Fine-grained questions about long videos ask how many times an action occurs, in what order events unfold, or exactly when a state changes. Answering them can require finding brief moments scattered across hours of footage. Agentic models that decide when and where to look promise to find these moments, yet this ability is hard both to measure and to learn. Sparse supporting evidence may establish that an action occurred without establishing its total count; verifying timeline-wide claims requires checking for missed occurrences and double counting. Answer-only supervision, in turn, does not explicitly supervise the search for evidence. We present NexSight, which builds verifiable tasks to both evaluate and train agentic long-video understanding. Its agentic pipeline, SightForge, accepts a task only if a verifier blind to the draft answer can reconstruct it from evidence densely reacquired from the raw video and counterexample checks fail to refute it. This yields SightLine, 310 evidence-traceable tasks from 108 long videos, evaluated through SightLoop, a shared multiscale harness supporting temporal exploration, local inspection, evidence integration, and revisiting while recording returned observations. On disjoint videos, solver trajectories on verified tasks become step-level supervision for observation decisions. The strongest model reaches only 54.52% on SightLine, yet 79.35% when given the annotated evidence. With under 6K examples, Sight-35B nearly doubles its base model's interactive accuracy (21.29% to 42.26%), outperforms the 397B model of its family, begins using explicit spatial inspection after post-training, which its base model never uses, and gains over 8 points on two external long-video benchmarks.
Figures & tables
Figure 1: NexSight overview. (1) Verifiable task construction (SightForge): dense raw-video evidence, answer-hidden verification, and counterexample checks establish reliable tasks. (2) Multiscale active evaluation (SightLine and SightLoop): a 310-task benchmark and interactive harness assess temporal exploration, local inspection, and evidence integration. (3) Agentic post-training (Sight-35B): QA supervision and separately collected solver trajectories improve interactive accuracy by 20.97 percentage points over Base and support Direct-64 transfer to Video-MME Long and LVBench (123-question subset).
Table 1: From task construction to evaluation and learning: documented capabilities of representative long-video benchmarks and agentic systems. Symbols indicate whether each work meets the stated criterion, not relative performance.
Figure 2: Six core capabilities and task distribution of SightLine. The benchmark covers unlocalized fact retrieval without timestamp cues, ordering of separated events, counting across temporal windows without duplication, identity tracking across occlusions and shot changes, state/attribute change recognition, and temporal boundary localization. The donut chart shows the task count and percentage for each capability (310 tasks in total).
Figure 3: SightForge: verification before acceptance. The verifier reconstructs an answer from the question and reacquired evidence, without seeing the draft answer or generation reasoning. Comparison, independent review, and counterexample checks gate acceptance; unresolved candidates return for additional evidence or revision, or are rejected. Solver trajectories are collected separately on training tasks.
Figure 4: Overall model performance on SightLine. Accuracy over 310 tasks across fixed-input Direct and interactive SightLoop configurations (Ledger On/Off). Each model has two overlaid bars: Direct-64/128 and Ledger On/Off. The top and rounded internal boundary mark scores, not their sum. Reference models are ordered by their highest score, followed by Sight-35B. Model sources: GPT-5.6 Sol [ 20 ] , DeepSeek V4 Flash Vision [ 6 ] , Nex-N2.5-Mini [ 19 ] , Qwen3.5-397B-A17B [ 24 ] , and Qwen3.5-35B-A3B-Base [ 23 ] .
Figure 5: Temporal selection under matched budgets. Only window selection varies across a fixed 128-frame budget. Bars show 310-task accuracy and temporal evidence recall ( Rhit ); error bars denote 95% video-cluster bootstrap confidence intervals. Random averages three seeds per question. The paired Agentic–Fixed temporal-hit difference interval is positive (+20.79 points), while the accuracy-difference interval crosses zero (+4.84 points; Appendix F ).
Figure 6: Evidence-availability diagnostic. GPT-5.6 Sol performance with actively selected observations versus supplied ground-truth (GT) evidence under the same semantic rubric over all 310 tasks. Evidence Oracle supplies annotated GT evidence artifacts in a single pass with xhigh reasoning. Agentic On/Off denotes SightLoop with Ledger On/Off.
Figure 7: SightLoop (Ledger On) capability accuracy (%) on the 310-task SightLine benchmark; row labels give task counts. Sight-35B gains most over Base in cross-window count and state change under combined post-training, complementing the temporal-localization diagnostic in Sec. 6.2 .
Model / deployment
Direct-64
Agentic
Qwen3.5-35B-A3B-Base
80 / 25.81
66 / 21.29
Qwen3.5-35B-A3B-Inst
103 / 33.23
87 / 28.06
Sight-35B (ours)
88 / 28.39
131 / 42.26
Table 2: Fixed-input and agentic performance of 35B models on SightLine. Entries report correct answers / accuracy (%) over 310 tasks; inference budgets vary across deployments.
Benchmark
Questions
35B Base
Sight-35B
Δ
Video-MME Long
900
477 / 53.00
551 / 61.22
+8.22
LVBench subset
123
55 / 44.72
65 / 52.85
+8.13
Table 3: Direct-64 cross-benchmark transfer: correct answers / accuracy (%), with gains in percentage points.
Appendix figures & tables20 assets
Supplementary material from the paper’s appendix.
Appendix
Figure A1 : SightLine benchmark overview and distributions. The benchmark contains 310 tasks from 108 long videos across six fine-grained capabilities. The figure summarizes benchmark statistics, task proportions by capability and content domain, and the distribution of reference evidence artifacts per task.
Content domain
Videos
Questions
Art Performance
22
60
Film / TV
22
58
Knowledge
21
66
Life / Vlog
22
68
Sports Competition
21
58
Total
108
310
Appendix
Table 4: Content-domain distribution of SightLine.
Table 6: Formal capability identifiers and task counts.
Figure A2 : Human review outcomes for SightForge candidate tasks. Of 370 candidates, 188 were accepted as is, 122 were retained after minor wording edits, and 60 were rejected. The resulting SightLine benchmark contains 310 tasks, corresponding to an 83.8% retention rate.
Figure A3 : SightLoop multiscale active observation. Given a question and video overview, the video agent selects temporal windows, invokes an image agent for local inspection, integrates returned evidence, and revisits observations as needed. The image agent uses the same evaluated VLM. Optional Ledger memory retains observations across turns; calls, frames, pixels, and tokens are governed by explicit budgets, while actions, observations, and timestamps are logged.
Checkpoint
Direct-64
SightLoop Ledger On
Qwen3.5-35B-A3B-Base
80 / 25.81
66 / 21.29
Sight-35B (ours)
88 / 28.39
131 / 42.26
Appendix
Table 7: 35B checkpoint scores (correct / accuracy %) over all 310 tasks.
Model / checkpoint
Ledger
Acc. (%)
Model calls
Tool calls
Observed frames
Recorded tokens
GPT-5.6 Sol
On
53.23
11.99
19.72
195.99
331,446
GPT-5.6 Sol
Off
54.52
9.64
17.82
211.20
252,835
DeepSeek V4 Flash Vision
On
49.68
11.26
14.63
149.08
268,386
DeepSeek V4 Flash Vision
Off
48.39
10.16
13.66
159.14
232,136
Qwen3.5-397B-A17B
On
31.61
16.67
14.06
274.19
506,441
Qwen3.5-397B-A17B
Off
34.84
15.08
12.16
248.36
408,774
Appendix
Table 8: SightLoop accuracy and mean recorded resources per task (310-task normalization).
Model
Ledger
Valid / 310
Semantic correct
Full-set Acc. (%)
GPT-5.6 Sol
On
310/310
165
53.23
GPT-5.6 Sol
Off
310/310
169
54.52
DeepSeek V4 Flash Vision
On
310/310
154
49.68
DeepSeek V4 Flash Vision
Off
310/310
150
48.39
Qwen3.5-397B-A17B
On
303/310
98
31.61
Qwen3.5-397B-A17B
Off
299/310
108
34.84
Appendix
Table 9: SightLoop response validity and accuracy across all 310 tasks.
Table 11: Primary observation-tool calls per recorded SightLoop trajectory ( n ). Spatial operations comprise crop, zoom, highlight, and region comparison. Totals in Table 8 also include memory, metadata, and file tools and use 310-task normalization. Image-subagent invocations are separate from child tool calls; these columns are not additive.
Profile
Model calls
Tool calls
watch_clip calls
Time (min)
B1
8
16
10
12
B2
10
22
14
18
B3
14
36
24
25
B4
14
36
24
60
B5
14
36
24
120
B6
24
60
32
70
Appendix
Table 12: Full-SightLoop evaluation budget profiles. Top: configuration limits. Bottom: counts for one selected trajectory per matched question ( n ), excluding unmatched questions and additional attempts. For 30 configuration–question entries, multiple traces match the prediction; their budget attribution depends on the selected trace. Accuracy retains all 310 questions. Training-trajectory collection and matched temporal selection use separate configurations (Appendix B ).
Figure A5 : Qualitative diagnosis of agentic long-video problem solving (GPT-5.6 Sol). Left: SightLoop timeline ticks indicate returned overview frames, and blue bars show the union of local observation windows. Neither GT evidence group (at ∼ 81 s and 172 s) is intersected by the explored windows, whereas supplying GT frames recovers the correct answer. Right: Complete GT evidence is supplied, yet the model response omits the required posture transition.
Figure A6 : Qualitative comparison of fixed vs. active evidence acquisition (GPT-5.6 Sol). Under an identical observation budget, Fixed-Window sampling misses both annotated evidence points and answers incorrectly. Active selection redirects subsequent windows toward the end-card region after observing the cooldown segment, capturing the target promotional URL ( BODiPreviews.com ) and answering correctly ( Rihit=1/2 ).
Condition
N
Valid
Unresolved
Correct
Acc. (%)
Hit (%)
Fixed
310
309
0
112
36.13
19.84
Random (seed 0)
310
307
3
102
32.90
20.38
Random (seed 1)
310
310
1
97
31.29
17.47
Random (seed 2)
310
310
2
105
33.87
19.43
Agentic
310
310
1
127
40.97
40.63
Appendix
Table 13: Per-condition output coverage and seed-level statistics across all 310 questions.
Condition
Physical calls
Unique frames
API images
Tokens
Repeat (%)
Overlap (%)
Fixed
8.14
127.95
665.96
156,160
0.04
0.00
Random (mean)
8.10
127.91
661.97
155,960
0.07
3.14
Agentic
8.07
126.81
658.43
155,326
0.93
9.00
Appendix
Table 14: Per-question mean resource accounting in the matched-budget study.
Benchmark
Questions
35B Base
Sight-35B
Δ
LVBench subset
123
37 / 30.08
46 / 37.40
+7.32
Appendix
Table 15: LVBench (123-question subset) evaluation under the interactive harness: correct answers / accuracy (%).
Figure A9 : Source-video overlap across training and evaluation sets. SightLine contains 310 new tasks from 108 training-disjoint videos, corresponding to 324 original Video-MME questions. Of 300 Video-MME Long videos, 100 overlap with the 789 training sources. The 67-video LVBench subset shares three videos (five questions) with SightLine and one video (two questions) with training. Areas are not to scale; question counts and results appear in Tables 16 and 17 .
Video sources
Questions
Base correct
Sight-35B correct
Net gain
Δ (pp)
Overlapping training sources
300
173
191
+18
+6.00
Disjoint from training sources
600
304
360
+56
+9.33
All
900
477
551
+74
+8.22
Appendix
Table 16: Video-MME Long Direct-64 performance by training-source overlap. Counts refer to correct answers on the same questions; percentage-point gains are computed from the count difference within each row.
Video sources
Questions
Base correct
Sight-35B correct
Net gain
Overlapping training sources
2
0
1
+1
Disjoint from training sources
121
55
64
+9
All
123
55
65
+10
Appendix
Table 17: LVBench subset Direct-64 performance by training-source overlap. Both models are evaluated on the same questions within each row. Net gain is the difference in the number of correct answers.
College of Computer Science, Zhejiang University of Technology, Zhejiang, China · Zhejiang Key Laboratory of Visual Information Intelligent Processing, Zhejiang, China · UC Berkeley, CA, USA +1