup followed by cosine decay\. The 2B model uses full tuning at2×10−62\\times 10^\{\-6\}in both stages\. At 4B/9B, LoRA uses rank 16, scaling 32, and zero dropout, targeting attention projections \(q/k/v/o\_proj\), linear\-attention input/output projections \(in\_proj\_qkv,in\_proj\_z,in\_proj\_b,in\_proj\_a, andout\_proj\), and MLP gate, up, and down projections\. The vocabulary head remains frozen; all trainable adapters share learning rates of2×10−42\\times 10^\{\-4\}for SFT and2×10−62\\times 10^\{\-6\}for refinement\.

##### Weights and optimization\.

For training countsnq​kn\_\{qk\}andnq=∑knq​kn\_\{q\}=\\sum\_\{k\}n\_\{qk\}, we set

\(8\)π^q​k=nq​k\+1nq\+Kq,cq​k=min⁡\(8,ηq/π^q​k\),\\widehat\{\\pi\}\_\{qk\}=\\frac\{n\_\{qk\}\+1\}\{n\_\{q\}\+K\_\{q\}\},\\qquad c\_\{qk\}=\\min\(8,\\eta\_\{q\}/\\widehat\{\\pi\}\_\{qk\}\),whereηq\\eta\_\{q\}ensures∑kπ^q​k​cq​k=1\\sum\_\{k\}\\widehat\{\\pi\}\_\{qk\}c\_\{qk\}=1\. ForNNtraining records,SSsources, effective record batch sizeBB, and source question set𝒬s\\mathcal\{Q\}\_\{s\},

\(9\)ai=NB​S​\|𝒬si\|​nqi\.a\_\{i\}=\\frac\{N\}\{BS\|\\mathcal\{Q\}\_\{s\_\{i\}\}\|n\_\{q\_\{i\}\}\}\.
The anchor appliesaia\_\{i\}per judgment to balance source/question observation frequency; the reward branch uses only their sumAℬA\_\{\\mathcal\{B\}\}in Eq\.[6](https://arxiv.org/html/2610.36689#S4.E6)\. Priors and weights derive from training counts; reference\-correctness gates also use training labels only\. Refinement initializes from the same\-scale final SFT checkpoint, keeping a frozen reference copy\. Each microbatch uses two Gaussian directions and their negatives, with all four reward evaluations sharing the same model logits\. The noise scale decreases linearly from0\.40\.4to0\.10\.1over refinement updates;τ=4\\tau=4andλ=0\.05\\lambda=0\.05\. Withuuthe pathwise reward gradient, the update accumulatesg−ug\-uover microbatches and clips the total before AdamW\. The refinement configuration was selected using validation only, without test results\.

### B\.3\.Baseline Protocols

All benchmark\-adapted baselines use the same cleaned records, question registry, and observed\-label masks\. Fixed\-head classifiers use question\-specific output heads, while description\-driven models preserve the original decision criteria in their input schemas\. Answer\-SFT models score complete legal\-answer sequences, including EOS\. All models use their final training checkpoints, except ModernBERT, which selects among three epochs based on validation F1, followed by NLL and earlier\-epoch tie\-breaking\. MentaLLaMA undergoes benchmark\-specific answer SFT from its existing mental\-health instruction\-tuned checkpoint\([Yang et al\., 2024](https://arxiv.org/html/2610.36689#bib.bib9)\)\.

##### Open\-source Laya\.

We reproduce Laya’s SFT and RLCD training pipeline on PsyJudgeBench using the released[Laya\-multilingual checkpoint](https://huggingface.co/convaiinnovations/laya-multilingual/tree/1720e3e3357cfe1e281542e223f8273b0890ca34)\(revision1720e3e3\), preserving its native architecture and RLCD objective\. We first perform one epoch of full\-parameter SFT \(524 updates, learning rate10−510^\{\-5\}\), followed by one epoch of RLCD \(524 updates, learning rate2×10−62\\times 10^\{\-6\}\)\. RLCD combines cross\-entropy with score\-function optimization over Gaussian\-perturbed answer logits, using four samples, a noise schedule from0\.40\.4to0\.10\.1, and spherical/ranked\-probability reward weights of0\.75/1\.00\.75/1\.0\. Both stages use the benchmark training partition and are evaluated using their final checkpoints\.

##### Jev API evaluation\.

We evaluatejev\-1\.14\.0through Typesafe’s official API on all 44,999 test judgments\. Each request includes the source text, question criterion, and legal answers, with the source text capped at 512 Qwen tokens\. Binary and multiclass questions use Noul and Choice, respectively\. The API is evaluated zero\-shot without benchmark fine\-tuning or demonstrations, and its returned probabilities are aligned with the benchmark’s semantic labels\. All 84 invalid responses remain in the evaluation as incorrect predictions; the invalid label is excluded from the legal\-class macro average when computing F1\. Results follow the same question\- and source\-level aggregation as the main evaluation\.

##### Matched refinement controls\.

All direct\-scoring 2B continuations start independently from the same final SFT checkpoint and retain the same records, registry, observed\-label masks, record schedule, trainable parameters, optimizer, and 524 additional updates\. Adapted RLCD combines ordinary CE with Laya’s categorical score\-function objective\([Laya Contributors, 2026](https://arxiv.org/html/2610.36689#bib.bib12)\): four independent centered Gaussian logit actions, advantages centered and normalized by their standard deviation, and reward weights of 0\.75 for the spherical term and 1\.0 for ranked probability\. It uses SFT source/record weights without the JA reference gate or question\-group aggregation\.

##### Ablation configurations\.

The anchor\-only variant removes reward optimization while retaining balanced supervision and selective reference KL\. Removing class costs setscq​k=1c\_\{qk\}=1in both the anchor and reward objectives, while removing reference KL setsλ=0\\lambda=0\. The score\-function variant replaces pathwise differentiation with paired score\-function gradients, retaining the same anchor, question\-group reward, and logit\-space perturbations\.

The balanced\-CE\-only control disables both reference KL and reward optimization, retaining the source/question weights and class costs\. The ordinary Brier control replaces the group/tail reward with one minus the class\-cost\-normalized mean per\-judgment half\-Brier error, retaining the anchor, perturbations, reward scale, and pathwise estimator\. This changes grouping and tail weighting jointly\. The answer\-sequence balanced\-CE control starts from its corresponding answer\-SFT checkpoint and uses full\-vocabulary CE on the answer letter and EOS\. It matches the direct balanced\-CE control’s pretrained backbone, record schedule, weights, optimizer, and two total training epochs, including 524 continuation updates\.

### B\.4\.Unified Timing Protocol

Figure 5\.Quality–latency comparison of the 16 models with average F1≥57%\\geq 57\\%in Table[1](https://arxiv.org/html/2610.36689#S5.T1)\.A scatter plot focuses on models with average F1 of at least 57 percent and uses a linear latency axis\. It shows PsyJev at 2B, 4B, and 9B, alongside thirteen baselines\. Dashed arrows pair PsyJev and Qwen3\.5 answer\-SFT at matching backbone sizes\.Model computation time is measured on a shared probe comprising 13 training texts \(the second record in each source’s stored training order\) and 64 observed judgments\. Each request contains one text and its observed questions and is repeated five times after warmup\. All models run on a single NVIDIA H100 80GB using BF16 and CUDA Graphs\. CUDA events measure GPU execution, including text prefill and answer scoring, while excluding loading, tokenization, data transfer, and graph capture\. Inputs reside on the GPU, and latency is reported as total measured time divided by the number of judgments\. Each model retains its native scoring mechanism under the same protocol\. For the matched 4B comparison, full\-answer and first\-token scoring share the Qwen3\.5\-4B answer\-SFT checkpoint, while PsyJev uses its final JA\-RLCD checkpoint\. Their respective latencies are 25\.25, 11\.87, and 12\.01 ms per judgment\. Figure[3](https://arxiv.org/html/2610.36689#acmlabel3)pairs these measurements with full\-test F1, while Figure[5](https://arxiv.org/html/2610.36689#acmlabel5)compares quality and latency across the models in Table[1](https://arxiv.org/html/2610.36689#S5.T1)\.

## Appendix CPrediction Diagnostics

### C\.1\.Probability Evaluation Protocol

All probability evaluations use the same 44,999 test judgments and legal answer sets\. Classification models normalize their answer scores over the legal options, while answer\-generation models normalize complete answer\-sequence likelihoods, including EOS\. No post\-hoc calibration is applied\. For a judgment withKKanswers, normalized NLL and Brier score are defined as:

\(10\)NLL=−log⁡pylog⁡K,Brier=∑k=1K\(pk−𝕀\[k=y\]\)2\.\\mathrm\{NLL\}=\-\\frac\{\\log p\_\{y\}\}\{\\log K\},\\qquad\\mathrm\{Brier\}=\\sum\_\{k=1\}^\{K\}\(p\_\{k\}\-\\mathbb\{I\}\[k=y\]\)^\{2\}\.
NLL is averaged over judgments within records and then equally across records and sources\. Brier score and 15\-bin ECE are computed at the source level and averaged equally across sources\.

For Figure[4](https://arxiv.org/html/2610.36689#acmlabel4), letns,bn\_\{s,b\},as,ba\_\{s,b\}, andcs,bc\_\{s,b\}denote the count, accuracy, and mean confidence in sourcessand binbb\. The source\-bin contribution and cumulative curve are

es,b=ns,bns​\|as,b−cs,b\|,E≤b=113​∑s∑j=1bes,j\.e\_\{s,b\}=\\frac\{n\_\{s,b\}\}\{n\_\{s\}\}\|a\_\{s,b\}\-c\_\{s,b\}\|,\\qquad E\_\{\\leq b\}=\\frac\{1\}\{13\}\\sum\_\{s\}\\sum\_\{j=1\}^\{b\}e\_\{s,j\}\.Empty bins contribute zero\. The denominatornsn\_\{s\}is the full source judgment count, soE≤15E\_\{\\leq 15\}equals the source\-mean ECE in Table[5](https://arxiv.org/html/2610.36689#S5.T5)\.

Table 11\.Probability quality of the final PsyJev\-4B by source\. The mean gives equal weight to all 13 sources\.
### C\.2\.Unseen\-Dataset Transfer

All models use their benchmark\-trained checkpoints on the complete 715\-example Dreaddit test set\([Turcan and McKeown, 2019](https://arxiv.org/html/2610.36689#bib.bib3)\), without Dreaddit supervision or further tuning\. The registered question is*Does the author express psychological stress in this text?*Legal answers are ordered \(no,yes\), mapping original labels 0 and 1, respectively\. Texts retain their first 512 tokens; prediction uses legal\-answer argmax, and F1 treatsyesas positive\. Auditing all 715 texts against the actual PsyJudgeBench training \(33,511\) and validation \(5,852\) records found no raw/normalized exact matches, no word\-level 4\-gram Jaccard matches≥0\.85\\geq 0\.85with uncapped candidate retrieval, and no full\-text containments with the shorter text at least 80 characters\. Table[12](https://arxiv.org/html/2610.36689#A3.T12)lists all 12 configurations\.

Table 12\.Complete Dreaddit transfer results \(F1, %\) without additional training on Dreaddit\.Table 13\.Different questions about the same text\. Outcomes \(%\) use identical opposite\-label binary pairs and equal source weights\. “Same answer” means exactly one prediction is wrong in this subset\.
### C\.3\.Questions about Shared Text

Table[13](https://arxiv.org/html/2610.36689#A3.T13)adds PsyJev\-4B SFT and both\-wrong outcomes to the main comparison\. All models use identical opposite\-label binary pairs, with three outcomes\. Results are source\-balanced, with paired text\-group resampling accounting for shared\-text dependence\.

### C\.4\.Case\-Level Probability Shifts

Table[14](https://arxiv.org/html/2610.36689#A3.T14)follows the three 4B cases in Table[6](https://arxiv.org/html/2610.36689#S5.T6), using the final JA\-RLCD checkpoint and its SFT parent\. In Case A, the Sadness probability crosses below 0\.5 while Emptiness and Loneliness remain positive\. In Cases B and C, the gold label overtakes the strongest competing label; both labels’ probabilities are shown to distinguish relative preference changes from uniform confidence increases\. These cases use DepressionEmo:16ll3jf, CAMS:1286, and TherapistQA:364\. Models receive the original inputs; display excerpts preserve their order, mark omissions, and normalize punctuation\.

Table 14\.Answer probabilities before and after JA\-RLCD for the three cases in Table[6](https://arxiv.org/html/2610.36689#S5.T6)\.