UniJev: Separating Accuracy from Calibration in Trimodal System One Decision Models

Jianing Wang; Youwen Jian; Derong Wang; Jinyang Yu; Qingyi Si; Chuanyu Qin; Yanghua Xiao; Hongcheng Guo

cs.CVOct 19, 2026

UniJev: Separating Accuracy from Calibration in Trimodal System One Decision Models

Authors: Jianing Wang, Youwen Jian, Derong Wang, Jinyang Yu, Qingyi Si, Chuanyu Qin, Yanghua Xiao, Hongcheng Guo

Organizations: Fudan University · Institute of Information Engineering, Chinese Academy of Sciences

Abstract

System One models produce probabilistic judgments that software can consume directly, making reliable confidence as important as answer accuracy. Yet joint training across text, images, and video remains underexplored, and the contributions of supervised fine-tuning (SFT) and reinforcement learning for calibrated decisions (RLCD) are difficult to separate. We introduce UniJev, a trimodal framework that scores candidate answers with a shared head under joint supervised training across text, images, and video. Supervised decision models improve accuracy over their backbone readouts across all three modalities, while the existing RLCD algorithms change accuracy little and have inconsistent effects on visual calibration across recipes. These findings motivate a separate supervised objective for confidence calibration. We introduce Cora (Confidence--Outcome Reliability Alignment), which groups current predictions by confidence and optimizes each prediction's confidence toward the empirical accuracy of its bin. At both 2B and 9B, Cora lowers image and video expected calibration error (ECE) while preserving accuracy, without a reinforcement stage. We release our code at https://github.com/wjnwjnwj/UniJev.

Figures & tables

Figure 1: Motivation for reliable trimodal decisions. Joint trimodal training remains underexplored, and the effects of SFT and RLCD on accuracy and calibration are difficult to separate.

Method	Training signal	Estimator	Reference constraint
TypeSafe Jev	–	–	–
Laya / Dohnuts	Log + spherical; RPS	Gaussian SF	Soft CE
Astro-RLCD	Log + spherical	Gaussian SF + LOO	None
Valen	$c-(p_{a}-c)^{2}$	Clipped SF + GN	KL + Brier
NanoJev	$2p_{y}-\lVert p\rVert^{2}$	Paired SF + LOO	None
Decider RL	Return + belief log score	Clipped SF + direct	Retention KL

Table 1: Training designs of the methods compared in Table 2 . The reference constraint column lists each recipe’s regularization toward a reference model or supervised objective, used to keep updates close to the initialization.

c

: sampled correctness;

p_{a}

: selected-answer probability;

y

: observed label; SF: score-function (REINFORCE-style) gradient; LOO: leave-one-out; GN: group normalization; RPS: ranked probability score. A listed recipe need not describe its released checkpoint; evaluation conventions and selected revisions appear in Appendix A .

Figure 2: Overview of UniJev and the main findings. A shared trimodal decision model first learns candidate selection through the supervised conversion, then CORA directly aligns confidence with the empirical accuracy of current confidence groups. Most of the observed accuracy gain occurs during the supervised conversion; the evaluated reward-driven continuations do not consistently improve visual calibration; and CORA lowers image and video ECE at both model scales while preserving accuracy.

General-purpose models
		Text (764)					Image (1,858)					Video (502)
Model / method	B	Acc.	Conf.	ECE $\downarrow$	NLL $\downarrow$	Brier $\downarrow$	Acc.	Conf.	ECE $\downarrow$	NLL $\downarrow$	Brier $\downarrow$	Acc.	Conf.	ECE $\downarrow$	NLL $\downarrow$	Brier $\downarrow$
Qwen3.5-2B	2	66.1	0.673	0.105	0.864	0.484	68.7	0.706	0.058	0.831	0.379	67.1	0.708	0.050	0.771	0.421
Qwen3.5-9B	9	76.8	0.828	0.060	0.587	0.322	76.7	0.811	0.043	0.637	0.296	73.3	0.748	0.031	0.586	0.327
Qwen3.8-27B	27	84.7	0.842	0.028	0.469	0.233	78.9	0.780	0.027	0.627	0.286	74.5	0.807	0.075	0.655	0.352
InternVL3.5-8B	8	76.4	0.893	0.132	0.823	0.365	83.0	0.881	0.053	0.516	0.229	71.5	0.798	0.084	0.703	0.369

Table 2: Model comparison on text (764 test questions), image (1,858), and video (502). Each modality reports accuracy (Acc., %), mean maximum candidate probability (Conf.), 15-bin ECE, NLL, and multiclass Brier. The CORA rows report the final checkpoints after 100 supervised calibration updates on the trimodal training set. B denotes nominal parameter scale in billions. – denotes an unsupported modality or undisclosed scale. Released checkpoints use their native interfaces without local retraining. Our supervised controls use hard-label SFT at 2B/9B and soft-label SFT at 2B/9B. Local 2B recipe rows use a shared soft-label SFT initialization and a 500-step budget.

Default
	Text (764)			Image (1,858)			Video (502)
Arm	Acc.	Conf.	ECE $\downarrow$	Acc.	Conf.	ECE $\downarrow$	Acc.	Conf.	ECE $\downarrow$
CORA (10 bins, group 256, support 20)	71.9	0.767	0.055	83.6	0.851	0.030	74.5	0.767	0.052
Primary single-factor contrasts
Bin accuracy $\to$ bin mean confidence	71.9	0.989	0.274	83.0	0.992	0.162	74.7	0.986	0.240
Confidence bins $\to$ one global target	72.1	0.875	0.158	82.7	0.830	0.127	74.5	0.823	0.090

Table 3: CORA ablations and sensitivity on the test suite. Every arm shares the soft-label SFT initialization, training mixture, and 100-update protocol; each row changes one factor relative to the default and reports its final checkpoint. The default row is identical to the 2B CORA row in Table 2 .

Appendix figures & tables9 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 3: CORA training dynamics over 100 updates. Left: per-update group loss (thin) and its 25-update running mean (bold), shown on a logarithmic scale. Center: fraction of each 256-question group included in the loss after applying the minimum-support rule. Right: periodic monitoring accuracy, binary Brier score, and 15-bin ECE.

Figure 4: Text cases T1–T8: three shared correct answers, three shared errors, and two mixed outcomes across science, knowledge, and emotion questions. Asterisks mark annotated answers; numbers are confidence percentages. Candidate letters are display aliases, preserving the order and probabilities used in evaluation.

Figure 5: Image cases I1–I8: three shared correct answers, three shared errors, and two mixed outcomes across spatial, interface, counting, document, scene-text, chart, and table questions. Predictions and confidences come from the same checkpoints as Table 2 .

Figure 6: Video cases V1–V6: two shared correct answers, two shared errors, and two mixed outcomes for actions, reactions, and scene context. Timestamps locate the displayed excerpts within each cached input.

General-purpose models
		Text (764)				Image (1,858)				Video (502)
Method	B	$\Delta$	${>}0.95$	REL	AUROC	$\Delta$	${>}0.95$	REL	AUROC	$\Delta$	${>}0.95$	REL	AUROC
Qwen3.5-2B	2	+0.012	13%	13.4	0.667	+0.019	35%	5.5	0.913	+0.037	20%	5.3	0.804
Qwen3.5-9B	9	+0.060	41%	6.7	0.831	+0.043	56%	4.2	0.916	+0.015	31%	2.2	0.865
Qwen3.8-27B	27	-0.004	46%	3.5	0.832	-0.009	43%	1.4	0.900	+0.062	39%	9.5	0.831
InternVL3.5-8B	8	+0.129	58%	24.6	0.810	+0.051	70%	7.1	0.942	+0.083	40%	10.1	0.835

Appendix

Table 4: Confidence profiles of every Table 2 row, computed from the published test predictions, in Table 2 order. B: nominal parameter scale in billions;

\Delta

: mean confidence minus accuracy (signed);

{>}0.95

: share of predictions with confidence above

0.95

; REL: the reliability (calibration) term of the binary Brier decomposition with 15 equal-width bins,

\times 10^{3}

; AUROC: area under the receiver operating characteristic curve for confidence against answer correctness; –: unsupported modality or undisclosed scale.

Figure 7: Per-bin reliability diagrams of the general-purpose models in Table 2 , in table order. Colored bars: empirical accuracy of each 15-bin confidence bin; gray bars: each bin’s share of the slice’s predictions; dashed diagonal: perfect calibration.

Figure 8: Per-bin reliability diagrams of the released decision models in Table 2 (first half, table order). Axes and markings follow Figure 7 ; modalities a model does not support are marked in place.

Figure 9: Per-bin reliability diagrams of the released decision models in Table 2 (second half, table order). Axes and markings follow Figure 7 .

Figure 10: Per-bin reliability diagrams of the local training baselines, supervised controls, and CORA (Table 2 order). The five RLCD baselines and CORA (2B) share the soft-label SFT (2B) initialization, enabling a controlled comparison of the continuation recipes; CORA (9B) starts from soft-label SFT (9B). Axes and markings follow Figure 7 .

Model Calibration

Model calibration aligns predicted confidence or model outputs with observed outcomes and measurements to support more reliable decisions.

See all 218 papers

Ranking-Aware Calibration for Reliable Multimodal Reinforcement Learning
May 16, 2026Peng Cui, Boyao Yang, Jun ZhuConfidence Estimation for VLMs Multimodal Robustness
Dept. of Comp. Sci. & Tech., Institute for AI, BNRist Center, Tsinghua-Bosch Joint ML Center, THBI Lab, Tsinghua University, Beijing 100084, China · Dept. of Automation, Tsinghua University, Beijing 100084, China
Jev thinks "I don't know'', but doesn't say it: Introducing Sys1Cal-v1 Dataset for Probability Calibration
Sep 28, 2026Riccardo PorceddaProbability Calibration Uncertainty Calibration
Department of Excellence L’EMbeDS, Sant’Anna School of Advanced Studies, Italy · Department of Computer Science, University of Pisa, Italy
MJ1: Multimodal Judgment via Grounded Verification
Mar 9, 2026Bhavesh Kumar, Dylan Feng, Leonard TangPosition Bias in VLMs MLLM Evaluation
Haize Labs

May 16, 2026J/K move · Enter open · S save

Ranking-Aware Calibration for Reliable Multimodal Reinforcement Learning

Peng Cui, Boyao Yang, Jun Zhu

Dept. of Comp. Sci. & Tech., Institute for AI, BNRist Center, Tsinghua-Bosch Joint ML Center, THBI Lab, Tsinghua University, Beijing 100084, China · Dept. of Automation, Tsinghua University, Beijing 100084, China

Confidence Estimation for VLMs Multimodal Robustness VLM Reasoning RL Post-Training VLM Calibration RL for VLMs

Open paper AskPDF arXiv

Reinforcement learning post-training has substantially improved the reasoning accuracy of vision-language models, yet the resulting policies remain poorly calibrated. Terminal correctness rewards provide no gradient that penalizes confident errors more than uncertain ones and no signal that ties confidence to the quality of visual evidence, a gap that becomes especially severe under corrupted or ambiguous inputs where models continue to report high confidence on incorrect answers. We introduce Ranking-Aware Calibration (RAC), a training-time framework that supervises confidence using two comparison signals that group-based RL already produces at no additional labeling cost. The ranking-aware group loss enforces that a better rollout receives higher confidence than a worse one within the same prompt. The clean--corrupted pairwise loss enforces that confidence attenuates as visual evidence degrades. Because the ranking signal forces the policy to distinguish between correct and incorrect reasoning paths, it also reinforces task accuracy beyond what correctness rewards alone produce. Both losses require no external confidence annotations and integrate naturally with group-based RL post-training. We instantiate RAC on Qwen2.5-VL and InternVL-3.5 backbones and evaluate on six multimodal reasoning benchmarks under clean and corrupted inputs. Empirical results show that the ranking-aware loss substantially improves task accuracy by teaching the policy to discriminate between better and worse reasoning, while the pairwise corruption loss reduces calibration error under degraded inputs. Their combination achieves the best calibration across all tested backbones while improving accuracy in the majority of settings.

Abstract

Model Calibration

Explore similar work

Ranking-Aware Calibration for Reliable Multimodal Reinforcement Learning

Jev thinks "I don't know'', but doesn't say it: Introducing Sys1Cal-v1 Dataset for Probability Calibration

MJ1: Multimodal Judgment via Grounded Verification