Paper deep dive
Sparse Evidence Can Suffice: Agentic Evidence Seeking for Multimodal Video Misinformation Detection
Haochen Zhao, Yongxiu Xu, Xinkui Lin, Dong Xie, Jiarui Lu, Yuqi Qian, Yubin Wang, Hongbo Xu, Gaopeng Gou
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 91%
Last extracted: 7/21/2026, 6:15:50 AM
Summary
The paper introduces SIEVE, a framework for multimodal video misinformation detection that decouples evidence acquisition from verification. It employs an agentic evidence-seeking component to actively explore and extract sparse, decision-relevant clues from visual, textual (OCR/ASR), and contextual channels, constructing a compact evidence package. This package is then used by a verifier to determine veracity. The agent is trained using supervised trajectories and reinforcement learning to prioritize informative evidence while minimizing redundancy.
Entities (10)
Relation Signals (9)
SIEVE → uses → Verifier
confidence 95% · which is then used by a verifier to determine veracity.
SIEVE → uses → Evidence Agent
confidence 95% · An evidence-seeking agent actively explores the available multimodal evidence and constructs a compact evidence package
SIEVE → evaluatedon → FakeSV
confidence 90% · Experiments on multiple video misinformation benchmarks show that SIEVE consistently outperforms the evaluated baselines
SIEVE → evaluatedon → FakeVV
confidence 90% · For cross-dataset evaluation on FakeVV, the associated title is used
Evidence Agent → reads → ASR
confidence 90% · The agent selectively reads video-derived evidence channels, including visual frames, OCR, and ASR
Evidence Agent → reads → OCR
confidence 90% · The agent selectively reads video-derived evidence channels, including visual frames, OCR, and ASR
Evidence Agent → trainedwith → GRPO
confidence 90% · During GRPO refinement, all trajectories sampled for the same input are first scored
Verifier → evaluatedby → Kimi k2.5
confidence 85% · Kimi K2.5 (Kimi Team et al. 2026) evaluates packages from both sources
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Multimodal video misinformation detection is commonly formulated as a holistic video-understanding task, where the entire video and its associated content are processed and judged in a single pass. However, real-world misinformation often exhibits a sparse and compositional evidence structure: a reliable decision may depend on only a few coupled clues, while most video content contributes limited additional information. Exhaustive multimodal reasoning may therefore introduce substantial redundancy and obscure decisive evidence. This motivates decoupling evidence acquisition from verification: first identifying sparse, decision-relevant clues and then judging veracity based on the acquired evidence. Accordingly, we propose SIEVE, a framework for Sparse Interactive Evidence Verification via Extraction in multimodal video misinformation detection. An evidence-seeking agent actively explores the available multimodal evidence and constructs a compact evidence package, which is then used by a verifier to determine veracity. The agent is trained with supervised evidence-seeking trajectories and an evidence-aware reinforcement learning objective that promotes informative evidence acquisition while discouraging unnecessary or invalid interactions. Experiments on multiple video misinformation benchmarks show that SIEVE consistently outperforms the evaluated baselines and supports reliable verification using compact evidence packages. Moreover, the resulting acquisition process provides an explicit and inspectable evidence trail, improving the transparency and groundedness of multimodal misinformation detection.
Tags
Links
- Source: https://arxiv.org/abs/2607.18080v1
- Canonical: https://arxiv.org/abs/2607.18080v1
PDF not stored locally. Use the link above to view on the source site.
Full Text
66,551 characters extracted from source content.
Expand or collapse full text
Sparse Evidence Can Suffice: Agentic Evidence Seeking for Multimodal Video Misinformation Detection Haochen Zhao1,2, Yongxiu Xu1,2 , Xinkui Lin1,2, Dong Xie3, Jiarui Lu1,2, Yuqi Qian1,2, Yubin Wang1,2, Hongbo Xu1, Gaopeng Gou1 Abstract Multimodal video misinformation detection is commonly formulated as a holistic video-understanding task, where the entire video and its associated content are processed and judged in a single pass. However, real-world misinformation often exhibits a sparse and compositional evidence structure: a reliable decision may depend on only a few coupled clues, while most video content contributes limited additional information. Exhaustive multimodal reasoning may therefore introduce substantial redundancy and obscure decisive evidence. This motivates decoupling evidence acquisition from verification: first identifying sparse, decision-relevant clues and then judging veracity based on the acquired evidence. Accordingly, we propose SIEVE, a framework for Sparse Interactive Evidence Verification via Extraction in multimodal video misinformation detection. An evidence-seeking agent actively explores the available multimodal evidence and constructs a compact evidence package, which is then used by a verifier to determine veracity. The agent is trained with supervised evidence-seeking trajectories and an evidence-aware reinforcement learning objective that promotes informative evidence acquisition while discouraging unnecessary or invalid interactions. Experiments on multiple video misinformation benchmarks show that SIEVE consistently outperforms the evaluated baselines and supports reliable verification using compact evidence packages. Moreover, the resulting acquisition process provides an explicit and inspectable evidence trail, improving the transparency and groundedness of multimodal misinformation detection. Introduction Figure 1: An example illustrating that sparse, claim-relevant evidence can suffice for video claim verification. Videos have become a dominant medium for online information consumption, especially on short-video platforms such as TikTok and YouTube, where user-generated videos can be rapidly created and widely disseminated (Bu et al. 2023; Venkatagiri et al. 2023). However, the same ease of video creation and dissemination also accelerates the spread of misinformation, threatening the credibility of online information ecosystems and public discourse (Bu et al. 2023; Lazer et al. 2018). Since the volume and velocity of online videos far exceed the capacity of manual fact-checking, automatic video misinformation detection has become an increasingly important research problem (Venkatagiri et al. 2023; Bu et al. 2023). Compared with text- or image-based misinformation, video misinformation is more challenging because videos contain heterogeneous visual, acoustic, textual, and social signals, as well as temporal dynamics and subtle cross-modal inconsistencies (Bu et al. 2023; Shang et al. 2021; Liu et al. 2023). Early video misinformation detection methods mainly rely on task-specific encoders and multimodal fusion strategies, such as modeling cross-modal correlations, social context, creative-process patterns, or neighborhood relations among event-related videos (Choi and Ko 2021; Liu et al. 2023; Qi et al. 2023a; Li et al. 2025; Bu et al. 2024; Qi et al. 2023b). Although these methods have achieved promising results, they are typically optimized as end-to-end discriminative classifiers, offering limited support for explicit, interpretable, and flexible multi-step evidence reasoning in open-world verification scenarios (Wang et al. 2025b; Zheng et al. 2025; Wang et al. 2026; Lang et al. 2026). Recently, vision-language models and multimodal large language models have shown strong potential for fake short-video news detection by leveraging their broad multimodal knowledge and reasoning capabilities (Zheng et al. 2025; Wang et al. 2025b; Hong et al. 2025; Zhang et al. 2025). Some methods introduce chain-of-thought reasoning to improve explainability (Hong et al. 2025), while others further explore long reasoning trajectories and reinforcement learning for video misinformation detection (Zhang et al. 2025; Li et al. 2026). Nevertheless, most existing methods still follow a single-pass or fixed-depth inference paradigm, where the model directly predicts a label from pre-extracted inputs. When critical evidence is sparse, fragmented, or insufficient, such models may rely on internal assumptions rather than actively acquiring targeted evidence, leading to unreliable or poorly grounded decisions (Yao et al. 2023; Zhang et al. 2023; Li et al. 2026; Wang et al. 2026). As shown in Figure 1, a video claim can sometimes be resolved from sparse and localized evidence without encoding the entire evidence space. The datasets provide Event and Description metadata, but SIEVE uses only the Description field as the initial claim context and excludes the Event field from all model inputs. In the example, one OCR clue, “A WOMAN BREASTFEEDING A CAT ON THE PLANE,” identifies the event asserted by the video post, while another, “IT WAS ALL A STAGING SHOT IN A STUDIO,” reveals that the depicted scene was staged. The former establishes what is being claimed, whereas the latter provides the evidence needed to refute it; considered together, these complementary clues are sufficient to reject the claim. This example highlights a challenge beyond multimodal fusion: how to identify and combine the sparse, decisive clues needed to resolve a claim within a largely redundant evidence space, where exhaustive evidence consumption is unnecessary and may obscure the evidence that determines the judgment. Accordingly, we formulate video misinformation detection as a budget-constrained evidence-seeking and verification process, where an acquisition budget encourages the agent to prioritize small sets of potentially decisive and complementary clues rather than exhaustively reading all available content. Specifically, the platform-provided Description field is treated as the initial claim context, while the dataset-provided Event field is not used. An evidence-seeking agent selectively reads video-derived evidence channels, including visual frames, OCR, and ASR, and constructs a compact evidence package in which complementary clues can be jointly interpreted. A verifier then makes the judgment from the Description and the acquired evidence. This design decouples evidence acquisition from verification and exposes both the acquired evidence and its compositional basis for inspection. Our main contributions are summarized as follows: • We formulate multimodal video misinformation detection as a selective evidence acquisition and verification problem and empirically investigate whether a compact subset of video-derived evidence can support reliable verification within a bounded multimodal context. • We propose SIEVE, an agentic framework that actively seeks potentially decisive and complementary multimodal evidence and constructs a compact evidence package for verification. The evidence agent is trained with supervised trajectories and evidence-aware reinforcement learning for grounded evidence acquisition. • We conduct experiments on multiple video misinformation benchmarks. SIEVE consistently outperforms the evaluated baselines and reaches near-saturated Macro-F1 with a moderate evidence budget. Its decoupled design also provides an explicit and inspectable evidence acquisition trail. Related Work Existing video misinformation detection methods usually formulate the task as holistic multimodal classification, jointly encoding visual, textual, acoustic, and social signals for prediction. Early studies investigate topic modeling, adversarial learning, and domain-specific multimodal fusion for fake news or COVID-19 misinformation videos (Choi and Ko 2021; Shang et al. 2021; Liu et al. 2023). Later works build short-video benchmarks and introduce richer contextual modeling. FakeSV incorporates multimodal content and social context for fake short-video news detection (Qi et al. 2023a), while subsequent methods exploit neighborhood relations among event-related videos (Qi et al. 2023b), creative-process cues (Bu et al. 2024), or heterogeneous social latent structures (Li et al. 2025). Despite these advances, most methods process pre-collected video inputs through a single discriminative pipeline, offering limited support for explicitly identifying which sparse multimodal clues are sufficient for grounded verification. Recent works leverage vision-language models and multimodal large language models to improve semantic reasoning and explainability in misinformation detection. Rationale-augmented VLM methods move from direct prediction to analysis-oriented detection (Zheng et al. 2025), while FakeSV-VLM adapts VLMs through progressive experts and event-level consistency checking (Wang et al. 2025b). ExMRD introduces a refining–retrieving–reasoning chain to produce explainable micro-video rumor predictions (Hong et al. 2025), and Fact-R1 further combines long-chain reasoning, preference alignment, and reinforcement learning for video misinformation detection (Zhang et al. 2025). More recent agentic frameworks, such as CSI and FactGuard, use multi-agent investigation or tool-augmented reasoning to strengthen evidence acquisition and decision reliability (Wang et al. 2026; Li et al. 2026). Retrieval-guided adaptation has also been explored for test-time fake news video detection (Lang et al. 2026). Despite these advances, most methods still reason over preconstructed global inputs or externally retrieved information, rather than learning how to seek compact, localized, and claim-relevant evidence from dense multimodal video content. Our work is also related to evidence-centered fact verification and selective perception. Textual fact verification has long emphasized the importance of retrieving supporting evidence before making a claim-level judgment (Thorne et al. 2018), and later work shows that evidence retrieval should be optimized for downstream verification utility rather than relevance alone (Zhang et al. 2023). In language-agent research, ReAct demonstrates that interleaving reasoning and actions enables models to acquire missing information instead of relying on internal assumptions (Yao et al. 2023). In video understanding, adaptive frame selection methods show that many video tasks can be solved by observing only a subset of informative frames (Wu et al. 2019). Unlike these lines of work, SIEVE studies video misinformation detection as a budgeted interactive evidence-seeking problem. It uses only the platform-provided Description field as the initial claim context, explicitly excludes the dataset-provided Event field, and actively acquires compact, claim-relevant evidence from OCR, ASR, and visual channels before evidence-based verification, shifting the task from fixed-input classification to budgeted active evidence acquisition. Figure 2: Overall framework of the proposed method. Method Additional details on evidence-space construction, training, reward design, and policy optimization are provided in the supplementary material. Overview Figure 2 presents SIEVE. Given a video and its Description, SIEVE constructs a structured multimodal evidence space. The Evidence Agent receives a one-time slot overview of thumbnails and channel availability for navigation only, then selectively reads evidence under a budget and emits STOPSTOP once sufficient. The overview is excluded from the evidence package and Verifier input. The Verifier predicts from the Description and acquired evidence, enabling compact, evidence-grounded verification. Problem Formulation Each instance is x=(,,c)x=(V,A,c), where V and A are the visual and audio streams and c is the fixed textual context. Both datasets provide Event and Description fields, while FakeSV additionally provides comments. SIEVE uses only Description, i.e., c=cdescriptionc=c_description, excluding Event and Comment from all model inputs. For cross-dataset evaluation on FakeVV, the associated title is used, i.e., c=ctitlec=c_title. The task is to predict y∈=Fake,Realy =\Fake,Real\. Conventional methods typically predict directly from the complete multimodal input. In contrast, we formulate detection as active evidence acquisition followed by evidence-based verification. The visual stream is divided into K temporal slots ()=sii=1KS(V)=\s_i\_i=1^K, each containing available channels from ℳ=OCR,ASR,CTX,KEYM=\OCR,ASR,CTX,KEY\. The evidence space is ℰ(,)=ei,m∣i∈1,…,K,m∈ℳi,E(V,A)=\e_i,m i∈\1,…,K\,\ m _i\, (1) where ℳi⊆ℳM_i denotes the channels available in slot sis_i. Rather than accessing all detailed contents of ℰE, SIEVE sequentially constructs an ordered evidence package TP_T and predicts y^=parse(gϕ(c,T)), y=parse\! (g_φ(c,P_T) ), (2) where gϕg_φ is the autoregressive Verifier. Evidence Space Construction SIEVE builds a slot-based evidence space from each video’s audiovisual content. TransNetV2 (Souček and Lokoč 2020) detects shot boundaries; short shots are merged and long ones subdivided. If over 10 slots remain, adjacent low-motion segments are merged. Each slot may contain OCR, ASR, CTX, and KEY evidence. PaddleOCR (Du et al. 2020) extracts OCR text ranked by temporal, spatial, and recognition-quality cues, while Whisper (Radford et al. 2023) produces slot-aligned transcripts. DINOv2 ViT-S/14 (Oquab et al. 2024) encodes visual candidates: CTX contains representative frames, whereas KEY contains one or two complementary frames selected by visual distinctiveness and OCR relevance. Before the first action, the environment provides a one-time navigation overview with one thumbnail and channel-availability indicators per slot; it is excluded from the evidence package and Verifier input. Channel state is tracked by t(i,m)∈−,,C_t(i,m)∈\ -, x, v\ for unavailable, unread, and read. Legal READs are x entries allowed by the remaining budget. Visual READs are capped at 24 accumulated CTX/KEY images, excluding actions that exceed the cap. A legal READ reveals the evidence and changes its state to v. Repeated, unavailable, or malformed actions immediately terminate the rollout, add no evidence, and incur an RL penalty. Interactive Evidence Acquisition We formulate evidence acquisition as a budgeted sequential decision process. At step t, the agent observes ot=(c,ω,ht,μt,t,bt)o_t=(c,ω,h_t, _t,C_t,b_t), where c is the claim context, ω is the one-time overview provided at t=1t=1, hth_t stores acquired evidence and source identifiers, μt _t is evidence memory, tC_t tracks channel state, and btb_t records the remaining READ budget, accumulated visual-image count, and visual READs blocked by the 24-image cap. The agent generates zt∼πθ(⋅∣ot)z_t _θ(· o_t) containing an action and, when applicable, a concise non-verdict memory update for the preceding legal READ. The environment stores the memory update, parses the command, and validates it against the legal read-action set treadA^read_t. A legal READ(slotsid,channel)READ(slot_sid,\channel\) reveals esid,channele_sid,channel and updates the package, history, coverage, and budget; the new evidence is summarized in the next response. STOPSTOP terminates acquisition. Repeated, unavailable, or malformed commands are invalid attempts: they add no evidence, immediately terminate the current acquisition rollout, and incur a behavioral penalty during reinforcement learning. Acquisition ends upon STOPSTOP, budget exhaustion, or invalid action. The terminal package TP_T contains only legally acquired evidence and source identifiers, excluding memory and reasoning traces. Evidence-Based Verification Given c and TP_T, the Verifier autoregressively generates label tokens: ~=gϕ(c,T),y^=parse(~). y=g_φ(c,P_T), y=parse( y). (3) The Verifier receives only the claim context and legally acquired evidence, without access to unread evidence or interaction states such as coverage, budget, or agent memory. This restricted interface encourages evidence-grounded prediction rather than reliance on procedural artifacts. The Verifier is a smaller VLM trained with LoRA and frozen during policy refinement. Training SIEVE is trained in four stages: teacher trajectory generation, agent supervised fine-tuning, verifier supervised fine-tuning, and agent policy refinement. All teacher trajectories, refreshed agent rollouts, verifier-training packages, and RL trajectories are generated exclusively from the training split. Teacher trajectory generation. We use GPT-5.5 (OpenAI 2026) as the teacher VLM and roll it out in the same evidence environment. For each (x,y)(x,y), the teacher receives oto_t and the ground-truth label y as privileged supervision and acquires evidence supporting y. The label is used only during teacher rollout and excluded from the stored observation and all Evidence Agent inputs. At each step, the teacher produces ztz_t, containing an action and, when applicable, a non-verdict memory update, with at=parse(zt)a_t=parse(z_t). A trajectory is τ=(o1,z1,a1,…,oTτ,zTτ,aTτ)τ=(o_1,z_1,a_1,…,o_T_τ,z_T_τ,a_T_τ). We sample two trajectories per video using different random seeds and decoding temperatures. Any rollout containing repeated, unavailable, or malformed actions, or explicit label leakage, is terminated and discarded; only trajectories with valid transitions and protocol-compliant actions are retained. Because the teacher is conditioned on y, terminal prediction correctness is not a filtering criterion. Agent supervised fine-tuning. The Evidence Agent is supervised on the retained teacher trajectories. Because each response contains an action decision and, when applicable, a memory update, we apply token-level supervision to the complete response: ℒSFT(θ)=−∑τ∑t=1Tτ∑ℓ=1Lτ,tlogπθ(zτ,t,ℓ∗∣oτ,t,zτ,t,<ℓ∗),L_SFT(θ)=- _τ _t=1^T_τ _ =1^L_τ,t _θ (z^*_τ,t, o_τ,t,z^*_τ,t,< ), (4) where zτ,t∗z_τ,t^* is the teacher response. This stage teaches evidence selection, memory updating, protocol following, and stopping behavior. Verifier supervised fine-tuning. We train the Verifier on terminal evidence packages from retained teacher trajectories and two refreshed SFT-agent rollouts per sample. Kimi K2.5 (Kimi Team et al. 2026) evaluates packages from both sources, and only those judged valid and consistent with the ground-truth label are retained. Let verD_ver denote the union of the retained teacher-terminal and agent-terminal packages, and let y1:Lyy_1:L_y denote the target label tokens. With next-token distribution pϕp_φ, the Verifier is optimized with ℒver(ϕ)=−∑(c,,y)∈ver∑ℓ=1Lylogpϕ(yℓ∣c,,y<ℓ).L_ver(φ)=- _(c,P,y) _ver _ =1^L_y p_φ(y_ c,P,y_< ). (5) Using both package sources reduces the distribution shift between Verifier training and subsequent agent rollouts. The trained Verifier is then frozen for policy refinement. Agent policy refinement. Finally, we refine the Evidence Agent with reinforcement learning using the frozen Verifier as the outcome evaluator. The reward combines verification outcome, evidence coverage, and interaction cost: R¯(τ) R(τ) =B(s)+0.80[s=correct]Q(τ)−P(s,τ), =B(s)+80\,I[s=correct]Q(τ)-P(s,τ), (6) R(τ) R(τ) =clip(R¯(τ),−1.2,1.0), =clip\! ( R(τ),-1.2,1.0 ), where B(s)B(s) assigns a positive base reward to correct verification and a negative reward otherwise. For correct trajectories, Q(τ)Q(τ) rewards textual novelty, visual evidence, slot–channel diversity, and evidence volume, while P(s,τ)P(s,τ) penalizes unnecessary reads, repeated or invalid actions, and forced budget termination. For incorrect trajectories, only repeated or invalid actions are penalized. The coverage bonus discourages premature stopping, especially for real videos without a single decisive contradiction. Correctness determines the primary ordering, while coverage and cost distinguish trajectories within each outcome branch. Online rollout filtering. During GRPO refinement, all trajectories sampled for the same input are first scored by the frozen Verifier-based reward pipeline. We retain only groups with informative within-group variation according to correctness-contrast and reward-spread criteria, while degenerate groups are skipped. This filtering does not modify the reward of any individual trajectory. For each x, we sample G trajectories τ(j)j=1G\τ^(j)\_j=1^G and compute A^(j)=R(τ(j))−meankR(τ(k))stdkR(τ(k)). A^(j)= R(τ^(j))-mean_kR(τ^(k))std_kR(τ^(k)). (7) The Evidence Agent generates a response zt(j)=(zt,1(j),…,zt,Lt(j)(j)),z_t^(j)= (z_t,1^(j),…,z_t,L_t^(j)^(j) ), (8) at each interaction step, containing an action command and, when applicable, a memory update. We therefore define the policy ratio at the token level: ρt,ℓ(j)=πθ(zt,ℓ(j)∣ot(j),zt,<ℓ(j))πold(zt,ℓ(j)∣ot(j),zt,<ℓ(j)). _t, ^(j)= _θ\! (z_t, ^(j) o_t^(j),z_t,< ^(j) ) _old\! (z_t, ^(j) o_t^(j),z_t,< ^(j) ). (9) The asymmetric dual-clipped surrogate is ρ¯t,ℓ(j) ρ_t, ^(j) =clip(ρt,ℓ(j),1−ϵcliplow,1+ϵcliphigh), =clip\! ( _t, ^(j),1- _clip^low,1+ _clip^high ), (10) ℓ~t,ℓ(j) _t, ^(j) =min(ρt,ℓ(j)A^(j),ρ¯t,ℓ(j)A^(j)), = \! ( _t, ^(j) A^(j), ρ_t, ^(j) A^(j) ), ℓt,ℓ(j) _t, ^(j) =ℓ~t,ℓ(j),A^(j)≥0,max(ℓ~t,ℓ(j),CdualA^(j)),A^(j)<0. = The reported runs disable KL regularization; the clipping values are given in the supplementary material. The final objective is GRPO(θ)=[1G∑j=1G1Zj∑t=1Tj∑ℓ=1Lt(j)ℓt,ℓ(j)],J_GRPO(θ)=E [ 1G _j=1^G 1Z_j _t=1^T_j _ =1^L_t^(j) _t, ^(j) ], (11) where Zj=∑t=1TjLt(j)Z_j= _t=1^T_jL_t^(j) is the generated-token count in τ(j)τ^(j). The same trajectory-level advantage is assigned to all tokens in a rollout, while token-count normalization prevents longer trajectories from dominating the update. We maximize GRPOJ_GRPO to refine the policy. Inference At inference, SIEVE alternates between the Evidence Agent and environment. The environment initializes the evidence space, coverage, history, memory, and budget, and provides a one-time slot overview. The agent then issues iterative READREAD actions or STOPSTOP. Upon STOPSTOP, budget exhaustion, or an invalid action, the frozen Verifier predicts y^=parse!(gϕ(c,T)). y=parse! (g_φ(c,P_T) ). (12) This yields a binary Fake/Real label and an interpretable trail of the acquired slots and channels. Table 1: Performance comparison on FakeSV and FakeTT. Best results are shown in bold, and second-best results are underlined. SIEVE results are averaged over three independent runs. Results marked with * are taken from FakeSV-VLM (Wang et al. 2025b). Paradigm Method -F1 -P -R -F1 -P -R MLLM-based GPT-4o-mini* (OpenAI 2024a) 68.08 68.05 69.88 69.49 61.54 61.20 64.41 65.89 GPT-4.1-mini* (OpenAI 2025) 70.30 70.25 70.61 70.87 49.16 48.54 62.50 59.70 Qwen2.5-VL* (Bai et al. 2025) 64.21 60.79 64.55 61.52 45.82 45.31 56.69 55.42 InternVL2.5* (Chen et al. 2024) 64.39 57.89 68.52 60.50 46.82 45.29 64.92 59.23 InternVL2.5-MPO* (Wang et al. 2024) 65.13 61.07 66.46 62.12 43.14 40.84 61.90 56.23 Unimodal ViT* (Dosovitskiy et al. 2021) 70.85 70.66 70.64 70.91 64.88 62.59 62.54 63.80 BERT* (Devlin et al. 2019) 78.41 78.25 78.17 78.52 70.90 69.00 68.71 70.60 Multimodal TikTec* (Shang et al. 2021) 73.06 72.79 72.73 72.93 66.56 65.55 66.50 68.62 FANVM* (Choi and Ko 2021) 79.88 78.91 80.98 78.42 71.91 70.85 71.21 73.90 SV-FEND* (Qi et al. 2023a) 80.81 80.19 81.08 79.84 77.26 75.55 74.94 77.13 FakingRecipe* (Bu et al. 2024) 84.69 84.39 84.57 84.25 79.26 77.53 76.86 78.89 Reasoning/ VLM-enhanced CA-FVD* (Wang et al. 2025a) 85.79 85.28 86.57 84.78 81.61 80.26 79.50 82.17 ExMRD* (Hong et al. 2025) 86.90 86.52 87.31 86.13 84.28 83.13 82.27 85.19 FakeSV-VLM* (Wang et al. 2025b) 90.22 89.97 90.55 89.64 89.30 87.98 87.80 88.17 92.62 92.41 93.25 91.96 91.64 90.72 90.12 91.45 Experiments Datasets. We evaluate SIEVE on FakeSV and FakeTT, with out-of-domain testing on FakeVV. FakeSV is a Chinese benchmark from Douyin and Kuaishou (Qi et al. 2023a), while FakeTT is an English TikTok benchmark introduced by FakingRecipe (Bu et al. 2024). Both provide Event and Description fields, and FakeSV additionally includes comments. SIEVE uses only Description as the initial claim context and excludes all other fields. Following prior work, we adopt a temporal 70%/15%/15% train/validation/test split. For out-of-domain evaluation, the FakeTT-trained checkpoint is directly applied to FakeVV (Zhang et al. 2025), where each query pairs a source video with an authentic or entity-manipulated title. SIEVE uses the title as the claim context and retrieves evidence only from the paired video, without using any FakeVV data for training, validation, model selection, or prompt adaptation. Baselines. For FakeSV and FakeTT, we compare SIEVE with four baseline groups: MLLMs, including GPT-4o-mini, GPT-4.1-mini, Qwen2.5-VL, InternVL2.5, and InternVL2.5-MPO (OpenAI 2024a, 2025; Bai et al. 2025; Chen et al. 2024; Wang et al. 2024); unimodal models, ViT and BERT (Dosovitskiy et al. 2021; Devlin et al. 2019); task-specific multimodal methods, TikTec, FANVM, SV-FEND, and FakingRecipe (Shang et al. 2021; Choi and Ko 2021; Qi et al. 2023a; Bu et al. 2024); and reasoning- or VLM-enhanced methods, CA-FVD, ExMRD, and FakeSV-VLM (Wang et al. 2025a; Hong et al. 2025; Wang et al. 2025b). All in-domain baseline results are taken from FakeSV-VLM (Wang et al. 2025b). For FakeVV, we compare against GPT-4o, DeepSeek-R1, Qwen2.5-VL-7B, QVQ-72B-Preview, InternVL2.5-8B, Qwen3-VL-8B-Instruct, and GPT-5.5 (OpenAI 2024b; Guo et al. 2025; Bai et al. 2025; Qwen Team 2024; Chen et al. 2024; Bai and others 2025; OpenAI 2026). Results for the first five are reported by Fact-R1 (Zhang et al. 2025), whereas the last two are evaluated by us under the same protocol. Fact-R1 is excluded because its reported model is trained on FakeVV, violating our no-FakeVV-training setting. Training Details. The agent and verifier use Qwen3-VL-8B-Instruct and Qwen3-VL-2B-Instruct (Bai and others 2025), respectively, with LoRA adapters. The agent undergoes one epoch of SFT on teacher trajectories, followed by 100 GRPO steps with group size n=8n=8, a read budget of 24, and a 32K-token context window. The verifier is trained for two epochs with up to 24 images per example. Learning rates are 1×10−41× 10^-4 for agent SFT and verifier training and 5×10−75× 10^-7 for RL, with per-device batch size 1 and gradient accumulation 4. We report the mean over three independent runs with seeds 42, 2027, and 3407, using a fixed test-time decoding temperature of 0.7. All experiments use 8 NVIDIA A800 80GB GPUs with bfloat16 precision. Main Results Table 1 reports the main results on FakeSV and FakeTT. SIEVE achieves the best performance across all metrics on both benchmarks. On FakeSV, it obtains 92.62% ACC and 92.41% Macro-F1, outperforming the strongest prior method, FakeSV-VLM, by 2.40 and 2.44 points, respectively. On FakeTT, SIEVE improves ACC from 89.30% to 91.64% and Macro-F1 from 87.98% to 90.72%, yielding gains of 2.34 and 2.74 points. Several trends can be observed. Under the evaluated settings, general-purpose MLLMs perform worse than task-specific methods, indicating that direct prompting alone remains challenging for fine-grained video misinformation detection. Stronger multimodal detectors generally outperform unimodal baselines, although early multimodal fusion methods do not always surpass a strong text-only model, suggesting that adding modalities does not by itself guarantee better performance. Reasoning- and VLM-enhanced methods further improve performance, yet they still mainly operate on preconstructed global inputs. In contrast, SIEVE explicitly separates evidence acquisition from veracity judgment and exposes the verifier only to selected OCR, ASR, and visual evidence. The consistent gains on both Chinese and English benchmarks demonstrate the effectiveness of the proposed acquisition-and-verification framework under the evaluated settings. Together with the evidence-budget analysis, these results indicate that compact, claim-relevant evidence can support strong verification performance without exposing all constructed evidence to the verifier. Table 2: Ablation study on FakeSV and FakeTT. Exhaustive evidence trains a separate verifier on all OCR/ASR evidence and up to 24 uniformly sampled images; all other variants share a fixed verifier. Ablation Type Setting -F1 -F1 Evidence Input Description-only 77.12 73.92 66.22 65.72 Same-video random 61.99 49.20 55.85 55.43 Exhaustive evidence 92.25 91.97 89.63 87.80 Agent Training Base agent 88.56 87.91 84.95 84.23 Agent SFT only 91.51 91.13 89.97 89.13 w/o rollout filtering 91.70 91.33 90.97 89.98 w/o evidence memory 91.88 91.63 90.97 89.78 Model SIEVE 92.62 92.41 91.64 90.72 Ablation Study To assess each component, we evaluate seven variants: (1) Description-only, using only the Description field; (2) Same-video random evidence, replacing the selected package with an equally sized random package from the same video; (3) Exhaustive evidence, training and evaluating a separate Verifier with all constructed OCR and ASR evidence and up to 24 images uniformly sampled from the full visual evidence space; (4) Base agent, removing agent-side SFT and RL; (5) Agent SFT only, removing RL; (6) w/o rollout filtering, disabling group-level filtering during RL; and (7) w/o evidence memory, removing cross-step evidence memory. Except for Exhaustive evidence, the Verifier is fixed across all variants. As shown in Table 2, Description-only confirms the complementary value of acquired evidence, while same-video random evidence shows that indiscriminate selection is insufficient. Despite being separately trained and evaluated with comprehensive textual evidence and uniformly sampled visual evidence, Exhaustive evidence still underperforms Full SIEVE on both datasets, particularly by 2.92 M-F1 points on FakeTT. This indicates that broad evidence coverage alone cannot replace targeted acquisition: redundant or weakly informative content may dilute decisive clues, whereas SIEVE constructs a compact and verification-effective evidence package. Agent SFT substantially improves over the base agent, validating trajectory-level supervision. Removing rollout filtering or evidence memory further degrades performance, supporting their roles in policy refinement and coherent acquisition. Overall, Full SIEVE performs best on both datasets, demonstrating the joint contribution of targeted acquisition and agent training. Can Sparse Evidence Suffice? Figure 3: Evidence sufficiency on FakeSV and FakeTT. (a) Stable correction of Description-only errors. (b) Macro-F1 under different inference-time evidence budgets. Figure 3 investigates whether reliable verification can be achieved from only a small subset of the available evidence. In Figure 3(a), we focus on samples misclassified under the Description-only setting and evaluate the frozen Verifier after each legal read. A sample is counted at step k only if its prediction remains correct for all subsequent evidence prefixes. Within four reads, 53.61% and 55.45% of the initial errors are stably corrected on FakeSV and FakeTT, respectively. The rates reach 60.55% and 67.08% within eight reads, accounting for approximately 97% of all eventually corrected cases on both datasets. This shows that, for a substantial fraction of samples requiring video evidence, only a few acquired clues are sufficient to establish a stable correct prediction. Figure 3(b) examines whether such compact evidence also preserves overall performance. Increasing the inference-time cap from 2 to 12 reads improves Macro-F1 from 82.90% to 90.55% on FakeSV and from 72.60% to 90.10% on FakeTT, whereas extending the cap from 12 to 24 yields only 1.86 and 0.62 additional points. Together, the rapid stable correction and early performance saturation provide direct empirical support for our central hypothesis: sparse, claim-relevant evidence can suffice for reliable video misinformation verification under the considered evidence space and Verifier interface. Figure 4: Evidence acquisition statistics on the test sets. The left panel shows the distribution of legal reads across evidence channels, and the right reports average total and visual reads. Evidence Acquisition Analysis Figure 4 summarizes the evidence acquisition behavior of SIEVE under the 24-read inference cap. The agent uses fewer than eight legal read actions per instance on average: 7.76 on FakeSV and 7.89 on FakeTT. Under the current action and Verifier interface, it therefore constructs a compact evidence package rather than reading every available slot–channel pair. Textual evidence is selected most frequently. On FakeSV, OCR accounts for 42.27% of legal reads and ASR for 23.07%. On FakeTT, OCR and ASR are more balanced at 34.35% and 32.60%, respectively. This pattern suggests that the learned policy adapts its channel preference to dataset-specific evidence distributions. Visual evidence is used more selectively. The agent averages 2.69 and 2.61 visual reads on FakeSV and FakeTT, respectively, and selects KEY frames more often than CTX frames. This selection pattern is consistent with the learned policy favoring visually distinctive KEY frames over generic contextual CTX frames. Overall, the results show that strong verification performance is obtained from compact, complementary textual and visual evidence packages within the considered evidence budget. Out-of-Domain Generalization Table 3 evaluates cross-dataset generalization by directly applying the FakeTT-trained SIEVE checkpoint to FakeVV, without using FakeVV for training, validation, model selection, or prompt adaptation. SIEVE achieves 73.20% ACC and 74.91% F1, outperforming all evaluated baselines under this no-FakeVV-training protocol. Its recall of 80.00% and precision of 70.42% indicate that SIEVE identifies most fake cases across domains, although this relatively high sensitivity is accompanied by some false-positive predictions. We exclude Fact-R1 itself because its reported model is trained on FakeVV and is therefore not directly comparable in this out-of-domain setting. Table 3: Out-of-domain performance on FakeVV under the no-FakeVV-training setting. Fake is the positive class. Best results are in bold; † marks results reported by Fact-R1 (Zhang et al. 2025). TableHeaderTopMethod 1 GPT-4o† (OpenAI 2024b) 56.0 60.4 35.0 44.3 DeepSeek-R1† (Guo et al. 2025) 53.5 58.1 25.2 35.1 Qwen2.5-VL-7B† (Bai et al. 2025) 52.9 51.1 51.1 51.1 QVQ-72B-Preview† (Qwen Team 2024) 53.5 52.6 52.6 52.6 InternVL2.5-8B† (Chen et al. 2024) 53.5 58.5 24.0 34.0 Qwen3-VL-8B-Instruct (Bai and others 2025) 67.3 66.3 72.5 69.2 GPT-5.5 (OpenAI 2026) 67.8 66.6 73.3 69.8 TableOursSIEVE (FakeTT ckpt.) 73.2 70.4 80.0 74.9 References S. Bai, K. Chen, X. Liu, J. Wang, W. Ge, S. Song, et al. (2025) Qwen2.5-vl technical report. arXiv preprint arXiv:2502.13923. External Links: Document Cited by: Table 1, Table 3, Experiments. S. Bai et al. (2025) Qwen3-vl technical report. External Links: 2511.21631, Link Cited by: Table 3, Experiments, Experiments. Y. Bu, Q. Sheng, J. Cao, P. Qi, D. Wang, and J. Li (2023) Combating online misinformation videos: characterization, detection, and future directions. In Proceedings of the 31st ACM International Conference on Multimedia, p. 8770–8780. External Links: Document Cited by: Introduction, Introduction. Y. Bu, Q. Sheng, J. Cao, P. Qi, D. Wang, and J. Li (2024) FakingRecipe: detecting fake news on short video platforms from the perspective of creative process. In Proceedings of the 32nd ACM International Conference on Multimedia, p. 1351–1360. External Links: Document Cited by: Introduction, Related Work, Table 1, Experiments, Experiments. Z. Chen, W. Wang, Y. Cao, Y. Liu, Z. Gao, E. Cui, et al. (2024) Expanding performance boundaries of open-source multimodal models with model, data, and test-time scaling. arXiv preprint arXiv:2412.05271. External Links: Document Cited by: Table 1, Table 3, Experiments. H. Choi and Y. Ko (2021) Using topic modeling and adversarial neural networks for fake news video detection. In Proceedings of the 30th ACM International Conference on Information & Knowledge Management, p. 2950–2954. External Links: Document Cited by: Introduction, Related Work, Table 1, Experiments. J. Devlin, M. Chang, K. Lee, and K. Toutanova (2019) BERT: pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, p. 4171–4186. External Links: Document Cited by: Table 1, Experiments. A. Dosovitskiy, L. Beyer, A. Kolesnikov, D. Weissenborn, X. Zhai, T. Unterthiner, M. Dehghani, M. Minderer, G. Heigold, S. Gelly, J. Uszkoreit, and N. Houlsby (2021) An image is worth 16x16 words: transformers for image recognition at scale. In International Conference on Learning Representations, Cited by: Table 1, Experiments. Y. Du, C. Li, R. Guo, X. Yin, W. Liu, J. Zhou, Y. Bai, Z. Yu, Y. Yang, Q. Dang, and H. Wang (2020) P-ocr: a practical ultra lightweight ocr system. arXiv preprint arXiv:2009.09941. Cited by: Appendix B, Evidence Space Construction. D. Guo, D. Yang, H. Zhang, J. Song, P. Wang, Q. Zhu, R. Xu, R. Zhang, S. Ma, X. Bi, X. Zhang, X. Yu, Y. Wu, et al. (2025) DeepSeek-R1 incentivizes reasoning in LLMs through reinforcement learning. Nature 645 (8081), p. 633–638. External Links: Document Cited by: Table 3, Experiments. R. Hong, J. Lang, J. Xu, Z. Cheng, T. Zhong, and F. Zhou (2025) Following clues, approaching the truth: explainable micro-video rumor detection via Chain-of-Thought reasoning. In Proceedings of the ACM on Web Conference 2025, p. 4684–4698. External Links: Document Cited by: Introduction, Related Work, Table 1, Experiments. Kimi Team, T. Bai, et al. (2026) Kimi k2.5: visual agentic intelligence. arXiv preprint arXiv:2602.02276. External Links: Document, Link Cited by: Verifier supervised fine-tuning.. J. Lang, R. Hong, T. Zhong, Y. Wang, and F. Zhou (2026) Nip rumors in the bud: retrieval-guided topic-level adaptation for test-time fake news video detection. In Proceedings of the 32nd ACM SIGKDD Conference on Knowledge Discovery and Data Mining, p. 600–611. External Links: Document Cited by: Introduction, Related Work. D. M. J. Lazer, M. A. Baum, Y. Benkler, A. J. Berinsky, K. M. Greenhill, F. Menczer, M. J. Metzger, B. Nyhan, G. Pennycook, D. Rothschild, M. Schudson, S. A. Sloman, C. R. Sunstein, E. A. Thorson, D. J. Watts, and J. L. Zittrain (2018) The science of fake news. Science 359 (6380), p. 1094–1096. External Links: Document Cited by: Introduction. M. Li, Y. Zhang, H. Xu, X. Li, C. Gao, and Z. Wang (2025) Learning complex heterogeneous multimodal fake news via social latent network inference. In Proceedings of the AAAI Conference on Artificial Intelligence, Vol. 39, p. 433–441. External Links: Document Cited by: Introduction, Related Work. Z. Li, H. Yu, H. Jiang, Q. Sheng, Y. Xu, B. Bi, Y. Li, Z. Yuan, Y. Cai, and Z. Wang (2026) FactGuard: agentic video misinformation detection via reinforcement learning. External Links: 2602.22963 Cited by: Introduction, Related Work. F. Liu, Y. Yacoob, and A. Shrivastava (2023) COVID-VTS: fact extraction and verification on short video platforms. In Proceedings of the 17th Conference of the European Chapter of the Association for Computational Linguistics, p. 178–188. External Links: Document Cited by: Introduction, Related Work. OpenAI (2024a) GPT-4o mini: advancing cost-efficient intelligence. Note: https://openai.com/index/gpt-4o-mini-advancing-cost-efficient-intelligence/Official release post, accessed 2026-07-05 Cited by: Table 1, Experiments. OpenAI (2024b) GPT-4o system card. External Links: 2410.21276 Cited by: Table 3, Experiments. OpenAI (2025) Introducing gpt-4.1 in the api. Note: https://openai.com/index/gpt-4-1/Official release post, accessed 2026-07-05 Cited by: Table 1, Experiments. OpenAI (2026) Introducing gpt-5.5. Note: https://openai.com/index/introducing-gpt-5-5/Accessed: 2026-07-09 Cited by: Teacher trajectory generation., Table 3, Experiments. M. Oquab, T. Darcet, T. Moutakanni, H. Vo, M. Szafraniec, V. Khalidov, P. Fernandez, D. Haziza, F. Massa, A. El-Nouby, M. Assran, N. Ballas, W. Galuba, R. Howes, P. Huang, S. Li, I. Misra, M. Rabbat, V. Sharma, G. Synnaeve, H. Xu, H. Jégou, J. Mairal, P. Labatut, A. Joulin, and P. Bojanowski (2024) DINOv2: learning robust visual features without supervision. Transactions on Machine Learning Research. Cited by: Appendix B, Evidence Space Construction. P. Qi, Y. Bu, J. Cao, W. Ji, R. Shui, J. Xiao, D. Wang, and T. Chua (2023a) FakeSV: a multimodal benchmark with rich social context for fake news detection on short video platforms. In Proceedings of the AAAI Conference on Artificial Intelligence, Vol. 37, p. 14444–14452. External Links: Document Cited by: Introduction, Related Work, Table 1, Experiments, Experiments. P. Qi, Y. Zhao, Y. Shen, W. Ji, J. Cao, and T. Chua (2023b) Two heads are better than one: improving fake news video detection by correlating with neighbors. In Findings of the Association for Computational Linguistics: ACL 2023, p. 11947–11959. External Links: Document Cited by: Introduction, Related Work. Qwen Team (2024) QVQ: to see the world with wisdom. Note: Accessed: 2026-07-13 External Links: Link Cited by: Table 3, Experiments. A. Radford, J. W. Kim, T. Xu, G. Brockman, C. McLeavey, and I. Sutskever (2023) Robust speech recognition via large-scale weak supervision. In Proceedings of the 40th International Conference on Machine Learning, Proceedings of Machine Learning Research, Vol. 202, p. 28492–28518. Cited by: Appendix B, Evidence Space Construction. L. Shang, Z. Kou, Y. Zhang, and D. Wang (2021) A multimodal misinformation detector for COVID-19 short videos on TikTok. In 2021 IEEE International Conference on Big Data, p. 899–908. External Links: Document Cited by: Introduction, Related Work, Table 1, Experiments. T. Souček and J. Lokoč (2020) TransNet v2: an effective deep network architecture for fast shot transition detection. arXiv preprint arXiv:2008.04838. Cited by: Appendix B, Evidence Space Construction. J. Thorne, A. Vlachos, C. Christodoulopoulos, and A. Mittal (2018) FEVER: a large-scale dataset for fact extraction and VERification. In Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, p. 809–819. External Links: Document Cited by: Related Work. S. Venkatagiri, J. S. Schafer, and S. Prochaska (2023) The challenges of studying misinformation on video-sharing platforms during crises and mass-convergence events. External Links: 2303.14309, Document Cited by: Introduction. J. Wang, J. Liu, N. Zhang, and Y. Wang (2025a) Consistency-aware fake videos detection on short video platforms. In Advanced Intelligent Computing Technology and Applications, Lecture Notes in Computer Science, Vol. 15859, p. 200–210. Note: ICIC 2025 External Links: Document Cited by: Table 1, Experiments. J. Wang, Y. Wang, L. Cheng, and Z. Zhong (2025b) FakeSV-VLM: taming VLM for detecting fake short-video news via progressive mixture-of-experts adapter. In Findings of the Association for Computational Linguistics: EMNLP 2025, p. 4782–4798. External Links: Document Cited by: Introduction, Introduction, Related Work, Table 1, Table 1, Experiments. W. Wang, Z. Chen, W. Wang, Y. Cao, Y. Liu, Z. Gao, J. Zhu, X. Zhu, L. Lu, Y. Qiao, and J. Dai (2024) Enhancing the reasoning ability of multimodal large language models via mixed preference optimization. arXiv preprint arXiv:2411.10442. External Links: Document Cited by: Table 1, Experiments. Y. Wang, Y. Yang, and H. Zhang (2026) CSI: an investigative multi-agent framework for explainable short video fake news detection. In Findings of the Association for Computational Linguistics: ACL 2026, p. 25508–25528. External Links: Document Cited by: Introduction, Introduction, Related Work. Z. Wu, C. Xiong, C. Ma, R. Socher, and L. S. Davis (2019) AdaFrame: adaptive frame selection for fast video recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), Cited by: Related Work. S. Yao, J. Zhao, D. Yu, N. Du, I. Shafran, K. Narasimhan, and Y. Cao (2023) ReAct: synergizing reasoning and acting in language models. In The Eleventh International Conference on Learning Representations, Cited by: Introduction, Related Work. F. Zhang, D. Li, Q. Zhang, J. Chen, G. Liu, J. Lin, J. Yan, J. Liu, and Z. Zha (2025) Fact-r1: towards explainable video misinformation detection with deep reasoning. In Advances in Neural Information Processing Systems, Vol. 38. Cited by: Introduction, Related Work, Table 3, Experiments, Experiments. H. Zhang, R. Zhang, J. Guo, M. de Rijke, Y. Fan, and X. Cheng (2023) From relevance to utility: evidence retrieval with feedback for fact verification. In Findings of the Association for Computational Linguistics: EMNLP 2023, p. 6373–6384. External Links: Document Cited by: Introduction, Related Work. X. Zheng, Z. Zeng, H. Wang, Y. Bai, Y. Liu, and M. Luo (2025) From predictions to analyses: rationale-augmented fake news detection with large vision-language models. In Proceedings of the ACM on Web Conference 2025, p. 5364–5375. External Links: Document Cited by: Introduction, Introduction, Related Work. Appendix A Dataset Splits and Statistics Table 4 summarizes the datasets used in our experiments. FakeSV and FakeTT are chronologically partitioned into training, validation, and test sets, while FakeVV is used only for out-of-domain evaluation with its balanced official test set. The class distributions of fake and real samples are also reported for completeness. Table 4: Dataset statistics used in our experiments. FakeSV and FakeTT are chronologically split into training, validation, and test sets with an approximately 70%/15%/15% ratio. FakeVV contains 102,000 video–text pairs in its official split; only its balanced test set of 2,000 query samples, associated with 997 unique source videos, is used for out-of-domain evaluation. F/R denotes the numbers of fake and real samples, respectively. Dataset Total Overall F/R Train Val. Test Test F/R FakeSV 3,624 1,810/1,814 2,536 546 542 304/238 FakeTT 1,991 1,172/819 1,393 299 299 200/99 FakeVV 102,000 51,000/51,000 100,000 – 2,000 1,000/1,000 Appendix B Evidence Space Construction: Implementation Details Shot detection and slot refinement. Shot boundaries are detected with TransNetV2 (Souček and Lokoč 2020). Duration-based refinement then proceeds in two passes. First, shots shorter than 0.80.8 s are merged into the temporally shorter of their two neighbors. Second, shots longer than 66 s are subdivided using motion-equalized splitting based on frame differences, with uniform-duration splitting used when motion estimates are unavailable. Depending on duration, a long shot is divided into two segments up to 1212 s, three segments up to 1818 s, and four segments beyond 1818 s. If the number of refined slots still exceeds Nmax=10N_ =10, we iteratively merge the adjacent pair with the lowest motion density, defined as the combined frame-difference motion divided by the combined duration. This preferentially merges low-information static segments while preserving short but dynamic slots. When per-frame motion cannot be computed, the shortest adjacent pair is merged instead. OCR channel. On-screen text is extracted with PaddleOCR (Du et al. 2020) at 2 FPS, using a minimum box duration of 0.50.5 s and a minimum area ratio of 1.5%1.5\%. Deduplicated OCR items within each slot are ranked by imp=0.25d^+0.30a^+0.20c+0.15s^+0.10ℓ^,imp=0.25 d+0.30 a+0.20c+0.15 s+0.10 , where d^=min(d/5 s,1) d= (d/5 s,1), a^=min(a/0.1,1) a= (a/0.1,1), c is the mean OCR confidence, s s is the temporal-stability score of the text box across consecutive detections, and ℓ^=min(ℓ/20,1) = ( /20,1) for character length ℓ . The OCR channel exposes the deduplicated text items in descending importance order. ASR channel. Full-video transcripts are produced by Whisper large-v3 (Radford et al. 2023) with word-level timestamps. Each transcript segment is matched to a slot using a 1-s context-extended window [tstart−1,tend+1][t_start-1,t_end+1]. A segment is designated primary when its midpoint lies within [tstart,tend][t_start,t_end] and boundary when it overlaps only the extended window. The ASR channel exposes only the concatenated primary text, assigning each segment to exactly one slot; boundary context is retained as metadata but is not shown to the agent. Visual channels: CTX and KEY. Candidate frames are encoded with DINOv2 ViT-S/14 (Oquab et al. 2024) using the 384-dimensional, ℓ2 _2-normalized CLS token, such that cosine similarity reduces to a dot product. CTX contains three representative frames: the slot is divided into three equal-duration bins, and one frame per bin is selected by jointly considering sharpness, measured by Laplacian variance, and dissimilarity to previously selected CTX frames. KEY contains up to two complementary frames: • key_diff is the candidate with the lowest maximum cosine similarity to the three CTX frames, with sharpness used as a tie-breaker. If all candidates have similarity above 0.970.97, the least-similar candidate is retained as a fallback. • key_ocr is the sharpest frame within the temporal window of the OCR item with the highest impimp score. If key_diff and key_ocr coincide, or if one is unavailable, KEY contains a single frame. Coverage matrix and legal actions. The coverage matrix t∈−,,||×|ℳ|C_t∈\ -, x, v\^|S|×|M| is initialized to x for non-empty channels and − - otherwise; v denotes a read channel. Let ntimgn_t^img be the number of accumulated CTX/KEY images and κi,m _i,m the images returned by a READ, with κi,m=0 _i,m=0 for OCR and ASR. With Bimg=24B_img=24, the legal read set is tread=READ(slotsid,channel)|t(sid,channel)=,btread>0,ntimg+κsid,channel≤Bimg.A^read_t= \READ(slot_sid,\channel\) | subarraycC_t(sid,channel)= x,\ b^read_t>0,\\ n_t^img+ _sid,channel≤ B_img subarray \. The visual READs excluded by this constraint are explicitly listed in <budget_state> for the agent. Each legal READ consumes one unit of the overall read budget and changes its coverage status from x to v. Repeated, unavailable, or malformed actions immediately terminate the current acquisition rollout, are excluded from the evidence package, and receive the behavioral penalty used during reinforcement learning. Appendix C Training and Optimization Details Reward Design Details The trajectory reward combines verification outcome, evidence coverage, and interaction cost. Let s(τ)s(τ) denote the terminal status of trajectory τ. As introduced in the main text, the reward is R¯(τ) R(τ) =B(s)+0.80[s=correct]Q(τ)−P(s,τ), =B(s)+80\,I[s=correct]Q(τ)-P(s,τ), (13) R(τ) R(τ) =clip(R¯(τ),−1.2,1.0). =clip\! ( R(τ),-1.2,1.0 ). Verification outcome. The base reward depends on the terminal verification outcome: B(s)=0.20,s=correct,−1.00,s=incorrect,−1.00,s=novalidevidence,−1.00,s=malformed.B(s)= cases0.20,&s=correct,\\ -1.00,&s=incorrect,\\ -1.00,&s=no\ valid\ evidence,\\ -1.00,&s=malformed. cases (14) Here, no valid evidence denotes termination without any legally acquired evidence, while malformed denotes a terminal Verifier output that cannot be parsed into a valid binary prediction. Evidence coverage. For a correct trajectory, the evidence-coverage score is Q(τ)= Q(τ)= 0.45qtext+0.25qvis+0.20qdiv+0.10qvol, 45q_text+25q_vis+20q_div+10q_vol, (15) qtext= q_text= clip(nnovel180,0,1), ( n_novel180,0,1 ), qvis= q_vis= clip(max(nv2,nimg4),0,1), ( ( n_v2, n_img4 ),0,1 ), qdiv= q_div= clip(0.5nslot2+0.5nch2,0,1), (0.5 n_slot2+0.5 n_ch2,0,1 ), qvol= q_vol= clip(nr3,0,1). ( n_r3,0,1 ). Here, nnoveln_novel is the number of OCR or ASR characters not already present in the initial claim context; nvn_v and nimgn_img denote the numbers of visual reads and acquired images; nslotn_slot and nchn_ch denote the numbers of distinct slots and evidence channels covered; and nrn_r is the number of valid reads. The coverage bonus is applied only to correct trajectories and discourages premature stopping. Interaction penalties. We define the efficiency, behavioral, and forced-termination penalties as Peff P_eff =min(0.15, 0.01nr+0.005nv), = (0.15,\,0.01n_r+0.005n_v ), (16) Pbeh P_beh =min(0.30, 0.15ninv)+min(0.20, 0.08nrep), = (0.30,\,0.15n_inv )+ (0.20,\,0.08n_rep ), Pstop P_stop =0.03[forced termination at budget exhaustion], =03\,I[forced termination at budget exhaustion], where ninvn_inv counts unavailable or malformed actions, excluding repeated reads, while nrepn_rep counts repeated read attempts. The two action categories are mutually exclusive. The branch-dependent penalty is P(s,τ)=Peff+Pbeh+Pstop,s=correct,Pbeh,s=incorrect,min(0.30,0.15ninv),s=novalidevidence,0,s=malformed.P(s,τ)= casesP_eff+P_beh+P_stop,&s=correct,\\ P_beh,&s=incorrect,\\ (0.30,0.15n_inv),&s=no\ valid\ evidence,\\ 0,&s=malformed. cases (17) This design makes terminal correctness the primary reward signal. Correct trajectories receive additional credit for acquiring complementary textual and visual evidence, while excessive reads, repeated actions, invalid actions, and budget-exhausted termination are discouraged. All final rewards are clipped to [−1.2,1.0][-1.2,1.0] for stable policy optimization. GRPO Optimization Details For completeness, we provide the sampled-token KL estimator supported by the GRPO implementation. Let ut,ℓ(j)=(ot(j),zt,<ℓ(j))u_t, ^(j)=(o_t^(j),z_t,< ^(j)). We define Δt,ℓ(j) _t, ^(j) =logπθ(zt,ℓ(j)∣ut,ℓ(j))−logπref(zt,ℓ(j)∣ut,ℓ(j)), = _θ\! (z_t, ^(j) u_t, ^(j) )- _ref\! (z_t, ^(j) u_t, ^(j) ), (18) D^KL(j,t,ℓ) D_KL^(j,t, ) =exp(−Δt,ℓ(j))+Δt,ℓ(j)−1. = \! (- _t, ^(j) )+ _t, ^(j)-1. Here, πref _ref is the fixed reference policy. In our formal runs the KL penalty is disabled (β=0β=0, i.e. disable_kl=true, kl_coef=0.0); we retain the estimator above for completeness since it is the framework’s default mechanism and may be enabled in future ablations. The policy-ratio clipping range follows the asymmetric PPO/DAPO convention: ϵcliplow=0.20,ϵcliphigh=0.30,Cdual=3.0, _clip^low=0.20, _clip^high=0.30, C_dual=3.0, (19) where CdualC_dual is the dual-clip constant applied when the advantage is strongly negative. The advantage estimator is GRPO (group-relative, no value network): for a group of n=8n=8 rollouts sharing a prompt, A^(j)=(R(j)−mean(R))/std(R) A^(j)=(R^(j)-mean(R))/std(R). Online rollout filtering. Before computing advantages, we discard low-information groups using a hybrid criterion operating on three per-rollout reward-manager outputs: the overall reward R(j)R^(j), the binary correctness a(j)∈0,1a^(j)∈\0,1\, and the evidence-quality score q(j)=Q(τ(j))q^(j)=Q(τ^(j)) (as defined in Eq. 15). For a group g of rollouts sharing a prompt, define the correctness-contrast indicator mixed(g)= (g)= [maxja(j)>0] \! [ _ja^(j)>0 ] (20) ∧[minja(j)≤0], \ I\! [ _ja^(j)≤ 0 ], i.e. whether the group contains both correct and incorrect trajectories, and the reward-spread indicator spread(g)= (g)= [maxjR(j)−minjR(j)≥ρrange] \! [ _jR^(j)- _jR^(j)≥ _range ] (21) ∨[stdj(R(j))≥ρstd], \ I\! [std_j\! (R^(j) )≥ _std ], i.e. whether the group’s overall-reward values are sufficiently dispersed, with ρrange=0.04 _range=0.04 and ρstd=0.015 _std=0.015. Writing j⋆=argmaxjR(j)j = _jR^(j) for the top-scoring rollout in the group, define the all-correct admissibility indicator good(g)= (g)= [minja(j)>0]∧spread(g) \! [ _ja^(j)>0 ] (g) (22) ∧[q(j⋆)≥τevi] \ I\! [q^(j )≥ _evi ] ∧[maxjR(j)>0.01] \ I\! [ _jR^(j)>0.01 ] ∧[minjR(j)<0.99], \ I\! [ _jR^(j)<0.99 ], with evidence floor τevi=0.40 _evi=0.40. A group is kept iff keep(g)=mixed(g)∨good(g).keep(g)=mixed(g)\ \ good(g). (23) That is, groups with a correctness contrast are always kept regardless of reward dispersion; all-correct groups are kept only if their top-scoring trajectory has adequate evidence support and the group is not reward-degenerate; all-incorrect groups are always discarded, so that an unresolved or wrong-verifier outcome is never learned as a relative improvement over another wrong outcome. Teacher decoding. Teacher trajectories for SFT supervision are sampled at two fixed decoding temperatures, T∈0.2,0.7T∈\0.2,0.7\, with one trajectory generated per (video, temperature) pair; the two temperature pools are merged and deduplicated before SFT data construction. LoRA configuration. Both the Agent and the Verifier use LoRA with rank r=32r=32, scaling α=64α=64, and dropout 0.050.05 during both the SFT and RL stages. The RL stage is initialized from the SFT adapter and retains a dropout value of 0.050.05. LoRA is applied to q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, and down_proj. The RL actor’s LoRA adapter is initialized from the Agent SFT checkpoint specified by init_lora_path, while the Verifier’s LoRA adapter remains frozen throughout RL. Optimizer. The RL actor is optimized with AdamW, learning rate 5×10−75× 10^-7, weight decay 10−210^-2, and no warmup (lr_warmup_ratio=0.0). Agent/Verifier SFT instead uses learning rate 10−410^-4 with a cosine schedule and a 3%3\% linear warmup. Rollout uses temperature 1.01.0, top-p 1.01.0, group size n=8n=8 during training, and temperature 0.70.7/top-p 0.90.9/n=1n=1 for validation. Appendix D Reward Ablation Study Table 5 examines how different reward components affect verification performance and evidence-acquisition efficiency. Using only the terminal outcome provides limited gains over SFT, indicating that correctness supervision alone is insufficient to learn an effective evidence-seeking policy. Removing the coverage reward reduces the average number of reads to 6.94 on FakeSV and 7.03 on FakeTT, but consistently degrades ACC and Macro-F1. It also lowers Real-class recall relative to the full reward, suggesting that coverage guidance helps prevent premature stopping when no single decisive contradiction is available. In contrast, removing interaction penalties substantially increases the average number of reads to 10.84 and 11.21, while still underperforming the full model. This shows that unrestricted evidence acquisition introduces redundant or less informative observations rather than reliably improving verification. The complete reward achieves the best ACC and Macro-F1 on both datasets with fewer than eight reads on average. Although some ablations obtain higher Real-class recall on FakeTT, their lower balanced metrics indicate a class-specific trade-off rather than superior overall verification. Overall, the coverage reward and interaction penalties play complementary roles: the former encourages sufficient evidence acquisition, whereas the latter discourages unnecessary exploration, jointly producing a more accurate and efficient policy. Table 5: Reward-component ablation on FakeSV and FakeTT. R-Rec denotes recall for the Real class. ↑ -F1 ↑ -Rec. ↑ . Reads ↓ FakeSV SFT only 91.51 91.13 80.67 8.34 Outcome only 91.51 91.19 82.35 8.62 w/o coverage 91.88 91.60 83.61 6.94 w/o penalties 92.25 92.01 85.29 10.84 92.62 92.41 86.55 7.76 FakeTT SFT only 89.97 89.13 93.94 8.51 Outcome only 90.30 89.43 92.93 8.78 w/o coverage 90.64 89.63 89.90 7.03 w/o penalties 91.30 90.37 90.91 11.21 91.64 90.72 90.91 7.89 Appendix E Prompts The Teacher prompt consists of a system prompt and a user prompt, shown in Figures 5 and 6, respectively. The system prompt provides the ground-truth label as privileged guidance for evidence acquisition and stopping. Specifically, for Fake claims, the Teacher is encouraged to stop once decisive falsifying evidence is found, whereas for Real claims, it is encouraged to gather broader supporting evidence. The user prompt specifies the claim and the interaction format. The Agent uses the same user prompt as the Teacher but a different system prompt that removes only ground-truth label information and label-dependent stopping guidance; all action-legality and budget rules remain unchanged. It must therefore determine its search strategy and stopping point solely from the claim and the evidence acquired during interaction. Finally, the Verifier predicts the claim label based on the claim and the collected evidence, using the prompt shown in Figure 7. Figure 5: System prompt for Teacher trajectory generation. Figure 6: User prompt for Teacher trajectory generation. Figure 7: Verifier prompt.