Fewer Tokens, More Answers: Verifiable-Reward Tuning of an Ornith-Style VLM for Multi-View Driving QA

T N Rickesh; Pranav Vinodh; Sowmya Kamath S

cs.CVOct 31, 2026

Fewer Tokens, More Answers: Verifiable-Reward Tuning of an Ornith-Style VLM for Multi-View Driving QA

Authors: T N Rickesh, Pranav Vinodh, Sowmya Kamath S

Organizations: natkal · National Institute of Technology Karnataka

Abstract

Reasoning-enabled model derivatives must turn intermediate deliberation into a usable answer within a deployment budget. We study that transition in an Ornith checkpoint adapted to multi-camera driving question answering with verifiable rewards. Paired archived outputs show that, on a selected 400-question development screen with a 256-token cap, GRPO increases answer extraction from 78.25% to 90.25% and accuracy from 55.50% to 63.75%. Of 39 wrong-to-correct transitions, 33 begin with an unextractable base response. The base's 87 unextractable responses all reach the token cap. Checkpoint-verified reconstruction on those same questions gives a much smaller first-pass accuracy gain at 512 tokens: +0.75 percentage points, versus +8.25 at 256. A later answer-cued continuation is analyzed separately; it helps the base more and narrows the adapter difference. On the full 2,511-question development set, this recovery-enabled pipeline reduces mean generated tokens from 108.6 to 47.1, while the paired accuracy interval includes zero gain. We frame these findings as adaptation of answer delivery under finite budgets, with explicit selection and provenance limits. Ornith's self-scaffolding lineage and newer reasoning-enabled derivatives such as Bonsai motivate the problem, but our measured results concern Ornith alone and do not establish a shared architectural mechanism or cross-model transfer.

Figures & tables

Model lineage	Documented distinction	Role in this paper
Ornith	Learned scaffolds and self-improvement	Recorded checkpoint is experimentally studied; release identity remains incomplete
Bonsai	Compressed multimodal backbone with thinking support	Related derivative; no WaymoQA evaluation or tuning result reported
Gemma	Multiple native multimodal designs; configurable thinking	Shared diagnostic reference, with tuning studied in the companion

Table 1: Primary-documentation context [ 8 , 9 , 10 , 5 ] . Properties can overlap; these are not mutually exclusive architecture classes. External release descriptions are not our driving measurements.

Checkpoint	Role	Correct / all	Acc. all (%)
Gemma-4-12B	Zero-shot reference	4,072 / 5,519	73.78 [E1]
Gemma-4-31B	Zero-shot reference	4,261 / 5,519	77.21 [E13]
Ornith + GRPO	Deployment observation	3,812 / 5,519	69.07 [E16]

Table 2: Shared-board reference context. [E1,E13,E16] Services and generation settings differ. Ornith is a tuned deployment observation without a corresponding base run on this board. These rows are not a paired Ornith effect or a controlled architecture ranking.

Figure 1: Correctness and answer availability on the selected 256-token screen. [E5] Correct fallback responses remain correct. Unextractable responses count against accuracy, but are not assigned a visual failure cause. This diagnostic differs from the companion’s almost fully extractable Gemma base.

Data / protocol	$n$	Base	GRPO	$\Delta$ (pp)	Scene CI (pp)
Screen / 256 first pass	400	55.50	63.75	+8.25	[+5.23, +11.14]
Screen / 512 first pass	400	68.50	69.25	+0.75	[-2.23, +3.80]
Screen / 512 + recovery	400	69.75	69.75	+0.00	[-2.77, +3.09]
Dev / 512 first pass	2,511	68.66	70.45	+1.79	[+0.45, +3.23]
Dev / 512 + recovery	2,511	70.53	71.68	+1.15	[-0.12, +2.42]

Table 3: Paired Ornith accuracy (%) by data and collection protocol. [E5–E6] Intervals are scene-cluster summaries conditional on the selected configuration. Screen questions are inside development; these rows are not independent evaluations.

Figure 2: Same-screen accuracy gains under distinct collection protocols, and recorded completion lengths. [E5–E6] The cost panel compares a single-pass screen protocol and the larger recovery-enabled development protocol; it does not estimate equal-workload latency.

Appendix figures & tables1 asset

Supplementary material from the paper’s appendix.

Appendix

Candidate learning rate	Correct	Gain (pp)	Source
$3\times 10^{-7}$	227 / 400	+1.25	E11
$10^{-6}$	226 / 400	+1.00	E12
$3\times 10^{-6}$	255 / 400	+8.25	E5

Appendix

Table 4: Archived 256-token screen candidates. [E5,E11–E12] Base accuracy is 222/400. The largest candidate was selected using this screen; its uncertainty is not a prospective confirmation or a general dose-response result.

VLM Reasoning

Vision-language model reasoning studies how multimodal systems use visual evidence and language to answer questions, plan, and explain.

See all 859 papers

Answer Correction with Verifiable Rewards: A Gemma-Family Study of Multi-View Driving QA
Oct 31, 2026T N Rickesh, Pranav Vinodh, Sowmya Kamath SVLM Evaluation VLMs for Autonomous Driving
natkal · National Institute of Technology Karnataka
VERS: Steering Reasoning Models by Editing Their Activation Explanations
Nov 2, 2026Ihor VitenkoEfficient Language Model Reasoning Activation Steering
CoT-X: An Adaptive Framework for Cross-Model Chain-of-Thought Transfer and Optimization
Nov 7, 2025Ziqian Bi, Yinzhi Wang, Tianyang Wang +6Efficient Language Model Reasoning LLM Inference Acceleration
1Purdue University, USA · 2Baruch College, City University of New York, USA · 3The Ohio State University, USA +6

Oct 31, 2026J/K move · Enter open · S save

Answer Correction with Verifiable Rewards: A Gemma-Family Study of Multi-View Driving QA

T N Rickesh, Pranav Vinodh, Sowmya Kamath S

natkal · National Institute of Technology Karnataka

VLM Evaluation VLMs for Autonomous Driving Reinforcement Learning with Verifiable Rewards

Open paper AskPDF arXiv

Multi-camera driving question answering provides a discrete test of vision-language models, but post-training gains can reflect changed selected answers or improved answer extraction. We investigate this distinction within the Gemma model family using archived, task-linked WaymoQA evaluations. A zero-shot diagnostic on 5,519 test questions spans effective-small, mixture-of-experts, unified, and dense checkpoints, with serving differences disclosed rather than interpreted as architecture effects. The strongest controlled evidence is a matched-harness Gemma-4-12B base/GRPO pair: accuracy rises from 73.78% to 75.43%, a gain of 1.65 percentage points with a scene-cluster bootstrap interval of [+0.84, +2.41]. There are 215 improvements and 124 regressions; 213 improvements start from an extractable but incorrect base answer, and only two recover previously unextractable responses. Crossing correctness with extraction state shows why correct noncompliant outputs cannot be counted as prediction errors. Category counts reveal a critical safety asymmetry: routine perception gains (+5.93 pp) mask a regression in safety-critical hazard prediction (-1.82 pp), demonstrating that unweighted outcome rewards can trade off rare edge cases for aggregate accuracy. Reward saturation describes sampled training dynamics rather than an architectural ceiling. A separate archived Gemma-31B thinking comparison yields no detected average improvement, with uncertainty that does not establish equivalence. These results support a modest answer-selection gain for the recorded tuning configuration and a reproducible diagnostic accounting, without attributing corrected labels to specific grounding mechanisms or claiming superiority over matched supervised tuning.

Abstract

VLM Reasoning

Explore similar work

Answer Correction with Verifiable Rewards: A Gemma-Family Study of Multi-View Driving QA

VERS: Steering Reasoning Models by Editing Their Activation Explanations

CoT-X: An Adaptive Framework for Cross-Model Chain-of-Thought Transfer and Optimization