Skip to content

EXPLORE-Bench: Egocentric Scene Prediction with Long-Horizon Reasoning

Conference: ECCV2026
Paper: ECCV
Code: https://github.com/JackYu6/EXPLORE-Bench/
Area: Multimodal VLM
Keywords: egocentric vision, long-horizon reasoning, benchmark, scene prediction, multimodal large language model

TL;DR

The paper formalizes a new task, egocentric scene prediction with long-horizon reasoning โ€” given an initial-scene image and a long sequence of atomic actions, a model must imagine and describe the final scene after all actions are executed. It builds EXPLORE-Bench (1,157 instances, action sequences averaging 113 steps and reaching 694, with structured object/attribute/relation annotations) and evaluates 20+ models with a unified object-coverage + attribute-score + relation-score metric, finding that the best model still trails humans by 7.39 points and that naively segmenting the action sequence for stepwise reasoning loses points because models describe only what changed.

Background & Motivation

The egocentric (first-person) view couples vision and action by construction: what we see determines what we do, and what we do changes what we will see next. Driven by exactly this perception-action loop, egocentric benchmarks such as EgoTaskQA, EgoMCQ, EgoSchema, EgoPlan-Bench, and ERQA have proliferated in recent years, covering recognition, memory, understanding, and planning. Yet they share one shape โ€” a video plus one short-horizon question. Their actions are short-range, their scene scope is regional, and they offer no structured account of the final scene. The parts that touch state prediction typically ask only "what state is this one object in after some visual cue," requiring neither a global roll-out of the physical consequences of an entire action sequence nor a complete picture of what the scene finally looks like.

Why the gap persisted is half blind spot and half engineering barrier. Conceptually, "predicting a future state" was assumed to mean short-range state tracking, not a global reasoning problem that requires maintaining, across dozens or hundreds of actions, what changed and what stayed the same. Engineering-wise, long-horizon evaluation needs three things at once: atomic action sequences that fully cover one task, structured object/attribute/relation annotations of the final scene, and an evaluation protocol that turns free-form text into quantifiable scores. The last one is the deadlock โ€” with only a generated scene description and an image, coarse text similarity cannot say what the model got right or wrong; yet having humans annotate every final-scene image as an object-attribute-relation graph is too expensive to scale.

This paper's design choice is to solve all three together: it reuses the timestamped atomic-action annotations of Ego4D and Ego-Exo4D plus self-recorded videos, anchoring "initial scene โ†’ final scene" instances by the start and end frames of the main activity; it replaces "one MLLM emits the whole annotation" with a multi-step pipeline that first tags objects, then grounds them, then has an MLLM generate attributes and relations, and finally corrects the result, pushing human cost from "annotate everything" down to "verify and fix"; and it borrows the directed-scene-graph captioning metric of CompreCap to decompose a description and align it at the object, attribute, and relation levels. Core idea: formalize egocentric scene prediction with long-horizon reasoning as "initial frame + full atomic-action sequence โ†’ final scene," and turn free-form descriptions of that future scene into comparable numbers through a three-level protocol of object-level soft coverage plus LLM-scored attributes and relations.

Method

Overall Architecture

What this paper delivers is an evaluation system rather than a model, and it consists of three pieces: instances, annotations, and an evaluation protocol. EXPLORE-Bench contains 1,157 instances drawn from 1,157 real first-person videos (Ego4D, Ego-Exo4D, and self-recorded footage), with an average video duration of 358.28 seconds and a maximum of 1,524.80 seconds. Each instance holds three things: (i) an image representing the initial scene; (ii) a sequence of atomic action descriptions ranging from 11 to 694 entries and averaging 113; and (iii) structured annotations of the final-scene image covering object categories, visual attributes, and inter-object relations. The benchmark annotates 23,771 objects spanning 1,612 categories, more than 20 objects per instance on average.

How the task is operationalized deserves a note of its own. Given the initial image and the action sequence, the model must emit a natural-language description of what the scene looks like after all actions run โ€” the output format is exactly that of image captioning, and the protocol then decomposes that text and aligns it to the structured annotations. One deliberate choice is that the input is a sequence of action descriptions rather than a video. This simulates the moment when an agent has planned its actions but has not executed them, so that evaluation probes reasoning about future physical consequences instead of video sampling and long-context modeling.

Key Designs

1. Task definition and instance construction: anchoring "long-horizon + global scene" to the start and end frames of a real first-person main activity

To ask "what will the scene look like after this action sequence," one must first guarantee that the sequence really describes one complete task and that the two end frames form a legitimate causal pair. The authors filter Ego4D and Ego-Exo4D by recorded activity: cooking or bicycle repair counts as valid because it has a clear goal and terminal state, whereas dancing or chatting with someone counts as invalid because no decidable end state exists. For each selected video they take the start and end frames of the main activity as the initial- and final-scene images, then use the timestamps of those two frames to retrieve the timestamped atomic-action annotations stored in the source datasets and cut out the sequence in between; self-recorded videos are annotated by hand at a comparable granularity. Instances obtained this way satisfy "the atomic actions completely describe a long-horizon task" โ€” the completeness column of the paper's comparison table. Existing EgoTaskQA/EgoMCQ/EgoSchema/EgoPlan-Bench/ERQA all fail that column, being short-action and regional, whereas EXPLORE-Bench is the only one checking both "long actions" and "global scene type."

Human quality control runs throughout. When extracting the start and end frames, the authors require no motion blur, only the main activity occurring in between, and no drastic camera-viewpoint change across the two scenes; ambiguous instances where the action descriptions are insufficient to infer the final scene (underspecified positions, severe occlusion) are filtered out. These constraints look mundane, but they decide whether the benchmark is solvable at all: if the camera wandered across half the room between the initial and the final frame, the model would no longer be reasoning about action consequences but guessing a new view that cannot be derived from the input.

2. A scalable multi-step scene annotation pipeline: replacing "let one MLLM annotate everything" with five controllable substeps

Asking a single MLLM to read the final-scene image and emit a complete "objects + attributes + relations" annotation runs into two problems at once: object coverage is insufficient (models gravitate to salient items and drop small peripheral ones), and attribute and relation accuracy is unstable. The authors therefore split annotation into five steps. Object tagging: RAM++ (Recognize Anything Plus Model) extracts object tags from the final-scene image while spaCy extracts nouns from the atomic action descriptions; the two are merged into a union to maximize coverage, and a scheme combining an LLM-based filter with rule-based filtering removes invalid tags such as synonyms and human body parts. Object grounding: the final-scene image and the filtered tags go into Grounding DINO to obtain bounding boxes, and each detected instance is identified as category plus index (e.g. bowl.2) so that attributes and relations can later be attached to a specific instance rather than to a category in general. Attribute generation: prompted with the final-scene image and the boxed objects, Qwen3-VL-235B-A22B-Instruct produces salient visual attributes per object โ€” shape, color, size, texture, state, and so on. Relation generation: the same model produces (object.0, relation, object.1) triplets covering both spatial relations (e.g. under) and interaction relations (e.g. holding). Integration and correction: objects are joined with their attributes and relations into complete annotations โ€” note that different instances of the same category in one image may carry entirely different attributes and relations, which is why instance-level indices matter โ€” and the image plus annotations are finally passed to GPT-5.2 for correction and augmentation.

The value of this pipeline lies not in the novelty of any single step but in converting open-ended generation into controlled retrieval, matching, and validation, which cuts human effort from "annotate from scratch" to "verify and fix." The accompanying human-in-the-loop is not decorative either: prompts and filtering rules were refined through multiple rounds of human verification, the final scene annotations are corrected by human annotators, and the annotation guidelines were updated several times based on annotator feedback.

3. A three-level evaluation protocol and a unified score: making an imagined scene scorable

The model's output is free-form text, so a structure for alignment must come first. The protocol begins with description decomposition: the generated description is split into sub-descriptions at sentence separators such as periods, then spaCy extracts and lemmatizes nouns into a candidate object set. Object-level evaluation then uses Sentence-BERT to extract word embeddings and computes a similarity matrix between candidate and annotated objects; a match is declared when the two are mutual maxima along rows and columns (i.e. bidirectional greedy matching), and the refined matrix yields the object-level coverage \(S_{\text{obj}}\), which quantifies how much of the annotated objects is softly covered by the generated description. Attribute-level evaluation concatenates all sub-descriptions that refer to the same annotated object and uses a standardized prompt to have an LLM score the attribute description against the annotated phrases on a 0โ€“5 scale, averaging over all matched objects to obtain \(S_{\text{att}}\). Relation-level evaluation reuses exactly the same scoring flow to obtain \(S_{\text{rel}}\).

The three terms are linearly rescaled to 0โ€“100 and combined with weights; because \(S_{\text{att}}\) and \(S_{\text{rel}}\) are natively on a 0โ€“5 scale they are multiplied by 20 to match units, and the weights follow CompreCap (w1 = 0.25, w2 = 0.35, w3 = 0.40):

\[S_{\text{uni}} = w_1 S_{\text{obj}} + w_2 (20\,S_{\text{att}}) + w_3 (20\,S_{\text{rel}})\]

The weighting itself takes a position worth noting: relations weigh 0.40 and attributes 0.35, for 0.75 combined, while object coverage carries only 0.25. Getting the inventory of objects right matters far less than getting what they look like and how they relate, which closes off the shortcut of padding a description with nouns. Metric reliability was checked as well: with Qwen3-8B as the LLM scorer, the Spearman correlation against human scores is 0.919, while pairwise correlations among human evaluators range from 0.912 to 0.936 โ€” the automatic scorer sits inside the human-consistency band.

The same scoring logic extends to abnormal scenes. For 20 videos in which executing the action sequence produces an abnormal final state (two categories: environmental damage such as objects falling, and safety hazards such as a faucet left running), the authors additionally annotate the key object states that reflect the anomaly, score them with the same procedure used for attributes to obtain a 0โ€“5 \(S_{\text{key}}\), and combine:

\[S_{\text{abn}} = (1-\lambda)\,S_{\text{uni}} + \lambda\,(20\,S_{\text{key}}),\quad \lambda = 0.9\]

โš ๏ธ Refer to the original paper: the main text writes \(S_{\text{abn}}=(1-\lambda)S_{\text{uni}}+\lambda S_{\text{key}}\) without the ร—20 rescaling of \(S_{\text{key}}\); back-computing from the reported numbers (e.g. GPT-5.2-Chat with \(S_{\text{uni}}\) = 69.94 and \(S_{\text{key}}\) = 3.10 gives \(S_{\text{abn}}\) = 62.79) shows the ร—20 is in fact applied, so the implementable form is written here. With ฮป = 0.9, ninety percent of the abnormal-scene score rests on the key states and the ordinary scene description carries only a tenth.

4. Stepwise-reasoning strategies for long horizons: how to segment, and why single-turn inference backfires

A long action sequence invites the dynamic-programming pattern of decomposing into subproblems and performing state transitions one at a time, and the authors design two groups of experiments around it. Segmentation comes in two flavors: evenly partitioning each instance's action sequence into a preset number of segments segment_num (simple, but blind to variation in sequence length), or sliding a window of window_size atomic actions so that every window_size actions form one segment, with the last segment possibly shorter โ€” longer sequences thus naturally yield more segments.

On top of segmentation sit two inference strategies whose outcomes differ sharply. Single-turn inference places all segments in one conversation and asks the model to predict the scene after each segment in turn, conditioning each prediction on the previous segment's scene description plus the initial-scene image. This route not only fails to help but drops substantially relative to the default setting that directly predicts the final scene, and the drop worsens as the number of segments grows or the window shrinks. The case study explains why: under this prompt the model generates descriptions that track only what changed in the scene and discard everything that stayed the same, so the final description becomes markedly shorter โ€” while object coverage and unchanged attributes both depend on restating the whole scene. Multi-turn inference narrows each round's information to one segment and asks for the complete scene description after that segment, in effect performing one final-scene prediction per round. As the number of rounds grows the description becomes substantially longer and the score recovers: finer decomposition (larger segment_num, smaller window_size) works better. The mechanism is straightforward โ€” in a single turn the model shares an implicit prior that "only the changes need mentioning," whereas round-by-round restatement forces it to maintain a complete scene state every time.

The returns, however, diminish and the cost does not. segment_num = 11 and window_size = 10 score within a hair of each other on the full set (52.34 vs 52.16), the former doing better on the Short subset (53.26 vs 52.31) and the latter markedly stronger on the Long subset (51.41 vs 48.33); multi-turn with window_size = 10 beats the default setting by 3.41 on Long. But once segment_num falls below 7 or window_size exceeds 25, scores drop back below the default setting, and multi-turn inference multiplies inference time. The authors' conclusion is that adaptively segmenting action sequences by length is likely the right use of long-horizon stepwise reasoning, rather than cutting ever finer.

A Worked Example: one instance where the bottle falls over

Figure 7 of the paper walks the whole pipeline through one case. The initial-scene image is a small round table with a light green and white checkered tablecloth, holding a pyramid of stacked bottles, a lunch box, a snack bag, a tissue package, and a milk carton. The action sequence has 14 entries, from "C draws a tissue from the tissue package with his right hand" and "C wipes the tablecloth with the tissue" through "C kneads the tissue," "C picks up the lunch box... places the lunch box on the tablecloth," "C places the spoon on the lunch box," "C puts the snack bag between the lunch box and the bottle pyramid," "C puts the milk carton on the right side of the bottle pyramid," "C puts the tissue package in front of the bottle pyramid," to "C picks up the bottle in the lower right corner of the bottle pyramid with his right hand" and "C puts the bottle next to the tissue package." The key state in the final-scene annotation is a single sentence: the bottle without a label lies horizontally on the tablecloth โ€” a bottle was pulled from the pyramid, the remaining structure lost its support, and this is the canonical environmental-damage anomaly.

On \(S_{\text{uni}}\) the two leading models do not look bad here (Qwen3-VL-8B-Thinking 70.71, GPT-5.2-Chat 69.94); on \(S_{\text{key}}\) they collapse: humans 4.65, GPT-5.2-Chat 3.10, Qwen3-VL-8B-Thinking 2.84. Comparing the generated descriptions makes it concrete โ€” Qwen3-VL-8B-Thinking writes that the two remaining bottles stand in "a stable base," and GPT-5.2-Chat writes that the remaining bottle stack still stands upright, whereas the bottle whose base was removed is in fact lying down. Both models recognized that objects had been rearranged and described the table, curtain, and tablecloth pattern in fine detail, yet on the one step that actually carries the physical consequence they produced answers violating physical commonsense. The example demonstrates two things at once: the three-level structured metric can localize which semantic level is wrong, and a fluent, long LLM description can mask a failure of physical reasoning.

Key Experimental Results

Main Results

Evaluation covers more than 20 models: on the proprietary side GPT-5.2-Chat, Gemini-3-Flash, and Gemini-3-Pro; on the open-source side Qwen2-VL-7B, Qwen2.5-VL-3B/7B, Qwen3-VL-2B/8B (non-thinking and thinking), InternVL3.5-2B/8B, LLaVA-OneVision-1.5-4B/8B, Keye-VL-1.5-8B, Ovis2.5-2B/9B, MiMo-VL-7B-RL-2508, MiniCPM-V-4.5 (8B), GLM-4.6V-Flash (9B), and Step3-VL-10B; plus two models trained specifically for embodied or egocentric reasoning, Embodied-Reasoner (7B) and EgoThinker (7B). All models share the same prompt and use greedy decoding. Instances are partitioned by action-sequence length into Short (11โ€“99 actions, 599 instances), Medium (100โ€“199, 430), and Long (200โ€“694, 128); the human baseline is evaluated on EXPLORE-Bench (tiny), 100 instances sampled at a 52:37:11 ratio across the three subsets.

Model Short \(S_{\text{uni}}\) Medium \(S_{\text{uni}}\) Long \(S_{\text{uni}}\) Full \(S_{\text{uni}}\) Full \(S_{\text{obj}}\)
Human โ€  61.56 57.87 51.50 59.08 71.38
Qwen3-VL-8B-Thinking 52.08 50.56 47.07 50.96 62.70
Qwen3-VL-8B-Instruct 51.23 50.64 48.00 50.65 60.63
Gemini-3-Pro 50.11 49.44 48.31 49.66 60.94
Gemini-3-Flash 50.27 48.84 48.18 49.51 59.72
Step3-VL-10B 49.87 48.01 47.09 48.87 60.61
GLM-4.6V-Flash (9B) 49.32 47.67 45.75 48.31 59.22
GPT-5.2-Chat 48.71 47.85 46.91 48.19 59.69
MiniCPM-V-4.5 (8B) 47.67 46.49 44.95 46.93 57.87
InternVL3.5-8B 47.00 45.30 43.99 46.04 56.57
EgoThinker (7B) 39.39 37.89 35.36 38.38 46.71
Qwen2-VL-7B-Instruct 39.14 37.73 36.73 38.35 46.62
Embodied-Reasoner (7B) 32.65 32.56 32.23 32.57 39.44

โ€  The human row is from EXPLORE-Bench (tiny) and is not strictly comparable with the full-set rows. Numbers are taken from Table 2 of the paper; only representative rows are kept for length, and the complete leaderboard also includes Qwen3-VL-2B, Qwen2.5-VL-3B/7B, LLaVA-OneVision-1.5-4B/8B, Ovis2.5-2B/9B, Keye-VL-1.5-8B, MiMo-VL-7B-RL-2508, InternVL3.5-2B, and Qwen3-VL-2B-Thinking.

Ablation Study

The paper has no model-level ablation; in its place is a systematic analysis of test-time strategies, using Qwen3-VL-8B-Instruct to compare the default setting against the two stepwise-reasoning strategies across the three subsets.

Inference config Short \(S_{\text{uni}}\) Full \(S_{\text{uni}}\) Long \(S_{\text{uni}}\) Note
Default setting (direct final-scene prediction) 51.23 50.65 48.00 One-shot complete final-scene description
Single-turn stepwise (various segment_num/window_size) falls steadily as segmentation gets finer below default below default Descriptions track only what changed, drop unchanged content, and get markedly shorter
Multi-turn, even segmentation segment_num = 11 53.26 52.34 48.33 Best on the Short subset
Multi-turn, sliding window window_size = 10 52.31 52.16 51.41 Best on the Long subset, 3.41 above the default

Key Findings

  • Humans lead but their absolute score is not high, and the gap narrows on Long: humans reach 59.08 on the tiny full set, 7.39 above the best model. Their edge is clear on Short/Medium and shrinks noticeably on Long (51.50 vs Gemini-3-Pro's 47.97), which the authors attribute to MLLMs' relative advantage on very long action sequences, especially in predicting inter-object relations. Humans themselves scoring only 59 shows how hard the benchmark is.
  • Open-source models are competitive; proprietary ones do not dominate: the Qwen3-VL-8B family overtakes Gemini-3-Pro on the full set (50.65 / 50.96 vs 49.66) while Gemini-3-Pro is stronger on the Long subset (48.31); Gemini-3-Flash tracks Pro overall and edges out GPT-5.2-Chat. The open-source side as a whole still trails humans by a wide margin.
  • Models trained specifically for embodied/egocentric reasoning do worse: Embodied-Reasoner (32.57) and EgoThinker (38.38) are no better than โ€” or fall below โ€” their base model Qwen2-VL-7B-Instruct (38.35). This is the paper's most striking observation: existing embodied/egocentric training does not cover the ability to roll out the consequences of long action sequences.
  • Description length is uncorrelated with score: the top three models average 545.77, 312.46, and 186.60 words respectively โ€” Gemini-3-Pro is both accurate and concise. Qwen2.5-VL-3B-Instruct writes 305.58 words for a far lower score; Embodied-Reasoner and EgoThinker write 812.00 and 808.92 words and sit at the bottom; Step3-VL-10B writes 747.27 words yet stays competitive; Ovis2.5-9B in thinking mode is shortest at 121.69 words.
  • Thinking mode is not a cure-all: enabling thinking lifts Keye-VL-1.5-8B by 5.77 on the full set and makes Qwen3-VL-8B-Thinking one of the strongest open-source entries, but the thinking variants of Ovis2.5-2B, Ovis2.5-9B, and MiMo-VL-7B-RL-2508 all score below their non-thinking counterparts.
  • Abnormal scenes: inflated ordinary scores, collapsed key-state scores: abnormal instances have much shorter action sequences (17 on average) and fewer objects (9 per scene), so \(S_{\text{uni}}\) runs higher than on the full set (GPT-5.2-Chat 69.94, Qwen3-VL-8B-Thinking 70.71), yet \(S_{\text{key}}\) sits between 1.45 and 3.10 against 4.65 for humans. On the ฮป = 0.9 composite \(S_{\text{abn}}\), the best model GPT-5.2-Chat reaches only 62.79 against 91.64 for humans.
Method \(S_{\text{obj}}\) \(S_{\text{att}}\) \(S_{\text{rel}}\) \(S_{\text{uni}}\) \(S_{\text{key}}\) \(S_{\text{abn}}\)
Human 95.19 3.64 3.88 80.32 4.65 91.64
GPT-5.2-Chat 87.32 2.82 3.55 69.94 3.10 62.79
Qwen3-VL-8B-Thinking 86.03 2.89 3.62 70.71 2.84 58.22
Gemini-3-Pro 80.98 2.70 3.41 66.47 2.74 56.00
Qwen3-VL-8B-Instruct 83.83 2.82 3.51 68.75 2.54 52.63
EgoThinker 72.33 2.21 2.92 56.92 2.38 48.44
Qwen2-VL-7B-Instruct 68.84 2.13 2.98 55.98 2.20 45.20
Embodied-Reasoner 63.75 1.62 2.75 49.28 1.45 30.95

Highlights & Insights

  • Future-scene prediction made quantifiable, with a validated scorer: the three-level protocol decomposes "did it say the right thing" into object coverage, attribute agreement, and relation agreement, then closes the loop with an LLM judge whose agreement with humans (ฯ = 0.919) sits inside the human-to-human band (0.912โ€“0.936). This "structured annotation + LLM scoring + human-consistency check" pattern transfers to any open-ended generation evaluation.
  • Feeding action sequences instead of video is a smart framing: with video input, long-horizon evaluation would entangle sampling strategy, long context, and frame selection with the thing actually being measured. Replacing it with a planned sequence of atomic action descriptions pins evaluation on "deriving world state from actions" and aligns it with the real decision moment when an agent has planned but not yet acted.
  • The counter-intuitive multi-turn result is worth remembering: within a single turn, asking the model to update the scene segment by segment actually hurts, because the model treats "describe the changes" as the task and discards everything unchanged. When an evaluation needs complete state, the prompt must explicitly demand full restatement โ€” a lesson that transfers directly to memory maintenance in long-horizon agents.
  • The structured annotations are an asset in themselves: 23,771 objects, 1,612 categories, and instance-indexed attributes and relation triplets form a ready-made scene-graph-style supervision signal. The authors list building a dedicated training set as future work, so this data has room to serve training as well.
  • The \(S_{\text{key}}\) idea transfers to safety evaluation: scoring safety-relevant key states separately exposes physical-commonsense errors that fluent long descriptions would otherwise mask โ€” an approach equally applicable to failure-mode evaluation for robot manipulation and autonomous driving.

Limitations & Future Work

  • The authors admit three limitations: test-time scaling was explored with only a small set of inference strategies, leaving more effective and efficient methods for long-horizon reasoning open; although EXPLORE-Bench covers diverse everyday activities, rare or abnormal cases (unexpected interactions) remain under-represented; and there is no dedicated training set, which becomes valuable to construct as more data becomes available.
  • On the protocol itself: object matching relies on Sentence-BERT embeddings with mutual-maximum matching, so synonyms and hypernyms (bowl vs container) can be mismatched or missed; attributes and relations are both scored by a single LLM on a 0โ€“5 scale whose empirical distribution is squeezed into a narrow band (\(S_{\text{att}}\) mostly 1.0โ€“2.9, \(S_{\text{rel}}\) mostly 2.0โ€“3.6), limiting discrimination โ€” which is also why the 0.25 object-coverage term effectively dominates the variance of the unified score.
  • The human baseline covers only 100 instances (the tiny set) and is not strictly comparable with the full leaderboard; the paper marks it with โ€ , but conclusions such as "Gemini-3-Pro approaches human performance on Long" remain constrained by that sample size.
  • With ฮป = 0.9, \(S_{\text{abn}}\) is 90% determined by \(S_{\text{key}}\) and \(S_{\text{uni}}\) contributes only a tenth, so the score is better read as a proxy for key-state hits than as overall scene quality.
  • Improvement directions: augment attribute and relation annotation with spatial coordinates or a scene-graph structure to reduce text-matching ambiguity, and actually implement the length-adaptive segmentation the paper only observes.
  • vs EgoTaskQA / EgoMCQ / EgoSchema / EgoPlan-Bench / ERQA: these are the most widely cited egocentric/embodied benchmarks (Table 1 of the paper), and they share short actions, regional scene scope, and no guarantee that the atomic actions describe a complete long-horizon task. EXPLORE-Bench is the only one simultaneously long-action, global-scene, and completeness-satisfying, and it adds object/attribute/relation evaluation with a scene-level score.
  • vs CompreCap: this paper's protocol (object coverage, attribute score, relation score, and the 0.25/0.35/0.40 weights) is explicitly borrowed from CompreCap's directed scene graph captioning metric. The difference is that CompreCap scores a description of an image that exists, whereas this paper scores a model's imagined, not-yet-existing future scene, with the input changed from an image to an action sequence โ€” the same metric form, measuring derivation instead of perception.
  • vs EOC-Bench: EOC-Bench also works in the egocentric setting on identifying, recalling, and forecasting objects, but its unit is the object; EXPLORE-Bench requires the final state of an entire scene (objects + attributes + relations) and stretches the reasoning span to hundreds of actions.
  • vs embodied/egocentric specialists (Embodied-Reasoner, EgoThinker): such models show strong embodied reasoning within their training distribution, yet here they fall below even their own base model. This signals that "embodied ability" evaluation is heavily task-dependent โ€” improvements on one benchmark do not extrapolate to the ability to roll out the physical consequences of actions.

Rating

  • Novelty: โญโญโญโญ the task framing (action sequence โ†’ final scene) and the object/attribute/relation scene-level evaluation are new among egocentric benchmarks, though methodologically it is a systematic combination of existing modules (RAM++, Grounding DINO, the CompreCap metric).
  • Experimental Thoroughness: โญโญโญโญ 20+ models plus a human baseline, a stepwise-reasoning analysis, and an abnormal-scene analysis; the human baseline has only 100 instances and the test-time strategy space explored is narrow.
  • Writing Quality: โญโญโญโญ clear structure and informative tables; equation (2) does not match the implementation, and several symbols are garbled in the cached text.
  • Value: โญโญโญโญ provides a reusable yardstick for long-horizon egocentric scene prediction and exposes a clear weakness of current embodied/egocentric specialists, giving direct guidance for future training sets and test-time strategies.