Faithful-MR1: A Framework for Faithful Multimodal Reasoning

• Faithful-MR1 is a two-stage training framework that enforces faithful perception and utilization of visual evidence in multimodal reasoning tasks.
• It uses an Anchoring stage for region-level visual grounding and a Reinforcing stage with counterfactual masking to ensure attention aligns with causally-relevant image regions.
• Evaluated on math and VQA benchmarks, the framework improves accuracy and response efficiency while using substantially fewer training examples.

Faithful-MR1 is a two-stage training framework for multimodal LLMs that targets faithful multimodal reasoning: both faithful perception of task-relevant visual evidence in the image and faithful use of that evidence during the reasoning chain. It was introduced to address a perception–reasoning disconnect in multimodal reinforcement learning with verifiable rewards (RLVR), where models may correctly perceive relevant visual facts yet drop, distort, or override them with language priors during subsequent reasoning. The framework combines an Anchoring stage that supervises a dedicated <Focus> token directly on image regions and a Reinforcing stage that uses counterfactual image intervention to reward answer-correct trajectories whose visual attention concentrates where vision causally matters. On Qwen2.5-VL-Instruct 3B and 7B backbones, Faithful-MR1 improves performance on seven multimodal reasoning benchmarks while using 19.2K training examples, and it is positioned as a training-signal-centric alternative to text-only perception supervision and answer-only multimodal RLVR (Tian et al., 21 May 2026).

1. Conceptual basis and target failure mode

Faithful-MR1 defines faithfulness in multimodal reasoning as a paired requirement: the model must first perceive the relevant visual evidence and then use that evidence faithfully in the derivation. The motivating diagnosis is the perception–reasoning disconnect (PRD): correctly perceived evidence is often dropped or contradicted during reasoning, which helps explain why recent extensions of RLVR to multimodal LLMs have yielded smaller gains than comparable text-only systems (Tian et al., 21 May 2026).

This framing distinguishes Faithful-MR1 from multimodal RLVR variants such as Vision-R1, Vision-SR1, VPPO, and Perception-R1. Those methods are described as largely supervising perception through generated textual descriptions of images or through answer-level reinforcement signals. Faithful-MR1 instead treats region-level grounding and causal use as separate training targets. The first gap is that textual perception signals are indirect: they do not teach the model how to extract and localize task-relevant evidence natively in the image. The second gap is that faithful use is largely overlooked: even when perception succeeds, nothing in a standard outcome reward guarantees that the evidence remains load-bearing in the reasoning chain (Tian et al., 21 May 2026).

A central implication is that higher perceptual coverage alone is not sufficient. The paper’s PRD analysis on MathVerse Vision-Only explicitly reports that high Coverage does not necessarily translate into higher accuracy, and that Faithful-MR1 improves Faithful Use while producing shorter, more token-efficient responses. This suggests that the framework is not merely increasing visual mention frequency, but is attempting to convert perception into causal reliance during reasoning (Tian et al., 21 May 2026).

2. Anchoring: explicit pre-reasoning perception supervision

The Anchoring stage turns perception into an explicit pre-reasoning subtask. The prompt prepends the instruction: “Perceive the visual cues relevant to the above question and capture them into <Focus>.” The <Focus> token is not trained to emit text. Its function is to collect and localize evidence by attending to the correct visual patches, and supervision is placed on its attention pattern rather than on textual output (Tian et al., 21 May 2026).

For each training example, question-relevant evidence is annotated as bounding boxes,

$\mathcal{B} = \{b_1,\dots,b_M\}.$

Each box $b_m$ maps to a set of visual patch tokens $\mathcal{P}(b_m)$, and the target region is

$$\mathcal{P}_{\mathrm{box}} = \bigcup_{m=1}^{M} \mathcal{P}(b_m).$$

Let $a_{\mathrm{Focus}, i}^{(\ell,h)}$ denote the <Focus> token’s attention to token $i$ at layer $\ell$ and head $h$. The attention mass on the target region is

$$s^{(\ell,h)} = \sum_{i \in \mathcal{P}_{\mathrm{box}}} a_{\mathrm{Focus}, i}^{(\ell,h)}.$$

Faithful-MR1 supervises this quantity toward a target ratio $\tau$ with a Bernoulli cross-entropy,

$b_m$0

and averages over a selected set $b_m$1 of layer–head pairs:

$b_m$2

This attention supervision is combined with the standard autoregressive answer loss,

$b_m$3

to form the Anchoring-stage objective

$b_m$4

In the reported experiments, Anchoring uses 6K SFT examples, one epoch, learning rate $b_m$5, auxiliary loss weight $b_m$6, target attention mass $b_m$7, and supervision on the last 4 attention layers (Tian et al., 21 May 2026).

The significance of this stage lies in its native grounding. Rather than reward textual descriptions of what the image contains, it aligns the model’s internal visual selection with task-relevant patches. The paper argues that this reduces hallucination and decorational descriptions that do not support the derivation, while making the pre-reasoning focus step an internal mechanism that conditions the subsequent chain (Tian et al., 21 May 2026).

3. Reinforcing: counterfactual faithful use of visual evidence

The Reinforcing stage targets the second half of faithfulness: whether the model actually uses visual evidence where vision causally matters. The method constructs a perturbed image by masking the annotated bounding-box regions while leaving the rest unchanged. For each sampled trajectory, the model is run on both the original and masked images, yielding next-token distributions $b_m$8 and $b_m$9. The per-token visual sensitivity score is

$\mathcal{P}(b_m)$0

Tokens with high $\mathcal{P}(b_m)$1 are treated as vision-dependent, since their predictions shift when relevant evidence is removed (Tian et al., 21 May 2026).

Faithful-MR1 then computes average attention from vision-dependent tokens to visual tokens, $\mathcal{P}(b_m)$2, and from non-vision-dependent response tokens, $\mathcal{P}(b_m)$3, and forms the attention-concentration ratio

$\mathcal{P}(b_m)$4

This produces a binary auxiliary reward

$\mathcal{P}(b_m)$5

with default threshold $\mathcal{P}(b_m)$6. The attention reward is explicitly gated on answer correctness: only trajectories with correct final answers can receive attention credit. This constraint is intended to ensure that rewards reflect faithful use in correct derivations rather than arbitrary attention concentration (Tian et al., 21 May 2026).

The RL update is integrated through split GRPO. Answer correctness reward $\mathcal{P}(b_m)$7 and attention reward $\mathcal{P}(b_m)$8 are normalized separately,

$\mathcal{P}(b_m)$9

and combined as

$$\mathcal{P}_{\mathrm{box}} = \bigcup_{m=1}^{M} \mathcal{P}(b_m).$$0

The policy objective is a clipped GRPO-style surrogate with KL regularization to a reference policy:

$$\mathcal{P}_{\mathrm{box}} = \bigcup_{m=1}^{M} \mathcal{P}(b_m).$$1

where

$$\mathcal{P}_{\mathrm{box}} = \bigcup_{m=1}^{M} \mathcal{P}(b_m).$$2

The reported Reinforcing configuration initializes from the Anchoring checkpoint, uses 13.2K RL examples, 8 rollouts per prompt, rollout batch size 128, learning rate $$\mathcal{P}_{\mathrm{box}} = \bigcup_{m=1}^{M} \mathcal{P}(b_m).$$3, $$\mathcal{P}_{\mathrm{box}} = \bigcup_{m=1}^{M} \mathcal{P}(b_m).$$4 for 7B and $$\mathcal{P}_{\mathrm{box}} = \bigcup_{m=1}^{M} \mathcal{P}(b_m).$$5 for 3B, bf16 precision, rollout temperature 1.0, and EasyR1/HybridFlow. Training was run on 32 AMD MI308X GPUs under ROCm/HIP (Tian et al., 21 May 2026).

4. Data, benchmarks, and evaluation protocol

Faithful-MR1 was trained on 19.2K examples constructed from Vision-SR1-47K with bounding-box annotations of question-relevant regions, automated with Gemini-3-Flash. The corpus is split into 6K SFT examples for Anchoring and 13.2K RL examples for Reinforcing. The benchmark suite contains seven tasks: MATH-Vision, MathVerse, MathVista, WeMath, DynaMath, MMMU-Pro, and HallusionBench (Tian et al., 21 May 2026).

These tasks emphasize visual reasoning rather than generic captioning. Problems require reading labels, geometric configurations, chart or table values, or spatial relations and then using them in multi-step derivations. DynaMath uses Worst Case Accuracy (WCA), where a question counts as correct only if all 10 visual variants are answered correctly. MathVerse Vision-Only removes figure-related text, forcing reliance on the image (Tian et al., 21 May 2026).

Answer correctness is verified with a rule-based answer reward and a format reward, including answer enclosure in $$\mathcal{P}_{\mathrm{box}} = \bigcup_{m=1}^{M} \mathcal{P}(b_m).$$6 and correct option-letter formatting for multiple-choice tasks. For evaluation, the framework uses Qwen3-235B-A22B-Instruct as an LLM-as-a-judge within EvalScope, and all evaluations use greedy decoding (Tian et al., 21 May 2026).

A concise summary of headline overall results is:

Backbone	Base Overall	Faithful-MR1 Overall
Qwen2.5-VL-3B-Instruct	38.8	43.9
Qwen2.5-VL-7B-Instruct	45.5	51.3

For the 3B model, +GRPO reaches 39.6 and +VPPO 40.8 before Faithful-MR1 reaches 43.9. For the 7B model, +GRPO reaches 47.9 and +VPPO 49.5 before Faithful-MR1 reaches 51.3. On DynaMath WCA, the 3B model rises from 13.6 to 18.6, and the 7B model rises from 20.4 to 26.8. Math Avg likewise increases from 35.8 to 41.6 for 3B and from 43.9 to 49.9 for 7B (Tian et al., 21 May 2026).

5. Empirical findings and ablation evidence

The paper reports that Faithful-MR1 outperforms recent multimodal reasoning baselines on both Qwen2.5-VL-Instruct 3B and 7B backbones while using substantially less training data. On public 7B checkpoints trained with much more data, Faithful-MR1 surpasses Vision-R1-7B, Perception-R1-7B, and Vision-SR1-7B on Overall and most benchmarks, with particularly strong gains on math reasoning, where faithful use of visual evidence matters most (Tian et al., 21 May 2026).

The ablation study attributes the gain across two stages. On the 7B backbone, the sequence is: Base 45.5, +Vanilla GRPO 47.9, +Vanilla SFT on Anchoring data + GRPO 48.6, +Anchoring (<Focus> + attention loss) + GRPO 49.2, and full Faithful-MR1 51.3. The largest single delta comes from Reinforcing: +2.1 over Anchoring+GRPO and +3.4 over Vanilla GRPO. The paper interprets this as evidence that final accuracy is more directly influenced by faithful use than by perception alone, although Anchoring also contributes beyond pure data effects through direct region-grounded supervision (Tian et al., 21 May 2026).

The robustness sweeps on DynaMath show an inverted-U relationship for visual-attention guidance. Anchoring weight $$\mathcal{P}_{\mathrm{box}} = \bigcup_{m=1}^{M} \mathcal{P}(b_m).$$7 yields peak robustness 35.9, which is +3.3 over $$\mathcal{P}_{\mathrm{box}} = \bigcup_{m=1}^{M} \mathcal{P}(b_m).$$8. Reinforcing weight $$\mathcal{P}_{\mathrm{box}} = \bigcup_{m=1}^{M} \mathcal{P}(b_m).$$9 yields peak robustness 37.7, which is +1.6 over $a_{\mathrm{Focus}, i}^{(\ell,h)}$0. Excessive weights hurt robustness, which the paper describes as consistent with both stages acting on shared attention mechanisms (Tian et al., 21 May 2026).

The PRD analysis on 788 MathVerse Vision-Only responses is especially diagnostic. Vision-R1-7B reports Coverage 70.2, Faithful Use 54.3, Length 526, and Accuracy 41.4. Vision-SR1-7B reports Coverage 74.6, Faithful Use 49.4, Length 597, and Accuracy 41.8. The Qwen2.5-VL-7B base model reports Coverage 65.2, Faithful Use 63.3, Length 401, and Accuracy 26.5. +VPPO reports Coverage 65.1, Faithful Use 62.9, Length 382, and Accuracy 41.9. Faithful-MR1 reports Coverage 61.3, Faithful Use 66.7, Length 325, and Accuracy 45.1. The paper’s stated conclusion is that high Coverage alone does not translate to higher accuracy, whereas Faithful-MR1 achieves the highest Faithful Use and the best accuracy under matched backbone and data, with shorter, more token-efficient responses (Tian et al., 21 May 2026).

6. Interpretation, limitations, and research context

Faithful-MR1 is best understood as a training framework rather than a standalone metric. Its faithfulness signal is interventional and behavior-level: region-grounded attention in Anchoring, and counterfactual causal localization through $a_{\mathrm{Focus}, i}^{(\ell,h)}$1 in Reinforcing. This places it in a broader movement away from outcome-only or surface-faithfulness criteria. FaithRL similarly argues that outcome-only RLVR provides little supervision over intermediate steps and encourages spurious reasoning and hallucinations, proposing step-level faithfulness maximization instead (Gui et al., 3 Feb 2026). Coverage-aware grounded-generation work likewise shows that precision-only faithfulness metrics reward abstention and can misorder systems when recall of relevant facts is ignored (Santillana, 8 Jun 2026). Meta-evaluation work on chain-of-thought metrics goes further, reporting that most proposed faithfulness metrics perform near chance on a benchmark with ground-truth labels (Gur-Arieh et al., 24 May 2026). In multimodal affective reasoning, FACR adopts a related interventional strategy by enforcing counterfactual consistency between invoked action units, predictions, and a structural causal graph (Huynh et al., 14 Jun 2026).

Several limitations are explicit. Anchoring depends on bounding-box annotations of question-relevant regions, and the 19.2K corpus was automatically constructed with Gemini-3-Flash, so annotation quality and coverage can vary by domain. Reinforcing requires an extra masked-image forward pass per rollout, creating training-time overhead, although the paper states that inference incurs no extra cost and that responses are shorter on average. Counterfactual masking assumes the annotated regions capture all causally relevant evidence, so partial or noisy annotations may mislocalize $a_{\mathrm{Focus}, i}^{(\ell,h)}$2. The paper is also explicit that faithful attention concentration is a strong behavioral signature but does not prove deep causal reasoning in all cases. Residual failures remain where perception is correct and attention concentrates yet reasoning fails because of algebraic mistakes or flawed geometry. Attention concentration therefore indicates necessary reliance, not sufficiency for correctness (Tian et al., 21 May 2026).

These limitations define the framework’s scope. Faithful-MR1 does not require explicit visual operations in the output; instead, it modifies the training signals that shape internal visual selection and its causal deployment. A plausible implication is that its main contribution is methodological: it operationalizes faithful multimodal reasoning as a coupling of region-level perception supervision and counterfactual evidence-utilization rewards, and then shows that this coupling can improve both accuracy and data efficiency on math reasoning and general VQA tasks (Tian et al., 21 May 2026).