{"id":"414f34df-56d5-4ca5-ada4-07af21fca12a","arxiv_id":"2608.06699","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"AgentPatch is a training-free, coarse-to-fine repair method that merges agentic multimodal LLMs into one static checkpoint while recovering weak-task and behavior-critical capabilities.","lead":"AgentPatch merges three specialized AI agents, one for image search, one for controlling screens, and one for detailed vision, into a single model by patching only the weak parts. It is a training-free merging recipe that could let developers combine expert AI models without retraining or per-request routing.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Selection-and-tuning on small dev splits plus unseen-test hyperparameters may inflate the GUI recovery claim; a re-split stability check is the decisive test.","rationale":"I read the paper as a serious, well-specified empirical contribution. The method is concrete and training-free, the ablations are informative, and the internal logic of coarse-to-fine repair is coherent: Stage 1 picks a recipient, Stage 2 restores weak-task-exclusive coordinates, Stage 3 patches behavior-critical FFN neurons under Guardian exclusion. I take DGC's use of gpt-5.5 and offline analysis-agent calls as an implementation choice, not a correctness violation, since the final checkpoint is static and the prompts and selector audit are described in enough detail to be auditable. The strongest load-bearing condition for the central claim is not any single equation; it is that the small fixed development partitions are representative enough to support both target selection and hyperparameter choice. That condition is least secure for the smallest and most behavior-sensitive benchmarks: AndroidWorld dev (24 tasks) and OSWorld dev (74 tasks). The supplementary material explicitly fixes these partitions and reports dev shares as low as 6.7% of the pool, and Table 3 demonstrates that the Stage 3 outcome is highly sensitive to which neurons are edited. With no error bars, no seed variation, and no dev-split resampling, the reported test-set improvement of 2.1 points over TSVM, driven substantially by AndroidWorld (+4.4) and OSWorld (+1.3), could be a selected split outcome rather than a general property of the method. The paper's own ablations partially support the mechanism rather than undermining it, but they do not address statistical stability. The reader's weakest_assumption identified essentially this same concern, and I agree with the CONDITIONAL verdict. I would not REJECT because the method is well-motivated, the ablation logic supports the design, and the failure modes are empirical rather than logical. UNVERDICTED would understate how much the paper does demonstrate. The path to ACCEPT is concrete and feasible: release partition seeds and code, rerun across resampled dev splits, report variance, and include ARM as the closest related agentic-merging baseline. My concrete test is designed to settle the central question directly.","tokens_in":20733,"tokens_out":2151,"duration_ms":19032,"concrete_test":"Resample the dev/test split several times (e.g., 5 random 20/80 partitions per benchmark, keeping the same total dev sizes as Supplementary A.2), rerun the full AgentPatch pipeline on each dev split with fixed code and hyperparameters, and report the six-benchmark test average with per-split dispersion. Alternatively, hold the reported dev split fixed and bootstrap the 24 AndroidWorld + 74 OSWorld calibration trajectories 100 times to see the spread in patched-neuron sets and final OSWorld/AndroidWorld scores. If the 56.6 average and the AndroidWorld 63.6 / OSWorld 19.0 numbers shift by more than roughly 1.5 points or fail to dominate the TSVM backbone across most resamples, the headline weak-task recovery claim is not established.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim is that AgentPatch recovers the weak GUI capability while preserving the other two, with the headline gains being AndroidWorld 59.2→63.6 and OSWorld 17.7→19.0. The load-bearing support for that claim is a pipeline whose choices are made on fixed development partitions: Stage 1 selects TSVM as the backbone, Eq. 6 identifies GUI as the weak task, and Eq. 9's beta plus the Stage 3 neuron ratios are tuned on the same small partitions. The paper states these partitions are fixed (Supplementary A.2) but gives no evidence that the measured dev behavior is stable. AndroidWorld dev has only 24 tasks, and the DGC repair evidence uses only those 24 plus 74 OSWorld trajectories; a handful of trajectory-level behavior gaps can determine which repair groups are compiled and which neurons get edited. Table 3 shows the method is sensitive to these choices: random neuron localization drops GUI by 5.6 points and Guardian ablation drops it by 5.0 points. Because the dev partition also drives selection of beta, q_r, q_guard, and alpha, any dev-set noise can propagate to all three stages. The paper reports no variance across seeds, no repeated dev-split sampling, and no comparison to the closest agentic-merging baseline ARM (Feng et al. 2026), despite it being cited in Related Work. The absence of error bars matters for the OSWorld gain of 1.3 points, which is close to the level of individual-task noise. The concern is not that the method is circular; the concern is that the headline test-set gains may be a selected-outcome artifact of one small, unrepeated calibration partition. A re-split stability experiment, or equivalently a robustness check across alternate dev pools, would settle whether the gains replicate.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces Agentic MLLM Merging as a task and proposes AgentPatch, a training-free, coarse-to-fine pipeline that (i) selects a stable merged backbone among candidate merging operators, (ii) restores weak-task-exclusive residual signals via a parameter mask, and (iii) applies an agent-guided behavior-critical patch that localizes FFN neurons from compiled repair/protection behavior spans and interpolates them toward the weak-task expert. Experiments use Qwen2.5-VL-7B with three experts (search, GUI, visual processing) and six benchmarks, with dev/test partitions kept disjoint. The headline result is an improvement of the TSVM backbone from 54.5 to 56.6 average, with AndroidWorld rising from 59.2 to 63.6 and OSWorld from 17.7 to 19.0, while V*Bench is preserved at 84.8. The paper reports component ablations, design ablations, hyperparameter sensitivity, a second-recipient generalization check, and qualitative trajectory evidence.","tokens_in":21046,"tokens_out":4041,"duration_ms":41730,"significance":"If the reported results hold, the paper makes a useful practical contribution: it shows that a merged agentic MLLM can be repaired to recover a degraded interactive capability without retraining, routing, or ensembles, and it articulates two concrete failure modes (asymmetric capability preservation and behavior-critical forgetting). The evaluation is largely honest: backbone selection, weak-task identification, and hyperparameter tuning are performed on fixed development partitions and the reported Table 1 numbers are on disjoint test partitions, so the evaluation is not circular. The paper also provides clean ablations, a second-recipient generalization check, a detailed DGC evidence contract, and a code release. The main weakness is statistical: all tables report single runs without variance or significance tests on stochastic interactive benchmarks, and the dev partitions used for pipeline decisions are small, so the magnitude and reproducibility of the headline OSWorld gain are not yet established.","major_comments":[{"comment":"All evaluation tables report single-run scores with no variance, no multiple seeds, and no significance tests. This matters directly for the central claim because AndroidWorld and OSWorld are dynamic, stochastic environments and the headline OSWorld gain is only 1.3 points (17.7 to 19.0), which is close to plausible run-to-run noise. Please provide repeated evaluations (multiple seeds or repeated executions with the same checkpoint) with means and standard deviations, and, where feasible, a paired test across tasks or trajectories to establish that the GUI gains and the overall average gain are not noise.","section":"Experiments, Table 1"},{"comment":"The pipeline makes every load-bearing choice on fixed development partitions that are small relative to the test sets: AndroidWorld dev has 24 tasks and OSWorld dev has 74 tasks, and these partitions drive Stage 1 backbone selection, Eq. (6) weak-task identification, Eq. (9) beta, the Stage 3 ratios q_r and q_guard, and Eq. (15) alpha. Because Table 3 shows that the method is sensitive to neuron localization (GUI drops by 5.6 points with random neurons) and Guardian exclusion (GUI drops by 5.0 points without it), dev-set noise can propagate through the entire pipeline. The paper reports no re-split stability check, no bootstrap over dev partitions, and no evidence that the identified weak task and the chosen hyperparameters are stable. Please add a dev-split resampling or stability analysis showing that the GUI weak-task identification and the chosen coefficients are robust to the particular dev sample.","section":"Methodology (Eqs. 5, 6, 9, 15) and Supplementary A.2"},{"comment":"The Related Work section cites ARM (Feng et al. 2026) as the most recent method for merging interactive LLM agents, yet ARM is absent from the baseline comparison in Table 1. Since ARM is the closest existing approach to the paper's setting, its omission means the claim that AgentPatch 'improves diverse merged backbones' relative to the state of the art is not fully established. Please include ARM in Table 1, or explicitly justify its exclusion (for example, if its code or checkpoints are unavailable).","section":"Related Work and Table 1"}],"minor_comments":[{"comment":"The TIES-Merging row has a formatting error: '44.619.12.0' should show separate values for AndroidWorld, OSWorld, and V*Bench, and the OSWorld score '19.1' appears to be missing from the printed alignment.","section":"Table 1"},{"comment":"The figure label 'SeletedBackbone' contains a typo and should read 'SelectedBackbone'.","section":"Figure 5"},{"comment":"The abstract's claim that AgentPatch 'improves diverse merged backbones' is supported by only two recipients (TSVM and Iso-CTS) in Table 4. Consider testing additional merging operators as recipients, or tempering the wording to 'multiple merged backbones'.","section":"Abstract and Table 4"},{"comment":"The per-group repair ratios q_r (0.0075 for AndroidWorld and 0.0025 for OSWorld) appear only in the supplementary table; the main text should state at least the value range or refer explicitly to the supplementary table when introducing q_r in Eq. (12).","section":"Supplementary A.4"},{"comment":"The Limitations section acknowledges the offline preprocessing cost and the use of three experts, but does not mention the sensitivity of the method to the dev-partition hyperparameters beta, q_r, q_guard, and alpha; adding a sentence would make the limitations more complete.","section":"Limitations"},{"comment":"The set-difference notation M^ell = (union_r R^ell_r) \\ (union_p P^ell_p) is clear, but the text could define R^ell_r and P^ell_p explicitly before Eq. (14) rather than only in the surrounding prose; currently the reader must infer the symbols from the preceding paragraphs.","section":"Equation (14)"}],"recommendation":"major_revision","confidential_remarks":"The paper is within the scope of a general CS/AI venue and the task framing is reasonably novel, though the contribution is incremental relative to existing merging methods. The central concern is evidential rather than conceptual: the claims rest on single-run evaluations and small dev partitions. I found no unattributed overlap beyond the authors' own prior work (PivotMerge, EvoMAS), which are cited appropriately."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a solid empirical paper with a genuinely new problem formulation, and the main thing I want before believing the headline numbers is a re-split stability check, not an overhaul of the method.\n\nWhat's actually new: the paper formalizes Agentic MLLM Merging and identifies two real failure modes, asymmetric capability preservation and behavior-critical forgetting. The proposed two-stage repair—weak-task unique residual recovery plus an agent-guided behavior-critical neuron patch with guardian-style protection—is a sensible, training-free combination of known ingredients that has not been put together this way before. The writing is clear, the ablation logic is internally consistent (54.5 → 56.2 → 56.6, with the residual mask beating the full active mask and both ablations in Table 3 hurting GUI by 5+ points), and the supplementary is unusually detailed: hyperparameters, prompt templates, evidence provenance, even a compiled selectors audit. That alone puts it above the typical merging paper.\n\nI also want to credit the honesty of the evaluation. Backbone selection, weak-task identification, and tuning all happen on fixed dev partitions with disjoint test splits. The circularity burden is genuinely low; no equation reduces to a fitted constant, and the self-citations (PivotMerge, EvoMAS) are relevant prior work, not padding.\n\nNow the soft spots, in proportion. The stress-test note is mostly right, and it lands on the right place: AndroidWorld dev has 24 tasks, OSWorld dev has 74, and the same small pools drive the choice of recipient, the weak-task label, and hyperparameters like beta, q_r, q_guard, and alpha. With no variance across seeds and no repeated dev-split sampling, the headline OSWorld gain of 1.3 points and even the AndroidWorld gain of 4.4 points could be a selected-outcome artifact. Table 3 shows the method is sensitive to those choices, which makes the missing stability analysis more than a formality. The absence of a comparison to ARM, the closest agentic-merging baseline, is also a real gap given ARM is cited in Related Work; I don't see a reason for it beyond ARM being very recent. These are empirical-evidence problems, not methodological ones. The method itself holds together, and Table 4's generalization to Iso-CTS is a good sign, though both recipients are selected on the same dev partitions.\n\nBottom line: this deserves a serious referee. I'd send it to review with a request for partition IDs, variance across seeds, and either a re-split stability experiment or a robustness check with alternate dev pools—plus an ARM comparison if feasible. If those come back clean, the central claim stands. I'd bring it to reading group and would probably cite the formulation in my own merging work.","headline":"AgentPatch is a genuinely new, well-specified repair pipeline for merged agentic MLLMs; the headline gains rest on one small dev partition, so the decisive question is whether the numbers survive a re-split, not whether the method is sound.","tokens_in":21641,"tokens_out":1624,"would_cite":true,"duration_ms":18546,"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":"AgentPatch claims that a merged agentic multimodal model can be repaired, without retraining, to recover its weakest interactive capability while preserving search and visual processing, raising the six-benchmark average from 54.5 to 56.6.","keywords":["Agentic MLLM Merging","model merging","weak-task degradation","behavior-critical forgetting","task vectors","GUI interaction","training-free repair","static checkpoint"],"falsifier":"Hold out different development partitions for the same six benchmarks (for example, a different 24 AndroidWorld and 74 OSWorld tasks) and rerun the full AgentPatch pipeline without touching the current test sets; if GUI is no longer selected as the weakest capability or the AndroidWorld and OSWorld gains disappear, the reported repair depends on the specific calibration split.","tokens_in":20507,"feed_emoji":"🧩","tokens_out":11183,"duration_ms":95401,"temperature":0.7,"pith_summary":"Agentic multimodal large language models, models that see, plan, and act in dynamic environments, are usually specialized for one tool or environment, and merging them into a single generalist preserves their capabilities unevenly. This paper introduces Agentic MLLM Merging, names two failure modes, weak-task degradation and behavior-critical forgetting, and proposes AgentPatch, a training-free, coarse-to-fine repair that selects a stable merged backbone, restores weak-task-exclusive parameter updates, and patches behavior-critical neurons while protecting already-working capabilities. The result is one static checkpoint with no routing or ensembles. On six benchmarks it raises the six-task average from 54.5 to 56.6, lifting AndroidWorld from 59.2 to 63.6 and OSWorld from 17.7 to 19.0 while holding V*Bench at 84.8.","feed_headline":"A patch revives the weakest skill lost in merging agentic AI models","feed_subtitle":"Training-free repair lifts AndroidWorld from 59.2 to 63.6 and OSWorld from 17.7 to 19.0 without hurting search or vision.","key_machinery":"The machinery is a three-stage parameter surgery on task vectors. Stage 1 picks the best training-free merge on fixed development splits. Stage 2 builds the weak-task unique mask $U^p_w = A^p_w \\prod_{k \\ne w}(1-A^p_k)$, a boolean mask of coordinates active in the weak-task expert but not in any other expert, and applies $\\theta^{(2)} = \\theta^{(1)} + \\beta U_w \\odot (\\theta^w - \\theta^{(1)})$. Stage 3 compiles repair and protection behavior spans from paired recipient-expert trajectories using the Diagnoser-Guardian-Compiler workflow, scores each feed-forward neuron by trajectory-balanced activation times output-projection norm, takes top fractions per behavior group, removes Guardian-protected neurons, and interpolates whole neuron blocks toward the weak-task expert. The key identity is that restricting recovery to unique weak-task coordinates plus excluding protected neurons converts merging from global averaging into targeted, capability-aware repair.","core_discovery":"The paper claims that the bottleneck in merging agentic MLLMs is not parameter conflict in general but the asymmetric loss of interactive capabilities and the loss of single decisive actions. AgentPatch's central claim is that a merged model can be repaired without retraining: first choose the most stable merge, here TSVM, as recipient; then restore only the weak-task expert's unique residual, coordinates active in GUI-Owl-7B and inactive in the other experts, with coefficient $\\beta=0.15$; finally locate feed-forward neurons that fire on expert-success behaviors, exclude neurons tied to recipient-protected behaviors via a Guardian mask, and softly interpolate the remaining neurons toward the GUI expert with coefficient $\\alpha=0.02$. The reported result is a single static checkpoint that improves diverse merged backbones, alleviates weak-task degradation, and better balances weak-task recovery with preservation of search and visual capabilities.","pith_inferences":["We note that 'training-free' applies to the final checkpoint: the DGC workflow invokes an external analysis model on full trajectories during offline construction, and the paper does not report this one-time cost; a deployment would weigh it against the serving savings of a single checkpoint.","The unique-residual mask is a hard boolean product; a natural extension is a soft mask weighting coordinates by relative task-vector magnitude, which could reduce sensitivity to the $\\beta=0.15$ choice.","The same coarse-to-fine scheme could be applied recursively: after GUI is repaired, run Stages 2 and 3 with search or visual as the weak task on their own calibration sets to see whether the balance improves further.","A testable prediction for scaling: adding a fourth expert (for example, a web-browsing agent) should still leave GUI as the repair target as long as the new capability does not dominate the development-split average."],"forward_implications":["Merged agentic MLLMs can be repaired into one deployable checkpoint, removing the need for expert routing or ensemble inference.","Weak-task recovery is non-monotonic: intermediate residual strength ($\\beta=0.15$) balances recovery with preservation, while stronger injection ($\\beta=0.30$) returns to backbone-level performance.","Behavior-level repair transfers across environments of the same capability: the same patch improves both AndroidWorld and OSWorld.","Explicit capability protection is load-bearing: disabling the Guardian exclusion costs 5.0 points on GUI and 0.6 points on visual benchmarks.","The repair generalizes across merged recipients: applying the same pipeline to TSVM and Iso-CTS raises the six-benchmark average by 2.1 and 1.1 points respectively."],"supporting_citations":[{"why":"Supplies the MMSearch benchmark, one of the six evaluation tasks used in the reported average.","marker":"(Jiang et al. 2024)"},{"why":"Supplies MMSearch-R1-7B, the agentic search expert, and the FactualVQA benchmark used for search evaluation.","marker":"(Wu et al. 2025)"},{"why":"Supplies GUI-Owl-7B, the weak-task expert whose parameters and calibration trajectories drive the repair.","marker":"(Ye et al. 2025)"},{"why":"Supplies DeepEyes-7B, the agentic visual processing expert whose capability the method aims to preserve.","marker":"(Zheng et al. 2025)"},{"why":"Supplies AndroidWorld, the mobile GUI benchmark used for calibration and test evaluation.","marker":"(Rawles et al. 2024)"},{"why":"Supplies OSWorld, the desktop GUI benchmark used for calibration and test evaluation.","marker":"(Xie et al. 2024)"},{"why":"Supplies V*Bench, the fine-grained visual search benchmark used in the six-benchmark average.","marker":"(Wu and Xie 2023)"},{"why":"Supplies HR-Bench 8K, the high-resolution visual understanding benchmark used in the six-benchmark average.","marker":"(Wang et al. 2024)"},{"why":"Supplies TSVM, the merging operator selected in Stage 1 as the stable recipient backbone.","marker":"(Gargiulo et al. 2025)"},{"why":"Supplies the task-vector formulation (expert minus base) and Task Arithmetic, a candidate backbone and baseline.","marker":"(Ilharco et al. 2022)"}],"fun_headline_variants":["Training-free patch restores weak skills in merged agentic AI","Single checkpoint repairs merged agent models without any retraining","Forgotten skills get fixed: training-free patch for merged agents","AgentPatch repairs weak tasks with one checkpoint, no retraining"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The repair targets and all hyperparameters come from small fixed development partitions (24 AndroidWorld and 74 OSWorld tasks, 6.7% to 20.9% of each pool) and from experts built on one shared base model; if those partitions misrepresent which capability is weak or the chosen coefficients do not transfer, the reported test-set gains will not replicate.","fun_headline_variants_meta":{"raw":{"variants":["Training-free patch restores weak skills in merged agentic AI","Single checkpoint repairs merged agent models without any retraining","Forgotten skills get fixed: training-free patch for merged agents","AgentPatch repairs weak tasks with one checkpoint, no retraining"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000995,"raw_usage":{"total_tokens":4215,"prompt_tokens":947,"completion_tokens":3268,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":563,"completion_tokens_details":{"reasoning_tokens":3199}},"tokens_in":563,"tokens_out":3268,"duration_ms":22724,"temperature":1.0,"reasoning_tokens":3199,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T22:13:27.556592+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Hold out different development partitions for the same six benchmarks (for example, a different 24 AndroidWorld and 74 OSWorld tasks) and rerun the full AgentPatch pipeline without touching the current test sets; if GUI is no longer selected as the weakest capability or the AndroidWorld and OSWorld gains disappear, the reported repair depends on the specific calibration split.","supporting_citations":[],"review_version":1}