HART: High-Resolution Annotation-Free Reasoning Technique through a Closed-loop Framework

Conference: ECCV 2026
Paper: ECCV Original
Code: https://github.com/RL-MIND/HART
Area: Multimodal VLM / Visual Reasoning
Keywords: High-Resolution Visual Reasoning, Visual Grounding, Annotation-Free RL, Closed-Loop Self-Verification, AP-GRPO

TL;DR

HART first predicts key regions on a downsampled full image, then deliberately withholds the full image and lets the model answer using only those regions cropped from the original high-resolution image, so that "is the answer correct" becomes a self-verification signal for "is the localization accurate"; an advantage-weighted AP-GRPO then optimizes localization directly without any bounding-box annotations, lifting Qwen2.5-VL-7B from 42.3% to 62.4% on MME-RealWorld-Lite.

Background & Motivation

High-resolution inputs remain the weak spot of large multimodal models (LMMs). The number of visual tokens grows roughly quadratically with image resolution, while only a small subset is actually relevant, so mainstream architectures such as Qwen2.5-VL and InternVL3 simply cap the input at a maximum pixel budget. The cost is that critical details get squeezed out, and performance visibly degrades on remote sensing, autonomous driving, and document OCR — scenes where the answer hides in a small local patch. A widely adopted remedy mimics the fovea of the human eye through visual grounded reasoning: predict the key region of interest (ROI) from a downsampled image conditioned on the textual question, crop that region from the original high-resolution image, and answer using both the downsampled view and the crop, spending compute only on the pixels that carry the answer.

The difficulty lies in how to train such a model. One line of work supervises localization directly with ground-truth boxes (the Visual CoT family, VLM-R1, Point-RFT); localization is genuinely optimized, but box annotations are expensive to collect and hard to scale. Another line uses reinforcement learning: MGPO, DeepEyes, and Pixel-Reasoner reward only the correctness of the final answer and need no extra annotations at all. Yet that reward says nothing about localization quality. A model that answers correctly from a wrong region still receives a positive reward, which pushes the policy toward "box something arbitrarily and guess from language priors." The authors call this reward misspecification and quantify it in a pilot experiment on Visual CoT: after filtering out descriptive questions and yes/no questions, 4,042 questions remain (one ground-truth box each; a prediction counts as correct if at least one predicted box covers more than 0.3 of the ground-truth area). Qwen2.5-VL-7B answers 2,895 of them correctly, but 1,057 of those boxes are wrong — 36.5%; InternVL3-8B is worse, with 1,359 of 2,129 correct answers localized incorrectly — 63.8%. A large share of the positive rewards handed out by these "annotation-free" methods is therefore rewarding wrong localization, and none of them has any mechanism to notice.

The paper's way out is this: if localization error cannot be computed without boxes, then encode localization quality into the answer itself. If the cropped region really suffices, the model should still answer correctly after the full image is taken away; if the box is off, the answer should collapse without global context. "Withhold the full image and answer from the ROIs alone" thus becomes a self-verification step performed by the model itself, and the supervision signal requires no human annotation at all. Core idea: build a closed loop in which the full image is withheld and the model answers from the ROIs alone, turning final-answer correctness into a proxy supervision signal for localization quality, and use an advantage-preference weighted AP-GRPO to suppress the misleading gradients from "right answer, wrong box" samples — thereby optimizing visual grounding directly with zero box annotations.

Method

Overall Architecture

HART takes a high-resolution image \(I_f\) and a textual question \(q\) as input and produces an interpretable reasoning path of "bounding boxes + answer." Post-training has two stages. Stage 1 is closed-loop reinforcement learning: the model first outputs bounding boxes for the key regions relevant to the question on the downsampled full image; the system maps those boxes back to the original high-resolution image to crop sub-images \(I_s\) and deliberately withholds the full image, handing the model only the sub-images plus the original question; answer correctness is the reward, weighted per-sample by AP-GRPO and fed back to update the localization policy directly. Stage 2 is SFT: the full image is restored, and the model is supervised on the complete high-resolution input to recover the answering ability lost by withholding the full image.

Two points are the key to reading the whole paper. First, the feedback signal comes neither from external box annotations nor from any detection or segmentation model — it is the model's own answer under the condition of "no full image, only sub-images": accurate localization makes the crops self-sufficient and the answer correct; a shifted box leaves the crops short of information and the answer wrong. Second, high-resolution information enters the pipeline at the step where the ROIs are cropped from the original image: the model only ever consumes the downsampled image and never touches the maximum pixel budget, while genuine high-frequency detail is fed in locally and on demand through the crop window — this both bypasses the resolution cap and eliminates the redundant computation of encoding the whole image at high resolution.

%%{init: {'flowchart': {'rankSpacing': 24, 'nodeSpacing': 28, 'padding': 6, 'wrappingWidth': 400}}}%%
flowchart TD
    A["High-res image + question"] --> B["Downsampled full image<br/>predict key region ROI"]
    B --> C["Closed-loop self-verification<br/>withhold full image, answer from ROIs"]
    C -->|correct answer implies reliable localization| D["AP-GRPO<br/>advantage-weighted policy update"]
    D -->|re-localize after policy update| B
    D --> E["RL → SFT two stages<br/>recover high-res reasoning with full image"]
    E --> F["Output: bounding boxes + answer"]

Key Designs

1. Closed-loop self-verification: turning "is the answer right" into a supervision signal for "is the box right"

This is the hinge of the paper, and it targets exactly the pain point above: an answer-only reward cannot distinguish "right box, right answer" from "wrong box, still right answer." HART splits an ordinary multimodal question-answering pass into two steps. In step one, the model outputs the coordinates of the key regions given the downsampled full image and the question \([I_f, q]\), with the instruction "output the coordinates of the key image area relevant to the problem." In step two, the system crops those regions from the original high-resolution image, removes the full image, and rewrites the instruction as "you were supposed to answer a question based on a full image, but now the full image is not visible; you can only see some sub-regions relevant to the problem — answer the following question." The model must now answer from \(I_s\) and \(q\) alone. Correctness therefore implicitly entails that the region contains the information needed to answer — the reward function itself is unchanged (still a binary judgment of answer correctness), but the meaning of the reward has changed.

The authors formalize this shift in information-theoretic terms: with \(L\in\{0,1\}\) denoting whether localization is correct and \(R\in\{0,1\}\) whether the response is correct, the mutual information under HART is strictly larger than that of the baseline, i.e. \(I_{\text{HART}}(L;R) > I_{\text{baseline}}(L;R)\) (proof in Appendix A of the paper; ⚠️ refer to the original paper). The intuition is that the loop lowers the probability of the "correct answer with wrong localization" cell, shrinking the entropy of \(R\) and strengthening the causal dependency between \(L\) and \(R\). Measurements support this: under HART's answering protocol, the proportion of "correct answer but wrong localization" drops from 36.5% to 21.5% for Qwen2.5-VL-7B (a 15.0-point drop) and from 63.8% to 55.9% for InternVL3-8B (a 7.9-point drop).

The loop also removes two practical obstacles of high resolution: (a) localization quality can be self-verified, so no manual boxes are needed; (b) zooming into the key region exposes finer visual detail, bypassing the maximum pixel constraint of existing LMMs while avoiding the redundant computation of encoding the full image at high resolution.

2. AP-GRPO: advantage-based dynamic weighting to suppress the misleading "right answer, wrong box" gradients

The closed loop fixes the semantics of the reward but does not remove the other source of reward misspecification: under HART's protocol, a wrong response \(r_i=0\) may mean the localization was wrong, or it may mean localization was right but the reasoning fell short, so such samples carry markedly higher localization uncertainty. Vanilla GRPO treats every sample in the group alike (the advantage \(A_i\) is only group-normalized), so these high-uncertainty samples receive updates on the same scale as samples with trustworthy grounding. AP-GRPO (Advantage Preference Group Relative Policy Optimization) assigns each sample a dynamic weight derived from its own advantage: since a correct answer inside the loop is more likely to indicate a correct box, correct samples should contribute more and be constrained less.

Concretely, the group of candidate responses \(\{o_i\}\) receives binary rewards \(r_i\in\{0,1\}\) with group mean \(\text{mean}(\{r_i\})\), and the objective is

\[\mathcal{J}_{\text{AP-GRPO}}(\theta)=\frac{1}{G}\sum_{i=1}^{G}\Big(\mu_1^{(i)}\cdot\frac{\pi_\theta(o_i\mid q)}{\pi_{\theta_{\text{old}}}(o_i\mid q)}A_i-\beta\,\mu_2^{(i)}\cdot\mathbb{D}_{\text{KL}}\big[\pi_\theta\,\|\,\pi_{\text{ref}}\big]\Big)\]

where the two weight factors depend only on where the sample sits relative to the group mean: \(\mu_1=1+k\,(r_i-\text{mean}(\{r_i\}))\) and \(\mu_2=\big(1-k\,(r_i-\text{mean}(\{r_i\}))\big)^2\), with \(k\) the only hyper-parameter in HART (\(k=0.6\) in the main experiments). For a correct sample \(\mu_1>1\) amplifies the advantage term, while \(\mu_2<1\) weakens the KL penalty and allows a larger deviation from the reference model — a trajectory that answers correctly from the sub-images alone is itself evidence of trustworthy localization and should not be reined in by the reference model. ⚠️ Equation 6 in the extracted text is partially garbled; the squared form of \(\mu_2\) and the overall expression are reconstructed from context — refer to the original paper.

The authors further give a theoretical account via Proposition 2: with \(g_{\text{AP-GRPO}}\) and \(g_{\text{baseline}}\) the gradients of AP-GRPO and the baseline, there exists \(\alpha\in[0,1]\) such that \(g_{\text{AP-GRPO}} = g_{\text{baseline}} - \alpha\,P(L{=}0,r{=}1)\,\mathbb{E}_{L=0,r=1}[\nabla\log\cdot]\) in a single-step RL setting with binary rewards. In words, AP-GRPO explicitly subtracts the gradient contribution of the samples that are "wrong localization yet positively rewarded" — precisely the source of reward misspecification identified earlier. Compared with equal-weight vanilla GRPO it brings three benefits: it sharpens attention to the correct ROIs, directly optimizes localization without any visual supervision, and keeps the reasoning path interpretable.

3. RL → SFT in two stages: learn to look at the right place first, then recover high-resolution reasoning

Withholding the full image genuinely removes information, so answering accuracy necessarily suffers: in the ablation, AP-GRPO alone (\(k=0.6\)) reaches only 53.3/48.0 (perception/reasoning) on MME-RealWorld-Lite, while SFT alone gives 54.0% overall. HART therefore appends an SFT stage in which the original image is fully visible, restoring high-resolution reasoning through standard autoregressive supervision; only the two stages chained together reach 62.4% overall. The order is a hard constraint: "SFT then AP-GRPO" at \(k=0.6\) scores only 51.7/48.1, far below the 64.9/58.5 of "AP-GRPO then SFT." A natural explanation is that the RL stage first forces out the ability to answer from local regions, so the model learns to extract information from sub-images; doing SFT afterwards on the full image leaves it with both "pick the right region" and "reason over everything." Reversing the order freezes the model into its existing answering pattern, and the subsequent localization signal can hardly move it. Data-wise, the paper follows the clean separation of [4]: 10K examples are randomly drawn from the MME-RealWorld training set for \(\mathcal{D}_{\text{RL}}\) and the remainder goes to \(\mathcal{D}_{\text{SFT}}\), with no overlap between stages.

A Worked Example

Take the TreeBench example in Fig. 4 of the paper, where the task requires attending to the right-most man before reasoning. Suppose the image is 2152×1615:

1. Localization: the model sees only the downsampled full image and outputs one ROI box. If it places the box on the crowd in the middle, the crop taken from the original image does not contain the man on the far right.
2. Self-verification: the full image is withheld and the model receives only that sub-image plus the original question. The key person is absent, so the answer is almost certainly wrong and the response gets \(r_i=0\).
3. Weighted update: a group of \(G\) responses is sampled for the same question. The ones that boxed the right man and answered correctly from the sub-image alone get \(r_i=1\) and, through AP-GRPO's \(\mu_1/\mu_2\), a larger policy update; the wrong responses are down-weighted.
4. Iteration: over rounds, the model learns to move the box onto the region that actually carries the answer. No box annotation is involved anywhere — the signal comes entirely from the model's own answer under the "sub-images only" condition.

By contrast, a baseline that only rewards the final answer (with the full image always visible) may still answer this question correctly from context even with the box on the middle crowd, so the sample receives a positive reward and reinforces a wrong localization — which is exactly the situation the paper measures in 36.5% of Qwen2.5-VL-7B's positive rewards. ⚠️ The exact box positions above are illustrative of the paper's description; Fig. 4 does not give precise coordinates.

Loss & Training

The main objective is the AP-GRPO target above (Eq. 6): the advantage term is scaled by \(\mu_1\) and the KL penalty by \(\mu_2\), both determined by where a sample's reward sits relative to the group mean, with \(k\) as the only hyper-parameter (\(k=0.6\) in the main experiments). Stage 2 uses the standard autoregressive negative log-likelihood,

\[\mathcal{L}(\theta)=-\,\mathbb{E}_{(x,y)\sim\mathcal{D}_{\text{SFT}}}\sum_{t=1}^{T}\log P\big(y_t\mid x,y_{<t};\theta\big)\]

where \(T\) is the text length and \((x,y)\) is the query and target response in \(\mathcal{D}_{\text{SFT}}\). Training uses the TRL framework for distributed reinforcement fine-tuning with Qwen2.5-VL-7B as the base model; every compared reinforcement fine-tuning method uses the same binary reward (answer correctness) and shares the same MME-RealWorld training set for fairness. Training details and cost are in Appendix C of the paper (⚠️ the main text reports no compute figures or inference latency).

Key Experimental Results

Main Results

Table 1: Answer accuracy (%) on the in-distribution MME-RealWorld-Lite, with sub-task sample counts in parentheses.

Method	Params	Perception-Autonomous Driving (350)	Perception-Remote Sensing (150)	Reasoning-Monitoring (150)	Overall
Qwen2.5-VL-7B	7B	30.0	32.7	28.7	42.3
LLaVA-OneVision-7B	7B	39.4	40.0	38.0	43.7
InternVL3-8B	8B	36.9	49.3	40.0	47.9
GRPO + SFT	7B	49.7	55.3	60.0	58.1
MGPO + SFT	7B	55.4	59.3	63.3	60.5
Pixel-Reasoner-7B	7B	30.9	52.0	46.0	49.7
DeepEyes-7B	7B	33.4	52.7	44.0	53.2
HART-7B	7B	57.7	58.7	58.7	62.4
Qwen2.5-VL-32B	32B	40.7	40.7	27.3	45.6
Qwen2.5-VL-72B	72B	30.6	34.0	26.7	43.7
GPT-4o	—	22.4	28.9	36.5	45.2
Claude 3.5 Sonnet	—	40.8	25.7	41.8	51.6

Table 2: Accuracy (%) on the out-of-distribution TreeBench and across-resolution benchmarks

Benchmark	Capability	Qwen2.5-VL-7B	InternVL3-8B	LLaVA-OV-7B	HART-7B
TreeBench (overall)	High-resolution reasoning	37.0	38.8	37.3	43.7
BLINK	Multi-image VQA	55.2	55.5	48.2	56.8
Mantis-Eval	Multi-image VQA	70.8	70.1	64.2	72.8
MathVista	Low-resolution VQA	68.2	71.6	58.3	71.8
MMStar	Low-resolution VQA	59.3	53.5	56.7	62.8
V* Bench	High-resolution VQA	78.5	72.3	70.7	80.6
HR-Bench-4K	High-resolution VQA	70.1	70.8	64.3	71.1
HR-Bench-8K	High-resolution VQA	61.0	62.0	59.8	71.9

Table 3: Grounding accuracy (%) at a coverage threshold of 0.3

Method	TreeBench correct grounding ↑	Visual CoT correct grounding ↑
Qwen2.5-VL-7B	50.2	66.0
LLaVA-OneVision-7B	38.3	67.0
InternVL3-8B	15.1	28.9
GRPO	51.0	66.1
MGPO	51.3	66.3
AP-GRPO (Ours)	75.4	77.7

Ablation Study

Table 4: Component ablation and sensitivity to \(k\) (%)

Config	\(k\)	MME-RW-Lite Perception / Reasoning	TreeBench Perception / Reasoning
Qwen2.5-VL-7B (baseline)	—	46.4 / 35.9	43.6 / 33.2
GRPO	—	52.1 / 49.1	52.3 / 32.0
MGPO	—	58.0 / 50.5	53.7 / 31.2
AP-GRPO	0.15	51.2 / 47.5	51.0 / 32.4
AP-GRPO + SFT	0.15	67.0 / 56.9	56.4 / 31.6
AP-GRPO	0.30	53.6 / 49.9	55.0 / 33.6
AP-GRPO + SFT	0.30	64.1 / 56.0	57.7 / 34.8
AP-GRPO	0.60	53.3 / 48.0	52.3 / 33.6
SFT + AP-GRPO (order swapped)	0.60	51.7 / 48.1	55.0 / 32.4
AP-GRPO + SFT	0.60	64.9 / 58.5	57.0 / 35.9

Key Findings

• The closed loop pays off most on perception sub-tasks where information hides in a local patch: against the Qwen2.5-VL-7B base, Perception-Autonomous Driving gains +27.7, Perception-Remote Sensing +26.0, and Reasoning-Monitoring +30.0 points, while OCR, which relies on global semantics anyway, barely moves (87.6 → 89.6). The same pattern holds across resolutions: V* Bench +2.1 and HR-Bench-4K +1.0 (already strong), but HR-Bench-8K goes from 61.0 to 71.9 (+10.9) — the higher the resolution, the more the loop's on-demand zooming is worth.
• RL and SFT are both necessary, and the order cannot be reversed: SFT alone gives 54.0% overall and AP-GRPO alone (\(k=0.6\)) gives 53.3 perception, yet chaining them reaches 62.4%; swapping the order to SFT + AP-GRPO yields only 51.7/48.1. The RL stage forces out "answer correctly from local regions only," and the SFT stage then reattaches full-image reasoning.
• AP-GRPO's value shows up mainly in localization itself, not in single-stage answer accuracy: in Table 3 its TreeBench correct-grounding rate is 75.4%, 24.1 points above MGPO's 51.3% (and 77.7% vs 66.3% on Visual CoT), yet in Table 4 AP-GRPO alone (53.3/48.0) is worse than MGPO alone (58.0/50.5). In other words, the localization gain only converts into answer accuracy once SFT restores high-resolution reasoning — which also explains why answer-only methods cannot optimize localization directly: they lack the step that turns grounding quality into a usable signal.
• Reasoning on OOD data remains the hardest case: on TreeBench's reasoning dimension, GRPO (32.0) and MGPO (31.2) actually fall below the no-post-training baseline (33.2); AP-GRPO's 33.6 is only a small gain, and it is SFT that pulls reasoning up to 35.9. Combined with the \(k\) sensitivity, \(k=0.6\) is best on TreeBench while \(k=0.15\) is best on MME-RW-Lite perception (67.0), so a single \(k\) does not transfer to every distribution — though the gap stays within an acceptable range.
• Transferability: applying HART to InternVL3-8B lifts MME-RealWorld-Lite perception from 46.4 to 61.8 and TreeBench perception from 46.3 to 59.1, showing the loop is not tied to one base model.

Highlights & Insights

• Turning evaluation into a training signal: with no box annotations, localization error cannot be computed, so HART lets the answer itself carry the localization information — still answering correctly after the full image is removed means the box was right. This is an almost cost-free way to construct a supervision signal, and it unties the knot that "annotation-free methods cannot optimize localization." It is the paper's clearest "aha" moment.
• The reward function is untouched; only its meaning changes: the reward is still binary answer correctness, but by changing the observation condition at inference (whether the full image is given), the same reward shifts from "measuring the answer" to "measuring localization." This "change the context, not the loss" idea transfers to any task where the final metric is observable but intermediate quality is not — for example, supervising retrieval quality in retrieval-augmented generation by asking whether the model can still answer once the source documents are removed.
• AP-GRPO's two weight factors are frugal by design: \(\mu_1\) amplifies the advantage term and \(\mu_2\) relaxes the KL constraint, both depending only on the gap between a sample's reward and the group mean, with no extra network or annotation and exactly one hyper-parameter \(k\). This "tilt the equal-weight assumption of GRPO a little" recipe can be dropped into almost any RL fine-tuning pipeline that uses group normalization.
• The two-stage order is established empirically: RL first, SFT last, with strictly non-overlapping data. Swapping the order costs more than 13 points, a conclusion of direct value to later work on mixed RL + SFT post-training.

Limitations & Future Work

• Gains depend heavily on the training distribution: both \(\mathcal{D}_{\text{RL}}\) and \(\mathcal{D}_{\text{SFT}}\) come from the MME-RealWorld training set, and the in-distribution gain is large (+20.1) while out-of-distribution TreeBench gains only +6.7; on high-resolution benchmarks private models still lead (o3 54.8 and Gemini-2.5-Pro 54.1 versus HART-7B 43.7). The main text reports 7B-scale results only; 32B and Qwen3-VL-8B results sit in the appendix.
• The self-verification signal is a noisy proxy: the pilot experiment deliberately excludes yes/no questions, which shows that when an answer can be guessed by a coin flip, the "correct answer ⇔ correct box" correspondence weakens markedly. For open-ended generation or tasks where language priors can fake a correct answer, the reliability of this loop needs re-evaluation.
• Grounding is measured with a custom metric: a prediction counts as correct if at least one box covers more than 0.3 of the ground-truth area — intersection-over-ground-truth rather than standard IoU, with a fairly loose threshold. The appendix reports stricter thresholds and an IoU curve, but the main-text conclusions rest on the loose one, so these grounding numbers are not fully comparable with other work (⚠️ refer to the original paper for the exact definition).
• Multi-turn inference adds overhead: each question costs at least "localize → crop → answer from sub-images," yet neither latency nor token cost is reported in the cached text, so the cost-effectiveness of deployment is unassessed; the cap on the number of ROI boxes and the selection strategy are also unspecified in the main text.
• Possible improvements: extend the binary answer reward into a process reward (e.g. a continuous score for sub-image sufficiency) to damp the noise from lucky guesses; make \(k\) adaptive to the distribution instead of fixing it at 0.6; move the loop from single images to video and multi-image high-resolution settings, where "withhold the context" can take richer forms.

Related Work & Insights

• vs MGPO / DeepEyes / Pixel-Reasoner (annotation-free grounded RL): all three perform reinforcement fine-tuning for visual grounded reasoning with a reward that only sees the final answer; MGPO relies on a multi-turn grounding pipeline, DeepEyes treats cropping as a tool, and Pixel-Reasoner drives exploration with curiosity. Their shared weakness is reward misspecification — in Table 3, MGPO's TreeBench correct-grounding rate is only 51.3%, essentially flat against the base model. By answering with the full image withheld, HART folds localization quality into the answer and reaches 75.4% correct grounding without any extra tool-calling policy.
• vs methods supervised with ground-truth boxes (Visual CoT / VLM-R1 / Point-RFT): these update the policy from the discrepancy against ground-truth boxes, so localization optimization is explicit and clean, but they depend on large-scale, costly annotation. HART uses no box annotation at all; the price is an indirect proxy signal, which is why its gains outside high-resolution perception are modest (MMStar +3.5, MathVista +3.6, a normal increment).
• vs vanilla GRPO: GRPO weights the group equally and, inside the loop, treats "unreliable localization" samples no differently; AP-GRPO changes only the weight factors (amplify with \(\mu_1\), relax with \(\mu_2\)) and gains +24.4 points of TreeBench grounding over GRPO's 51.0% → 75.4%, showing that in this setting "who should be learned from more" matters more than a stronger optimizer.

Rating

• Novelty: ⭐⭐⭐⭐☆ Treating "withhold the full image, answer from the ROIs alone" as an annotation-free self-verification signal is a clean and uncommon angle on reward misspecification, and AP-GRPO is a low-cost realization of it.
• Experimental Thoroughness: ⭐⭐⭐⭐☆ Seven benchmarks spanning in-distribution, out-of-distribution, and multiple resolutions, plus a grounding-specific study, \(k\) sensitivity, a stage-order ablation, and cross-base-model transfer; but the main text reports only 7B, and inference cost and stricter-threshold main results are missing.
• Writing Quality: ⭐⭐⭐⭐☆ The pilot experiment lays out "36.5% of the rewards are misspecified" before presenting the method, making the motivation easy to follow; some formulas (Eq. 6) are garbled in extraction and several conclusions are pushed to the appendix, so readers must cross-reference.
• Value: ⭐⭐⭐⭐☆ It offers a general post-training recipe for optimizing localization directly without box annotations, and the "change the context rather than the loss" idea transfers to other tasks whose intermediate quality is unobservable.

Related Papers

• [ICLR 2026] MedVR: Annotation-Free Medical Visual Reasoning via Agentic Reinforcement Learning
• [ECCV 2026] SyncLoop: A Multimodal Dual-Loop Framework for Self-Improving Mathematical Reasoning
• [ECCV 2026] Reasoning with Memory: A Temporal Granularity-Adaptive Framework for Training-Free Long Video Understanding
• [CVPR 2026] DeepScan: A Training-Free Framework for Visually Grounded Reasoning in Large Vision-Language Models
• [CVPR 2026] Seeing What Matters: A Training-Free Self-Guided Framework for Multimodal Detail Perception and Reasoning