{"id":"23c520c1-bf48-483e-849d-a7a69fe9427e","arxiv_id":"2509.02322","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A generalist agent with shared shallow layers and task-separated deep experts outperforms single-domain GUI and embodied agents on AndroidControl, GUI-Odyssey, and LIBERO benchmarks.","lead":"OmniActor is a single AI model that drives both on-screen GUI actions (2D apps) and robot arm movements (3D worlds), reporting top results on standard benchmarks for both. It works by sharing the model's early layers across both worlds while splitting later layers into separate experts, a design the authors say exploits synergy and avoids conflict between the two task types.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"K=8 layer split rests on a single unvalidated update-similarity statistic; if that statistic is noise, the synergy/conflict mechanism is unsupported.","rationale":"The paper is a competent engineering contribution: the ablation ladder (full-share, full-separate, OmniActor) is well designed, and the OmniActor-vs-OmniActor-hard comparison is the right deconfounding experiment, showing a real advantage for the shallow-shared/deep-separated structure. However, the paper's scientific claim is that this structure works because of 'synergy' and 'conflict' measurable via parameter-update similarity. That statistic is the only evidence for K=8 and for the cerebrum/cerebellum story. It has no error bars, no null baseline, and the reported values are very small in absolute terms. The reader's weakest assumption correctly identifies this gap. I agree with the CONDITIONAL verdict: the mechanism and the K choice need validation before the paper's central explanation can be accepted. The performance claim itself would also benefit from seeded runs, but the mechanism is the more distinctive and load-bearing part of the paper's contribution.","tokens_in":13800,"tokens_out":4765,"duration_ms":54105,"concrete_test":"Run 5 independent single-domain fine-tunes (GUI-only, embodied-only) and compute the per-layer update similarity; compare the shallow-vs-deep difference against a null distribution (e.g., random layer permutation or same-domain updates from different seeds). Then sweep K in {4,8,12,16} for OmniActor-style training and report success rates with seeds. If the shallow-vs-deep difference is within seed noise, or if the best K is far from 8 and the OmniActor-vs-OmniActor-hard margin disappears, the synergy/conflict mechanism is not validated.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central architectural claim rests on Sec 3.3's parameter-update-similarity statistic. d_gui and d_robot are computed from two separately fine-tuned single-domain models (OmniActor-GUI and OmniActor-EA), and their cosine similarity (Figure 4, peaking ~0.01-0.02) is used to conclude 'shallow synergy, deep conflict' and to set K=8. This is load-bearing because without it there is no principled reason to share layers 1-8 and separate layers 9-28; the ablation (OmniActor vs OmniActor hard) shows shallow sharing helps, but only at this single K. The statistic is never validated: no seed variance, no null distribution, no sensitivity to K, and the cosine values are tiny (0.01-0.02), plausibly within optimization noise. If the measured similarity is not a reliable predictor of joint-training transfer, then the proposed mechanism and the K=8 design choice are post-hoc, and the paper's claimed explanation for the performance gain is unsupported even if the gain itself is real.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes OmniActor, a single 7B vision-language-action model that handles both GUI (2D) and embodied (3D) tasks. The two central contributions are (i) a unified data format and action space, converting GUI actions into text tokens and continuous embodied actions into discretized token IDs; and (ii) a Layer-heterogeneity MoE architecture that shares parameters in shallow layers and separates parameters in deep layers, motivated by a claimed 'shallow synergy, deep conflict' pattern measured by a parameter-update-similarity statistic. The paper reports experiments on LIBERO-90, AndroidControl-Low/High, and GUI-Odyssey, showing that the full OmniActor model outperforms single-domain specialists (OmniActor-GUI and OmniActor-EA) on average, and also outperforms a fully-separated MoE variant (OmniActor hard).","tokens_in":13894,"tokens_out":6186,"duration_ms":70130,"significance":"If the results hold, the paper would make a useful contribution to the growing area of generalist GUI-and-embodied agents. The strongest evidence is the internal comparison in Table 2: directly mixing GUI and embodied data hurts (65.3 avg), fully separating parameters recovers most of the loss (71.3), and the proposed shallow-share/deep-separate design improves further (74.9). This ablation is clean and supports the architectural direction. The paper also contributes a large-scale data unification effort, including a unified action space, and demonstrates the architecture transfers to a different base MLLM (Qwen2.5-VL, Table 3). However, several load-bearing aspects need strengthening: the proposed 'parameter update similarity' justification for choosing the layer split, the absence of error bars, a data-amount confound in the comparison to single-domain baselines, and an unfair baseline adaptation for Magma.","major_comments":[{"comment":"The choice of K=8 (shared layers 1-8) is motivated entirely by the 'parameter update similarity' statistic in Figure 4. The reported cosine similarities are tiny (peaking around 0.01-0.02), and the paper provides no null distribution, no seed-to-seed variance, and no sensitivity analysis with respect to K. As a result, it is not established that these values are distinguishable from optimization noise, nor that K=8 is the correct or even a robust threshold. The ablation in Table 2 shows that the specific K=8 configuration works, but it does not demonstrate that the similarity statistic is predictive of joint-training transfer. Please provide: (a) the same similarity analysis over multiple random seeds/repeats, (b) a null model (e.g., permuted data or random parameter masks), and (c) an ablation over K (e.g., K=4, 12, 16, full-share, full-separate). Without these, the proposed mechanistic","section":"§3.3, Figure 4, §4.1"},{"comment":"The comparison against OmniActor-GUI and OmniActor-EA is confounded by data quantity. Section 4.1 states that trajectory data is mixed at a GUI-to-embodied ratio of approximately 1:5, for a total of ~4.1M samples. Thus OmniActor-GUI is trained on roughly 0.68M GUI trajectory samples, and OmniActor-EA on ~3.4M embodied samples, whereas OmniActor is trained on the full 4.1M. The fact that OmniActor outperforms these single-domain baselines could therefore be due to larger total data, not to synergy between domains. To support the abstract's claim that OmniActor 'outperforms agents only trained by GUI or embodied data,' the authors should either train the single-domain baselines with a matched total number of update steps (e.g., by upsampling their own domain's data) or otherwise control for data quantity. The OmniActor-hard comparison already controls for data quantity among the MoE varian","section":null},{"comment":"The baseline protocol for Magma is problematic. The paper states that Magma does not report results on the selected benchmarks, and so the authors 'finetune the pre-trained model on LIBERO, AndroidControl, and GUI Odyssey to report the performance.' This gives Magma in-domain supervised fine-tuning on the target tasks, while most other baselines (e.g., SeeClick, Aria-UI, OS-Atlas) are presumably evaluated without such task-specific adaptation. This is not an apples-to-apples comparison and likely disadvantages or advantages Magma in an uncontrolled way. The authors should either evaluate all baselines under the same protocol (zero-shot or with equal per-task fine-tuning) or exclude Magma from the headline comparison and report it in a clearly labeled subsection.","section":null},{"comment":"No error bars, standard deviations, or significance tests are reported anywhere; all results appear to be single runs. This is particularly important because the headline 'outperforms single-domain agents' claim is supported by small differences. For example, OmniActor-GUI scores 89.4 on AndroidControl-Low while OmniActor scores 86.4, a decrease of 3.0 points; conversely OmniActor gains 3.0 points on GUI-Odyssey. Without multiple seeds or a statistical test, it is impossible to know whether these differences are meaningful. The paper should report at least 3 seeds (or bootstrapped confidence intervals) for the central comparisons in Tables 2 and 3.","section":null},{"comment":"The abstract's claim that OmniActor 'outperforms agents only trained by GUI or embodied data in GUI or embodied tasks' is an overstatement. In the GUI domain, OmniActor (86.4) is numerically worse than OmniActor-GUI (89.4) on AndroidControl-Low, and it only wins on AndroidControl-High and GUI-Odyssey. Similarly, the 'average 1.2% higher' in Section 4.2 is driven by averaging three benchmarks. The claim should be qualified as 'on average' or 'on long-horizon GUI tasks' to match the reported results.","section":null}],"minor_comments":[{"comment":"Typo: 'gragh user interface' should be 'graphical user interface' (or 'GUI').","section":"§1"},{"comment":"The phrase 'interleavely' is likely a typo for 'interleaved' or 'interleavingly.'","section":"§1, §4.3"},{"comment":"In the visualization paragraph, 'Nrobot indicates GUI sample count' should be 'embodied sample count.'","section":"§4.3, Figure 5"},{"comment":"The discretization parameter K (number of bins for embodied actions) is introduced but its value is never specified. Please state the value used in experiments and, if possible, include an ablation on this hyperparameter.","section":"§3.2"},{"comment":"The observation that switching from Qwen2-VL to Qwen2.5-VL decreases LIBERO-90 success (69.5 -> 65.2) is attributed to a guess ('We guess that...'). A brief investigation (e.g., action tokenization differences or data format mismatch) would be more informative than a guess.","section":"§4.3, Table 3"}],"recommendation":"major_revision","confidential_remarks":"The paper's core architectural idea is promising and the data unification effort is valuable. However, the central mechanistic claim (K=8 from parameter-update similarity) and several experimental protocols need substantial strengthening before the results can be accepted as stated. The comparison with single-domain specialists also needs to be made fair with respect to data quantity. I would encourage the editor to consider a major-revision decision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"I read this as a serious engineering paper, not a breakthrough. The genuinely new piece is using per-layer parameter-update cosine similarity to decide where to share and where to separate parameters across GUI and embodied data. The ablations support the broad direction: full-share averages 65.3, full-separate 71.3, their OmniActor 74.9 on the four benchmarks. That ordering is meaningful, and the unified action tokenization plus the collected data are solid, reproducible-in-principle contributions.\n\nThe soft spots are real but not fatal. The load-bearing K=8 split comes from Figure 4, which shows cosine similarities peaking around 0.01–0.02 in shallow layers, and nothing in the paper demonstrates these values are distinguishable from optimization noise. No seed variance, no null distribution, no sensitivity analysis around K. So the mechanism is suggestive, not established. The paper's own cross-base-model results add to the caution: moving to Qwen2.5-VL changes LIBERO by -4.3 and GUI-Odyssey by +15, which tells us the system is not statistically pinned down. The GUI-Odyssey evaluation protocol is also underspecified.\n\nThat said, the stress-test's stronger claim—that the whole synergy/conflict architecture is post-hoc—does not land. The ablation comparison stands independently of the similarity statistic: OmniActor beats both hard separation and full sharing at a single operating point, and the authors explicitly acknowledge in Sec 5 that the embodied scenarios are limited. The statistic may not prove the mechanism, but the empirical gain is there. I would not demand a proof of the mechanism for acceptance, but I would demand the missing validation.\n\nThis paper is for people working on generalist agents, multitask interference, and sparse MoE design. It deserves a serious referee, especially because it reports external benchmarks and not just fitted curves. My recommendation: send it to review, and push the authors to release code, report seeds and variance, ablate K, and compare the update-similarity values against a trivial null (e.g., random parameter splits or shuffled data). Those additions would move it from conditional to solid.","headline":"Worth a referee: the empirical result is plausible and the ablations informative, but the shallow-shared/deep-separated mechanism rests on one unvalidated similarity statistic.","tokens_in":14632,"tokens_out":1363,"would_cite":true,"duration_ms":16929,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Screens and robots need not fight: one 7B agent beats both single-domain specialists by sharing shallow layers and splitting deep ones.","keywords":["multimodal agent","GUI agent","embodied agent","mixture of experts","layer sharing","parameter update similarity","action space unification","data conflict"],"falsifier":"Run the same parameter-update similarity computation twice, replacing the GUI-versus-embodied comparison with two independent fine-tunes on the same GUI data. If the within-domain similarity profile is indistinguishable from the cross-domain profile in shallow layers, or if a model trained with a different layer-split threshold (say K=4 or K=20) matches or beats K=8, then the measured gains cannot be attributed to the shallow-synergy/deep-conflict mechanism.","tokens_in":13500,"feed_emoji":"🤖","tokens_out":8718,"duration_ms":91709,"temperature":0.7,"pith_summary":"The paper tries to establish that a single multimodal agent can master both 2D screen tasks and 3D robot manipulation without losing to specialists. Its central difficulty is data conflict: naively mixing GUI and embodied training data lowers performance on both. The authors diagnose the conflict by comparing parameter-update directions from single-domain fine-tuning, finding that shallow layers pull in similar directions while deep layers pull apart. They then build OmniActor, which shares the first eight layers across both domains and gives the remaining layers separate GUI and embodied expert branches, while unifying both action types into one token vocabulary. On LIBERO-90 and three GUI benchmarks, the joint model outperforms both its embodied-only and GUI-only counterparts, especially on long-horizon tasks.","feed_headline":"One 7B agent beats both screen-only and robot-only specialists","feed_subtitle":"Sharing shallow layers and splitting deep ones turns mixed 2D/3D training from a liability into a win.","key_machinery":"Layer-heterogeneity Mixture of Experts — a parameter-sharing policy driven by a 'parameter update similarity' metric. The metric computes the cosine similarity between per-layer weight differences from two fine-tunes, one trained on GUI data and one on embodied data; high similarity in shallow layers motivates sharing layers 1-8, while low similarity in deep layers motivates separate attention, FFN, and head parameters for the two domains. This metric is what converts the qualitative idea of synergy and conflict into a concrete architectural split.","core_discovery":"On the paper's own terms, the central discovery is that GUI and embodied data have a layer-dependent relationship: their parameter update directions agree in shallow layers and diverge in deep layers, and an architecture can exploit that split. OmniActor is a 7-billion-parameter vision-language model with a shared trunk (layers 1-8), separate deep expert branches for GUI and embodied actions, and separate prediction heads. Trained on unified data, where GUI actions are text tokens and embodied six-degree-of-freedom-plus-gripper actions are discretized into the same vocabulary, it reaches 69.5 success on LIBERO-90 versus 63.4 for the embodied-only agent, and 86.4/77.5/66.0 on AndroidControl-L","pith_inferences":["The same shallow-share/deep-split diagnostic could apply to other pairs of task families with shared perception and divergent action syntax, such as different robot embodiments or diverse UI platforms; the paper only demonstrates it for GUI versus embodied.","The long-horizon GUI gains suggest embodied trajectory data may teach a general notion of planning or state-change that helps GUI tasks; a clean test would fix GUI data and ablate embodied data to see how much of the AndroidControl-High/GUI-Odyssey improvement disappears.","Because the similarity statistic is computed from single-domain fine-tunes with one initialization, a stronger validation would compare it with repeat fine-tunes; if shallow-layer similarity is not much higher than run-to-run noise, the architectural split is still empirically useful but the stated mechanism is unsupported by the metric.","The model's parameter count grows only in the deep layers, so scaling the shared trunk or expert capacity independently is a natural next test of whether the synergy scales with breadth."],"forward_implications":["Joint training with shared shallow trunk and split deep experts yields higher average success than either single-domain specialist: +6.1 on LIBERO-90 over the embodied-only agent and +1.2 average on the GUI suite over the GUI-only agent.","Naive data mixing is a poor baseline; full parameter separation regains lost performance but misses transfer, so the layer-dependent sharing policy is what recovers the synergy.","The recipe transfers to a newer base model, raising the claimed average success from 74.9 to 78.4.","Unifying both action spaces into one token vocabulary lets a single model emit click/tap coordinates and continuous robot-end-effector moves with the same decoder.","The parameter-update similarity metric provides a reusable heuristic for deciding where to share and where to split in other multi-domain agent training setups."],"supporting_citations":[{"why":"Supplies the base 7B vision-language model that all OmniActor variants fine-tune.","marker":"Wang et al. (2024)"},{"why":"Supplies the LIBERO benchmark and embodied trajectory data used for robot evaluation and training.","marker":"Liu et al. (2023a)"},{"why":"Supplies AndroidControl-Low and AndroidControl-High, the main GUI evaluation tasks.","marker":"Li et al. (2024a)"},{"why":"Supplies GUI-Odyssey, the cross-application long-horizon GUI benchmark.","marker":"Lu et al. (2025)"},{"why":"Supplies Aguvis GUI trajectory training data and prior action-space unification for GUI agents.","marker":"Xu et al. (2025)"},{"why":"Supplies OS-Atlas multi-platform GUI grounding data used in the grounding pretraining stage.","marker":"Wu et al. (2025)"},{"why":"Supplies UGround GUI grounding data focused on human-like mouse and keyboard actions.","marker":"Gou et al. (2025)"},{"why":"Supplies Aria-UI synthetic grounding data used to scale GUI grounding training.","marker":"Yang et al. (2025b)"},{"why":"Supplies the two-stage training recipe and grounding data selection that OmniActor follows.","marker":"Huang et al. (2025)"},{"why":"Provides the GEA generalist baseline that motivates continuous multi-embodiment action tokenization and online RL comparison.","marker":"Szot et al. (2025)"}],"fun_headline_variants":["Split deep layers, share shallow: one agent for screens and robots","7B agent masters GUI and embodied tasks with a brain-inspired split","Layer-wise MoE resolves 2D-3D data clash, boosts both tasks","OmniActor: same model clicks and grasps, beating specialists","Shallow synergy, deep split: a generalist agent for 2D and 3D"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The entire design rests on trusting that measuring which layers change in similar directions when training on screens alone versus robots alone tells you which layers are safe to share when training on both; if that measurement mostly reflects random training noise, the mechanism falls apart.","fun_headline_variants_meta":{"raw":{"variants":["Split deep layers, share shallow: one agent for screens and robots","7B agent masters GUI and embodied tasks with a brain-inspired split","Layer-wise MoE resolves 2D-3D data clash, boosts both tasks","OmniActor: same model clicks and grasps, beating specialists","Shallow synergy, deep split: a generalist agent for 2D and 3D"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000832,"raw_usage":{"total_tokens":3493,"prompt_tokens":790,"completion_tokens":2703,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":534,"completion_tokens_details":{"reasoning_tokens":2602}},"tokens_in":534,"tokens_out":2703,"duration_ms":23491,"temperature":1.0,"reasoning_tokens":2602,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T11:36:00.151486+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same parameter-update similarity computation twice, replacing the GUI-versus-embodied comparison with two independent fine-tunes on the same GUI data. If the within-domain similarity profile is indistinguishable from the cross-domain profile in shallow layers, or if a model trained with a different layer-split threshold (say K=4 or K=20) matches or beats K=8, then the measured gains cannot be attributed to the shallow-synergy/deep-conflict mechanism.","supporting_citations":[],"review_version":1}