Graph learning traditionally relies on static approaches that process text-attributed graphs (TAGs) through predetermined feature extraction. Real-world applications like social networks, citation graphs, and e-commerce platforms contain rich semantic content alongside topological structure, but existing methods struggle to adaptively explore these multi-scale dependencies during inference.

Agentic Graph Learning (AGL) formulates graph learning as a sequential decision process. Given a TAG G = (V, A, T) where V is nodes, A is adjacency matrix, and T contains node texts, the agent must predict labels for node classification or link prediction by iteratively exploring the graph structure. The formal objective optimizes policy πθ to maximize expected reward:

Success is measured by classification accuracy on standard TAG benchmarks. The paper introduces evaluation across 7 datasets spanning citation networks (OGB-Arxiv, PubMed, Arxiv-2023), Amazon products (OGB-Products, Amazon-Photo, Amazon-Computers), and social networks (Reddit) with train/test splits following GraphICL protocols.

1. Graph-native search tools enabling multi-scale exploration: τ1HOP (local neighborhood), τ2HOP (2-hop neighborhood), τSS (structure salience using PPR scores), and τDENSE (semantic similarity-based search)

2. Two-stage reinforcement learning framework with LLM backbones (Qwen2.5-3B/7B-Instruct) optimized via GRPO and REINFORCE++ algorithms

3. Stage 1 - Graph-native policy bootstrapping using composite reward:

   \[R(τ) = r_{FMT}(τ) + r_{ACC}(ŷ, y) + r_{COV}(τ)\]

4. Stage 2 - Search-constrained thinking with retrospective termination trigger and cognitive density regularization to prevent over-searching while maintaining reasoning quality

5. Graph-conditioned curriculum learning (GCCL) that orders training examples by difficulty using topological (homophily, degree) and semantic priors

6. Interactive trajectory format with ... blocks containing tool calls via <

   begin_of_query

   >tool_name:query<

   end_of_query

   > interface

1. Stage 1 - Graph-native policy bootstrapping: Trains on OGB-Arxiv and OGB-Products with 3,000 training nodes each using coverage reward rCOV to encourage exploration of all four search tools, format reward rFMT for structured output, and accuracy reward rACC

2. Stage 2 - Search efficiency optimization: Transitions from exploration to exploitation by removing coverage reward and adding reasoning density penalty rdepth = α·I[Nshort = 0] - λd·Nshort to encourage deeper reasoning over additional searches

3. Graph-conditioned curriculum learning applied throughout both stages, ordering instances from easy to hard using analytical scoring functions based on neighbor label consistency for node classification and semantic-structural alignment for link prediction

4. Optimization via GRPO and REINFORCE++ algorithms with maximum search budget B=4 tools per episode

5. Hardware and wall-clock time not reported. Learning rates, batch sizes, and specific training schedules not specified.

Prior work: GraphGPT (2024) performs graph instruction tuning with static context, GraphCoT (2024) uses heuristic prompting for graph QA, GraphICL (2025) applies in-context learning to graphs. These methods rely on fixed graph representations provided at inference time.

Delta: This paper introduces the first RL-driven framework for agentic graph learning where an LLM agent dynamically explores graph structure through tool use. Key additions include:

1. formulating graph learning as sequential decision-making with graph-native search operators
2. two-stage curriculum-based RL training without step-wise supervision, and

3. search-constrained thinking to balance exploration and reasoning depth.

   Applied-specific assessment: The architectural idea of treating graph learning as agentic exploration is genuinely novel - prior GraphLLMs and GraphRAG methods use static context injection. Benchmark gains are substantial (up to 17.5% node classification, 28.4% link prediction) and hold across diverse datasets and model scales. However, comparisons may be somewhat unfair as baselines weren’t designed for this interactive paradigm. The approach requires significant computational overhead from multi-step tool usage and RL training.

   The gains likely depend on having sufficient exploration budget and would diminish with reduced compute resources. The method introduces a new problem formulation rather than just applying known techniques.

   Verdict: SIGNIFICANT — introduces a novel paradigm for graph learning that demonstrates clear advantages over static approaches, though computational requirements may limit adoption.

1. Node Classification (in-domain): AgentGL achieves 68.9% on OGB-Arxiv vs 66.4% best baseline (GraphICL), 77.0% on OGB-Products vs 70.4% best baseline, improvements of 2.5-6.6% absolute

2. Node Classification (zero-shot): Up to 17.5% absolute improvement over best baselines across PubMed, Amazon-Photo, Amazon-Computers, Arxiv-2023, Reddit datasets

3. Link Prediction (in-domain): 95.9% on OGB-Arxiv vs 86.6% best baseline (Search-R1), 96.5% on OGB-Products vs 89.1% best baseline, improvements of 7.4-9.3% absolute

4. Link Prediction (zero-shot): Up to 28.4% absolute improvement over baselines, with particularly strong gains on Amazon-Computers (94.3% vs 74.0%) and Reddit (88.5% vs 59.0%)

5. Consistent improvements across both Qwen2.5-3B and 7B backbones, with larger models showing better zero-shot transfer

6. Outperforms GNN baselines (GCN, RevGAT, GraphSAGE), GraphLLMs (GraphGPT, LLaGA, GraphPrompter, GraphICL), GraphRAG methods (LinearRAG, HippoRAG, GraphCoT), and standard agentic search (Search-R1, Search-O1)

1. Model size: Qwen2.5-3B and Qwen2.5-7B parameter backbones, specific parameter counts for final AgentGL system not reported

2. Training compute: GPU hours and hardware specifications not reported in paper

3. Inference speed/latency: Each episode allows maximum 4 tool calls, with average actual usage of 3.17-3.56 tools per instance after full training

4. Memory footprint: Not specified, depends on backbone LLM size

5. Deployment practicality: Requires significant computational overhead from multi-turn tool usage during inference and complex two-stage RL training procedure, making deployment challenging compared to static GraphLLM approaches

1. Evaluation conducted on real-world graph datasets including citation networks (scientific papers), e-commerce graphs (Amazon products with co-purchase links), and social networks (Reddit posts)

2. Zero-shot transfer experiments demonstrate generalization across different domains without retraining on target datasets

3. No mention of actual production deployments or integration with live graph systems

4. Method operates on text-attributed graphs with natural language node content, matching real-world data formats

5. Limited to relatively small subsampled graphs (1,000-3,000 nodes for evaluation) which may not reflect true large-scale deployment challenges

6. No discussion of real-time inference requirements or streaming graph updates

1. FUNDAMENTAL: Limited to text-attributed graphs, cannot handle multimodal node attributes despite many real graphs containing images, audio, or other data types

2. FUNDAMENTAL: Scalability unclear for very dense graphs or those requiring exploration beyond 4-hop neighborhoods

3. ENGINEERING: Critical dependence on careful data allocation between training stages, with MSO stage stability requiring precise hyperparameter tuning

4. ENGINEERING: Computational overhead from multi-turn tool usage makes inference significantly more expensive than static methods

5. EVALUATION: Evaluation limited to relatively small subsampled graphs, scalability to truly large graphs uncertain

6. EVALUATION: May alter distribution of tool usage during inference compared to training, though this wasn’t thoroughly investigated

   Failure modes:

7. Agent may get stuck in redundant search loops when graph structure provides insufficient signal
8. Performance likely degrades significantly when search budget is severely constrained below 3-4 tools per instance.

Long-duration videos contain critical cues scattered sparsely across time, making them intractable for multimodal large language models (MLLMs) with limited context windows. Understanding such videos requires identifying relevant moments that may be temporally distant yet contextually related. This problem is challenging because direct uniform sampling risks missing fine-grained details due to sparse sampling or overwhelming models with redundant visual noise.

The task is long video question answering. Input: video sequences (potentially hours long) and natural language queries. Output: textual answers requiring reasoning across multiple video segments. The core challenge is retrieving and organizing query-relevant visual evidence from extensive temporal contexts.

Success is measured by accuracy on video QA benchmarks including LongVideoBench (LV-Bench), MLVU, Video-MME-Long, and EgoSchema. These evaluate the model’s ability to answer questions requiring temporal reasoning, event understanding, and cross-modal comprehension.

The paper introduces IR-600K, a dataset with 605,676 training samples from ActivityNet-QA, NExT-QA, and STAR, specifically designed for learning frame-query intent alignment with 5-level relevance labels.

1. Event boundary detection: Videos are segmented into semantically coherent clips using PELT change-point detection applied to frame embeddings from Qwen2.5-VL vision tower.

2. Spatio-temporal graph construction: Each clip becomes a node with embedding $h_k = E_{vl}(s_k)$ from video-language encoder. Temporal edges connect adjacent clips with $w_{k,k+1} = 1$. Spatial edges connect all clips with weights $w_{i,j} = \cos(h_i, h_j)$.

3. Graph-based clip retrieval: Query embedding $h_q = E_{vl}(q)$ identifies top-N anchor clips by cosine similarity. Multi-hop traversal expands context via:

   \[V_{hop} = \{ v_j | d(v_j, V_{anc}) \leq L, w_{ij} \geq \eta \}\]

4. Intent-aware frame retrieval: Lightweight MLLM scorer $R_\theta$ (Qwen2.5-VL-3B with LoRA) outputs probability distribution over relevance levels 1-5:

   \[P_\theta(\ell | q, x_t, P_{intent}) = \frac{\exp(\pi_\theta(\ell | q, x_t, P_{intent}))}{\sum_{k=1}^5 \exp(\pi_\theta(k | q, x_t, P_{intent}))}\]

5. Relevance scoring: Final score computed as weighted expectation:

   \[r_t = R_\theta(q, x_t, P_{intent}) = \sum_{\ell=1}^5 \ell \cdot P_\theta(\ell | q, x_t, P_{intent})\]

   The core contribution is shifting from flattened semantic matching to structured, intent-aware retrieval that preserves spatio-temporal video topology.

1. Intent-relevance scorer training: - Data: IR-600K dataset with 605K training samples, teacher labels from Qwen2.5-VL-72B-Instruct - Student model: Qwen2.5-VL-3B-Instruct with LoRA (rank=16, scaling=32, dropout=0.05) - Optimizer: AdamW with lr=5×10⁻⁵, weight decay=0.05 - Schedule: Cosine with warmup ratio 0.05, 1 epoch, batch size=128, bf16 precision - Hardware: 8×A100 GPUs

2. Loss function: Cross-entropy between student prediction and teacher discrete label:

   \[L_{CE} = -\sum_{\ell=1}^5 \mathbf{1}[\ell = y_t] \log P_\theta(\ell | q, x_t, P_{intent})\]

3. Other components: - Event boundary detector: No additional training, uses pretrained Qwen2.5-VL vision tower - Graph construction: Uses pretrained Perception Encoder as video-language encoder - Hyperparameters: N=3 anchors, L=2 hops, η=0.4 edge threshold, κₛ=3.25 relevance threshold

Prior work:

1. Video-RAG (2025): Retrieves keyframes via semantic similarity with CLIP, enriches with external tools like OCR and object detectors.
2. TV-RAG (2025): Temporal-aware retrieval with semantic entropy weighting for long video understanding.

3. Existing long-video RAG methods: Flatten videos into independent segments, rank by embedding similarity from contrastive models.

   Delta: This paper adds two key components:

4. Spatio-temporal graph modeling that preserves video topology via temporal/spatial edges and multi-hop traversal, and

5. Intent-relevance scoring using MLLM to assess frame-query intent alignment beyond semantic similarity.

   Applied-specific assessment:

   • Architectural novelty: The spatio-temporal graph representation for video RAG is relatively novel, though graph-based retrieval exists in other domains. The intent-relevance scorer using MLLM for frame ranking is a reasonable but incremental extension.
   • Benchmark gains: Improvements are modest (1-6% across benchmarks). On EgoSchema, achieves 67.2% vs 66.2% for IG-VLM baseline - within potential noise margin.
   • Fair comparisons: Methods compared don’t use auxiliary tools, making comparisons reasonable. However, gains are small and may not hold without similar computational budget.
   • Scale dependency: The approach requires training the intent scorer on 605K samples and uses relatively large models (3B-72B parameters), suggesting gains may depend on this scale.

   Verdict: INCREMENTAL — Solid engineering combining known techniques (graph retrieval + MLLM scoring) with modest improvements that don’t clearly exceed noise margins on most benchmarks.

1. LongVideoBench (LV-Bench): Overall accuracy ranges from 52.1-66.0% across models, VideoStir gains 2.5-6.6% over baselines
2. MLVU: Overall accuracy 50.4-74.1%, VideoStir gains 0.8-3.2%
3. Video-MME-Long (w/o Sub): Overall accuracy 45.6-62.1%, VideoStir gains 1.0-2.9%

4. EgoSchema test set: 67.2% accuracy vs 66.2% for IG-VLM baseline (1.0% improvement)

   Results show consistent but modest improvements across benchmarks. The gains are relatively small (typically 1-3%) and could be within statistical noise for some comparisons. VideoStir performs competitively but doesn’t demonstrate breakthrough-level performance jumps. Notably absent are results on other major video understanding benchmarks like Video-ChatGPT or VideoQA datasets beyond those used for training data construction.

1. Model size: Intent-relevance scorer uses Qwen2.5-VL-3B (3 billion parameters) with 3.7M LoRA parameters for adaptation
2. Training compute: 8×A100 GPUs for 1 epoch on IR-600K dataset, additional components use pretrained models (Perception Encoder, Qwen2.5-VL vision tower)
3. Inference speed/latency: Not reported, but acknowledged as limitation - system incurs additional latency from video-to-graph conversion and multi-stage retrieval
4. Memory footprint: Not explicitly reported, but requires storing graph representations and multiple model components
5. Deployment practicality: Limited by multi-stage pipeline complexity, requires significant computational resources for both graph construction and MLLM scoring, making real-time deployment challenging

1. Dataset evaluation only: The method is evaluated exclusively on curated video QA benchmarks (LongVideoBench, MLVU, Video-MME-Long, EgoSchema) without real-world deployment experiments

2. No hardware experiments: No testing on specific robots, vehicles, or production environments reported

3. Limited production integration: No discussion of integration with real-world video analysis systems or streaming applications

4. Synthetic/curated data focus: The approach is tested on controlled benchmark datasets rather than noisy, real-world video content with varying quality, lighting, or recording conditions

1. System latency (ENGINEERING) - VideoStir introduces additional steps for video-to-graph conversion and multi-stage retrieval, increasing end-to-end processing time

2. Computational overhead (ENGINEERING) - Requires multiple model components (event detector, graph encoder, intent scorer) with significant memory and compute requirements

3. Hyperparameter sensitivity (ENGINEERING) - Performance depends on carefully tuned parameters (N=3, L=2, η=0.4, κₛ=3.25) that may not generalize across different video types or domains

4. Limited evaluation scope (EVALUATION) - Only tested on specific video QA benchmarks, lacks evaluation on diverse real-world video content or streaming scenarios

   Failure modes:

5. Poor event boundary detection - When video contains gradual transitions or continuous activities, segmentation may fail to identify meaningful clips
6. Intent-relevance misalignment - The scorer may assign high relevance to visually similar but contextually irrelevant frames, especially for complex reasoning queries requiring multi-step inference

Multi-agent embodied systems face critical challenges in spatial coordination, temporal reasoning, and shared workspace awareness when multiple robotic arms must collaborate on manipulation tasks. Unlike single-agent settings where policies focus solely on task completion, multi-agent collaboration demands intricate reasoning about inter-agent interactions, collision avoidance, and synchronized execution - particularly challenging when agents have different morphologies and must coordinate in shared physical workspaces.

The task involves multi-agent robotic manipulation where N agents with joint actions $\mathbf{a}_t = [a_t^{(1)}, \dots, a_t^{(N)}]$ must collaborate to achieve task goal $\mathcal{G}$. Input consists of multi-view RGBD observations $\mathbf{o}_t = {o_t^{(1)}, o_t^{(2)}, \dots, o_t^{(K)}}$ from K calibrated cameras. The system must convert these to simulator state $s_t \in \mathcal{S}$ via scene conversion operator:

Success is measured by subtask completion rates S_i and overall task success rate SR across manipulation benchmarks including cube stacking, bimanual ball pickup, cylinder handover, collaborative assembly, and coordinated sweeping tasks.

The paper evaluates on five real-world tasks with 2-3 heterogeneous robotic arms (Franka Research 3 and AgileX Piper platforms), with 10 trials per execution mode on each task.

1. Real-to-Sim Scene Reconstruction: Multi-view RGBD observations converted to simulator-ready scenes via Grounded SAM2 object detection, GPT-5 visual disambiguation, FoundationPose 6-DoF pose estimation with multi-view fusion:

   \[\hat{\mathbf{t}} = \frac{1}{K'} \sum_{k \in \mathcal{K}'} \mathbf{t}^{(k)}, \quad \hat{\mathbf{R}} = \mathcal{M}(\{\mathbf{R}^{(k)}\}_{k \in \mathcal{K}'})\]

2. VLM-Driven Action Synthesis: Hierarchical decomposition where GPT-5 planner decomposes task goal $\mathcal{G}$ into semantic sub-goals $\mathcal{H} = {h_1, h_2, \dots, h_L}$ then assigns execution plans $\mathcal{E} = {e_1, e_2, \dots, e_L}$ with action primitives $\rho_l \in {\textsc{Move}, \textsc{Grasp}, \textsc{Place}, \textsc{Rotate}}$

3. Adaptive Camera Control: Multi-viewpoint aggregation to handle occlusions:

   \[\hat{o}_t = \textsc{Aggregate}(\{\textsc{Render}(\mathbf{s}_t, c_j)\}_{j=1}^{J})\]

4. Dual Execution Modes: Interactive mode with closed-loop VLM feedback and checkpoint verification $\phi(e_l^{\textsc{ckpt}}, \mathbf{s}_t) \in {0, 1}$; Iterative mode using Claude Code for complete trajectory generation $\mathcal{P}^{(m)} = \textsc{CodeGen}(\mathbf{s}_0, \mathcal{E}, \mathcal{F}^{(m-1)})$

5. Collision-Aware Sim-to-Real Transfer: Joint space interpolation with pre-execution collision volume verification $\mathcal{V}^{(i)}_l \cap \mathcal{V}^{(j)}_l = \emptyset$ for all agent pairs

1. No Traditional Training: The system operates via prompting foundation models rather than training new parameters

2. Real-to-Sim Calibration: Iterative camera extrinsic calibration refinement by comparing rendered views against real-world captures across multiple trials

3. Asset Generation: 3D mesh generation using Meshy Model platform with pre-defined physical properties (mass, friction, collision geometry)

4. Foundation Model Usage: - GPT-5 for visual reasoning, task decomposition, and interactive planning - Claude Code for trajectory generation in iterative mode - Grounded SAM2 for object detection and segmentation - FoundationPose for 6-DoF pose estimation

5. Data Collection Pipeline: Validated trajectories stored in knowledge base $\mathcal{D} = {(\mathbf{s}_0^{(j)}, \mathcal{G}^{(j)}, \tau^{(j)})}$ for future demonstrations

   Hardware details: Intel RealSense D435i cameras, ManiSkill simulator built on SAPIEN physics engine. Wall-clock time and specific computational costs not reported.

Prior Work:

1. RoboFactory (2025) - explored collaborative assembly scenarios but lacked fine-grained spatial reasoning and collision avoidance
2. RoCo (2024) - multi-agent planning with LLMs but relied on textual representations disconnected from physical environment

3. MALMM (2025) - distributed planning and communication but processed agent viewpoints in isolation

   Delta: This paper introduces “compositional environment” - synergistic integration of real-world and simulation components specifically for multi-agent embodied collaboration. Key additions:

4. unified decision-making space bridging sim and real
5. adaptive camera control for occlusion handling
6. dual execution modes with checkpoint verification

7. collision-aware sim-to-real transfer.

   Applied-Specific Assessment:

   • Architectural idea: The compositional environment concept combines known components (sim-to-real, VLM planning) but the specific integration for multi-agent coordination addresses a real gap
   • Benchmark gains: 49% overall success rate across challenging multi-agent tasks is meaningful given task complexity, though baselines are limited
   • Fair comparisons: Limited baseline comparisons - mainly ablation studies rather than comparison to other multi-agent systems
   • Scalability concerns: Heavy reliance on expensive foundation models (GPT-5, Claude Code) may not scale without similar computational resources

   Verdict: INCREMENTAL - Solid engineering contribution that combines existing techniques effectively for multi-agent robotics, but lacks fundamental algorithmic novelty beyond integration of known components.

1. Cube Stacking: 75% overall success (Interactive: 6/10, Iterative: 9/10) - best performing task due to spatially separated subtasks

2. Ball Pickup: 50% overall success (Interactive: 4/10, Iterative: 6/10) - benefits from iterative mode’s precise trajectory control

3. Transfer Cylinder: 25% overall success (Interactive: 4/10, Iterative: 1/10) - most challenging due to tight bimanual handover constraints

4. Place Cucumber: 35% overall success (Interactive: 4/10, Iterative: 3/10) - three-agent task with heterogeneous robots

5. Brush Box: 60% overall success (Interactive: 7/10, Iterative: 5/10) - interactive mode excels at dynamic role coordination

   Ablation Results:

   • Without adaptive camera: 30% average success (vs 50% full system)
   • Without checkpoint verification: 20% average success (vs 50% full system)

   Missing Benchmarks: No comparison to other multi-agent robotic systems, single-agent baselines, or existing collaborative manipulation frameworks. Results are mixed across tasks with significant variance in difficulty.

1. Model Size: Uses foundation models (GPT-5, Claude Code, Grounded SAM2) - exact parameter counts not specified but likely billions of parameters

2. Training Compute: No training required - operates via inference on pre-trained foundation models

3. Inference Speed/Latency: Not reported, but system involves multiple VLM calls per primitive action which likely introduces significant latency

4. Memory Footprint: Not specified, but requires loading multiple large foundation models simultaneously

5. Deployment Practicality: Limited by dependence on expensive commercial APIs (GPT-5, Claude Code). Interactive mode averaging 1.5-2.5 episodes per session, iterative mode 9.5-17.5 episodes per session. Reset tokens account for 10-31% of total token consumption, indicating reasonable computational efficiency for data collection but high per-episode costs for real-time deployment.

1. Real Robot Deployment: Tested on physical Franka Research 3 arms and AgileX Piper dual-arm platforms in laboratory tabletop settings

2. Hardware Experiments: Five manipulation tasks across two hardware configurations - two-agent (Franka × 2) and three-agent (Franka + AgileX Piper × 2) setups

3. Environment Constraints: Robot bases remain fixed, operates in structured laboratory workspace with calibrated cameras, requires manual scene setup and object placement

4. Sim-to-Real Validation: Demonstrates collision-aware transfer with trajectory interpolation, but limited to relatively simple manipulation primitives

5. Production Limitations: High computational costs from foundation model APIs, requirement for multi-view camera calibration, and dependence on structured environments limit immediate production deployment. Success rates (25-75%) indicate system works but may need higher reliability for real applications.

1. Minor sim-to-real pose offsets cause contact-rich primitive failures (ENGINEERING - fixable with better calibration and pose estimation)

2. VLM planner enters repetitive re-planning cycles without sufficient strategy exploration (FUNDAMENTAL - inherent to current VLM reasoning capabilities)

3. Interactive mode accumulates drift over long action sequences (ENGINEERING - could be addressed with better state estimation and error correction)

4. Iterative mode struggles with reactive tasks requiring closed-loop adaptation (FUNDAMENTAL - code generation approach limits real-time reactivity)

5. Heavy dependence on expensive commercial foundation model APIs (ENGINEERING - could use open-source alternatives with performance tradeoffs)

6. Limited to structured laboratory environments with fixed robot bases (EVALUATION - testing scope restrictions)

   Failure Modes:

   • System fails when tight bimanual coordination required (e.g., cylinder handover with 25% success)
   • Occlusion handling breaks down in complex multi-agent scenes despite adaptive camera control

Vision-Language Models (VLMs) suffer from sparse credit assignment during alignment training when using reinforcement learning methods like GRPO/PPO. While standard RL training uses terminal outcome rewards, this creates attribution problems in multi-step reasoning where visual evidence must be linked to intermediate reasoning steps.

Task definition: Given multimodal inputs x = (I, q) where I is an image and q is a text instruction/question, a VLM policy πθ generates token sequences y = (y₁, …, yT). The alignment objective uses terminal rewards R(x, y) or preferences, but sparse signals lead to unstable optimization and visual hallucinations.

Evaluation: Performance measured on multimodal reasoning benchmarks including MathVista (mathematical visual reasoning), MMStar (multimodal understanding), MMMU (multimodal knowledge), MathVerse (math with visual elements), MathVision (geometry problems), and AI2D (diagram understanding). Metrics are accuracy percentages.

The paper uses curated multimodal mathematical reasoning datasets for training.

1. Policy model πθ: Base VLMs including Qwen2.5-VL-7B/32B and InternVL3.5-8B-MPO generating autoregressive token sequences

2. Repair model ρϕ: Initialized from policy model, trained in two stages: - Stage 1 SFT: Learn to repair incorrect outputs using small dataset Drep - Stage 2 RL: Optimize repair reward encouraging correctness with minimal edits

   \[r_{\text{rep}}(x, \mathbf{y}, \tilde{\mathbf{y}}) = g_{\text{aud}} \cdot g_{\text{vis}} \cdot \left( r_{\text{task}}(x, \tilde{\mathbf{y}}) - \lambda \cdot \frac{\Delta_{\text{edit}}(\mathbf{y}, \tilde{\mathbf{y}})}{\max(1, |\mathbf{y}|)} \right)\]

3. Difference alignment: Apply Levenshtein edit distance to get token-level masks m marking corrections needed

4. Gated training: For incorrect trajectories, apply difference masks to gate gradient updates, focusing negative advantages only on error positions

   \[\hat{A}^{\text{DF}}_t = g_t \cdot \hat{A}_t\]

   Core technical contribution: Automatic construction of process-level supervision through repair-then-mask pipeline, avoiding expensive step-level human annotation while providing fine-grained credit assignment.

1. Stage 1 - Repair model SFT: - Data: 1,000 manually annotated repair samples (x, y_incorrect, y_correct, answer) - Optimizer: Not reported - Hardware: 32 NVIDIA A800 GPUs - Learning rate: Not reported for SFT stage

2. Stage 2 - Repair model RL: - Data: Same as policy training data, synthetic repairs - Optimizer: Same as policy (GRPO/GSPO) - Edit penalty coefficient λ = 0.5 - Hardware: 32 NVIDIA A800 GPUs

3. Policy training with DF: - Data: Curated multimodal mathematical reasoning dataset (scale not reported) - Optimizer: DF-GSPO or DF-GRPO - Learning rate: 1×10⁻⁶ - Batch size: 128 (Qwen) or 512 (InternVL) rollout batch size - Temperature: 0.7 - Hardware: 32 NVIDIA A800 GPUs - Wall-clock time: Not reported, but compute-matched across methods

Prior work:

1. “Let’s verify step by step” (Lightman et al. 2023): Process Reward Models requiring expensive step-level human annotation
2. “Free process rewards without process labels” (Yuan et al. 2024): Implicit process rewards suffering from language-reasoning capability trade-offs

3. Various token-level reward redistribution methods (Chan et al. 2024, Li et al. 2024) that redistribute terminal rewards but don’t localize error positions

   Delta: This paper introduces repair-based difference masking to automatically generate process-level supervision without human step annotation. Instead of redistributing rewards, it identifies which tokens need correction via counterfactual repair trajectories.

   Applied-specific assessment:

   • Architectural idea: Novel application of repair models + edit alignment to create supervision masks, but builds on well-known edit distance and masking concepts
   • Benchmark gains: Consistent 2-3% improvements across multiple benchmarks and model scales, though modest in absolute terms
   • Fair comparisons: Uses compute-matching protocol accounting for repair model training overhead
   • Scale dependence: Approach should work without proprietary data, uses standard base models

   Verdict: INCREMENTAL — Solid engineering contribution that automatically generates process supervision, but the core ideas (repair models, edit masking) are known techniques applied to a new setting with modest empirical gains.

1. MathVista: DF-GSPO 7B achieves 76.1%, baseline GSPO 73.6% (+2.5% improvement)
2. MMStar: DF-GSPO 7B achieves 68.3%, baseline GSPO 65.9% (+2.4% improvement)
3. MMMU: DF-GSPO 7B achieves 58.4%, baseline GSPO 56.7% (+1.7% improvement)
4. MathVerse: DF-GSPO 7B achieves 53.7%, baseline GSPO 50.8% (+2.9% improvement)
5. MathVision: DF-GSPO 7B achieves 30.4%, baseline GSPO 27.4% (+3.0% improvement)

6. AI2D: DF-GSPO 7B achieves 85.6%, baseline GSPO 84.5% (+1.1% improvement)

   Similar consistent improvements at 32B scale and with InternVL backbone. Results show DF consistently outperforms baselines across all benchmarks, with improvements ranging 1-3%. The method also outperforms PRIME (implicit process rewards) across all metrics.

1. Model size: 7B, 32B parameter VLMs (Qwen2.5-VL, InternVL3.5-8B)
2. Training compute: 32 NVIDIA A800 GPUs, compute-matched protocol accounts for repair model training overhead
3. Inference speed: Additional repair model inference only for incorrect samples, expected cost ≈ Cost_base + p_err × Cost(repair decode)
4. Memory footprint: Not reported, but requires storing both policy and repair models
5. Deployment assessment: Favorable scalability compared to test-time search methods, no exponential branching during inference, repair triggered only for error samples so overhead decreases as training progresses

1. No deployment results reported on real-world applications
2. No hardware experiments on specific robots or vehicles
3. No production integration examples
4. No sim-to-real transfer discussion
5. Evaluation limited to standard academic benchmarks with curated datasets
6. Method designed for training-time improvement rather than real-world deployment scenarios

1. FUNDAMENTAL: Misalignment between surface-level edits and true causal error spans due to paraphrasing or reasoning style differences
2. ENGINEERING: Repair model quality dependency - inaccurate or overly large repairs introduce noisy difference masks
3. ENGINEERING: Additional inference overhead from repair model, though modest compared to test-time search
4. EVALUATION: Requires reliable correctness signals which may not be available for all tasks

5. FUNDAMENTAL: Edit distance alignment may not capture semantic correspondence in complex reasoning chains

   Failure modes:

6. Mask hallucination: When repairs involve compensatory reasoning changes that don’t reflect true error locations
7. Reward hacking: Model generating superficial repairs that appear correct but don’t address underlying reasoning errors

Long video understanding (LVU) represents a critical challenge in multimodal AI, as current vision-language models struggle to maintain coherent narrative understanding across extended temporal contexts. The practical need is pervasive across video analysis tasks including film understanding, surveillance, educational content, and entertainment media where narrative structure spans minutes to hours.

The task evaluates models’ ability to reason over scene-level contexts in long videos. Input consists of videos averaging 1,978 seconds (∼33 minutes) ranging from 1 minute to 4+ hours. Output varies by task:

1. SceneQA requires answering questions requiring multi-minute context integration
2. SceneQA-Audio adds audio modality requirements
3. I-VQA involves selecting correct questions given answers
4. Comment Prediction determines video-comment relevance
5. Title Prediction generates video summaries, and

6. ClipQA handles shorter temporal spans.

   Success is measured through accuracy on 8,903 question-answer pairs across 6 task types. The key innovation is measuring temporal dependency: SceneQA requires cues spanning at least 2 minutes, with average temporal distance of 262 seconds between questions and supporting evidence.

   SceneBench introduces 2,485 videos with comprehensive scene-level annotations, systematically evaluating how performance degrades as temporal distance increases between question context and answer evidence.

Scene-RAG introduces a three-stage retrieval-augmented generation framework:

1. Scene Tiling: Detects coherent visual segments using Total Variation with L1 regularization (TV-L1). The optimization minimizes:

   \[\min_{x \in \mathbb{R}^n} \frac{1}{2} \sum_{t=1}^n (x_t - s_t)^2 + \lambda \sum_{t=2}^n |x_t - x_{t-1}|\]

2. Memory Construction: Encodes scene segments using InternVideo2-6B for visual representation and Qwen-Audio2 for audio captioning. Creates multimodal memory bank aligning visual and audio features temporally.

3. Query Retrieval: Decomposes user queries using Qwen3-14B, performs similarity search over scene embeddings, retrieves top-K=10 relevant scenes for final reasoning.

   The core contribution distinguishes this from prior RAG approaches by organizing memory around semantic scenes rather than uniform clips, enabling retrieval of temporally distant but contextually related segments.

Scene-RAG utilizes pre-trained components without additional training:

1. Visual encoding: InternVideo2-6B backbone frozen during inference, using stage-2 pretrained weights
2. Audio processing: Qwen-Audio2 for speech and ambient sound transcription

3. LLM reasoning: Qwen3-14B for query decomposition and final generation

   Hyperparameters: TV-L1 regularization λ=1.5, sensitivity α=1.5, minimum segment length L_min=3s, retrieval top-K=10, retrieved-to-sampled frame ratio 0.5.

   Hardware: Single H800 80G GPU. Runtime breakdown shows 276.72s offline preprocessing (273.52s visual encoding, 3.20s audio) and 62.29s online inference (1.04s visual, 0.19s audio, 61.06s LLM) for 2,767s video.

   No training data, learning rates, or optimization details reported as method uses frozen pretrained components.

Prior work: Video-RAG (2024) provides frame-level retrieval for long videos but uses uniform sampling. VideoRAG (2024) adopts straightforward frame sampling and vector database storage. MovieChat (2023) introduces memory modules for recursive video reasoning.

Delta: This paper adds (1) scene-based memory organization using TV-L1 regularization instead of uniform frame sampling, (2) multimodal memory combining visual and audio representations, (3) systematic benchmark revealing sharp performance drops with increasing temporal distance.

Applied-specific assessment: The architectural TV-L1 scene detection is a standard signal processing technique applied to video segmentation - not architecturally novel. Benchmark gains are modest (+1.1-1.7% on SceneBench, +8.6-11.8% on Video-MME) but consistent across models. Comparisons appear fair using same base models. However, the core contribution is more benchmark than method - Scene-RAG improvements are incremental over Video-RAG.

The real value lies in systematically demonstrating the temporal forgetting phenomenon and providing a comprehensive evaluation framework. The benchmark design exposing scene-level reasoning challenges is solid, but the retrieval method itself applies known techniques.

Verdict: INCREMENTAL — Solid benchmark contribution with systematic evaluation of temporal forgetting, but Scene-RAG method applies standard techniques with modest improvements.

1. SceneBench Title Prediction: Scene-RAG achieves 97.7-98.3% vs baseline 96.2-96.8%, improvement +0.9-1.5pp
2. SceneBench Comment Prediction: Scene-RAG achieves 71.5-85.4% vs baseline 71.1-82.0%, improvement +0.4-3.4pp
3. SceneBench ClipQA: Scene-RAG achieves 54.6-71.9% vs baseline 54.3-70.1%, improvement +0.3-1.8pp
4. SceneBench SceneQA: Scene-RAG achieves 25.6-28.9% vs baseline 23.6-27.7%, improvement +1.2-2.0pp
5. SceneBench SceneQA-Audio: Scene-RAG achieves 28.0-32.4% vs baseline 25.4-29.5%, improvement +2.6-2.9pp
6. SceneBench I-VQA: Scene-RAG achieves 37.0-44.1% vs baseline 33.0-46.5%, improvement +1.4-4.0pp
7. Video-MME Overall: Scene-RAG achieves 50.6-63.4% vs baseline 43.0-52.0%, improvement +7.6-11.4pp

8. MLVU: Scene-RAG achieves 74.1% vs baseline 70.8%, improvement +3.3pp

   Results show consistent but modest gains, with largest improvements on audio-augmented tasks. Performance drops sharply as temporal distance increases, confirming the forgetting hypothesis. Notable absence of evaluation on standard video QA benchmarks like MVBench.

1. Model size: Uses frozen InternVideo2-6B (6B parameters) + Qwen-Audio2 + Qwen3-14B (total ~20B parameters)
2. Training compute: No training required, uses pretrained frozen components
3. Inference speed: 62.29s online inference + 276.72s offline preprocessing for 2,767s video (46-minute video)
4. Memory footprint: Dynamic memory bank size scaling with video length, top-K=10 retrieval reduces storage needs
5. Deployment practicality: Reasonable for offline processing but 277s preprocessing makes real-time applications challenging. Single H800 80G requirement limits accessibility.

1. Dataset uses real YouTube videos across films, vlogs, documentaries with manual verification for legal redistribution
2. No deployment results or production integration reported
3. No hardware experiments beyond single GPU evaluation
4. Evaluation limited to curated benchmark datasets rather than diverse real-world scenarios
5. Manual annotation process taking 36 minutes per QA pair suggests significant human effort for scaling
6. Focuses on post-hoc analysis rather than real-time video understanding applications

1. FUNDAMENTAL: Scene detection via TV-L1 may miss semantically related but temporally distant segments, as it relies on local visual similarity
2. ENGINEERING: 277s preprocessing time per video makes real-time applications impractical
3. EVALUATION: Benchmark limited to 6 video genres, may not generalize to other domains like surveillance or medical video
4. FUNDAMENTAL: Manual annotation requirement (36 min/QA pair) severely limits scalability
5. ENGINEERING: Requires 80GB GPU memory, limiting accessibility

6. EVALUATION: No comparison to human performance on scene-level tasks

   Failure modes:

7. Visually similar scenes (different classrooms) cause retrieval confusion
8. Ambiguous references (“person in red” with multiple candidates) lead to incorrect context retrieval.