{"id":"964a7c43-894a-403d-b932-b700fce9119e","arxiv_id":"2607.19971","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"Training prediction and planning on disjoint parameter subsets of a compact shared encoder, then sparse-merging their task vectors, improves both tasks on JRDB and JTA crowd-navigation benchmarks.","lead":"Robot navigation needs to predict where people will move and plan a safe path at the same time, often on small onboard computers. This paper shows that training the two skills on separate parts of one network, then merging only the largest changes, sharply improves both tasks on standard crowd benchmarks.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The causal chain rests on gradient-magnitude top-K selection, but no random-mask control tests whether those coordinates are truly task-critical.","rationale":"The reader's weakest assumption is exactly the load-bearing premise I identify: Algorithm 1 and Eq. 10 assume that top-K gradient/task-vector magnitude marks each task's core parameters. The paper's own causal evidence (Supp. Sec. 6.4) is internal to that same magnitude-selection framework and therefore does not independently validate it. The missing control is a random-mask comparison, which would distinguish 'disjointness helps because it protects task-critical coordinates' from 'disjointness helps by itself, e.g., as an inductive bias or implicit regularization.'\n\nI agree with the reader's overall CONDITIONAL assessment. The empirical narrative is coherent, cross-architecture replication and sensitivity analyses are present, and the reported effect sizes are large. But because the central mechanism is not yet isolated from the magnitude proxy, the verification step I propose would either strengthen or weaken the causal claim. Since the reader already conditioned acceptance on missing reproducibility items, this additional missing control does not move the verdict; it reinforces the conditional status. I am not raising concerns about author integrity or consensus; this is a concrete gap in the evidence chain for the paper's own proposed mechanism.","tokens_in":26819,"tokens_out":5649,"duration_ms":65473,"concrete_test":"Run a random-mask ablation of Algorithm 1: replace line 6's top-K gradient-magnitude selection with uniform random selection among unassigned parameters, holding fixed the mask-growth schedule, allocation ratio, learning rates, merge K=1%, and all other training settings on JRDB. Then apply Eq. 10 and report planning/prediction ADE and collision rate. If random masks reproduce the reported 0.4044/0.5952 values within seed noise, then gradient-magnitude saliency is not load-bearing and the Skill Conflict mechanism is not the operative explanation. If random masks regress substantially toward Sparse-Merging-without-DPT (0.9246/0.6587 or worse), the magnitude-saliency premise is independently supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim has two linked parts: (1) Skill Conflict — overlapping top-K parameter assignments degrade prediction and planning — and (2) the remedy, DPT plus Sparse Merging, works because it isolates each task's 'core skill' in the largest-magnitude coordinates. Both parts rely on magnitude as a proxy for task relevance. Algorithm 1 (Sec. 3.4) selects the top-K unassigned parameters by |∇ΘLt| during alternating training, and Eq. 10 keeps only the K=1% largest-magnitude entries of each task vector. If task-relevant computation is distributed across many small co-adapted coordinates, or if large-magnitude deltas are not the coordinates that actually determine the outputs, then the overlap diagnostic (Eq. 9), the causal overlap experiment (Supp. Sec. 6.4, Tab. 5), and the proposed remedy all inherit the same unvalidated proxy.\n\nThe causal evidence in Supp. Tab. 5 does not resolve this: it forcibly introduces overlap among masks that were themselves selected by gradient magnitude. It shows that overlap among these particular coordinates hurts performance, but it never compares magnitude-selected coordinates against randomly selected ones. Thus the large DPT+SM gains (e.g., planning ADE 0.9246→0.4044 versus Sparse Merging alone in Tab. 1) could in principle come from disjointness, from regularization toward the pretrained model, or from the DPT-specific hyperparameters, rather than from preserving the tasks' 'core' parameters. This is an internal validation gap in the causal chain, not a disagreement with external consensus.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper addresses the problem of jointly performing motion prediction and motion planning in compact, shared-encoder models for social robot navigation. It identifies a phenomenon called \"Skill Conflict\": when the top-magnitude parameter regions used by the two tasks overlap, joint performance degrades. It provides empirical evidence for this via mask-ratio sweeps (Fig. 2), parameter-overlap visualizations (Fig. 3), and a forced-overlap experiment (Supp. Tab. 5). As a remedy, the paper proposes Disjoint Parameter Training (DPT), which trains each task on disjoint, progressively grown top-gradient-magnitude masks (Algorithm 1), followed by Sparse Merging, which keeps only the top 1% of each task vector by magnitude (Eq. 10). The paper claims large improvements over existing unified models and merging baselines on JRDB and JTA (e.g., planning ADE 0.8048→0.4044 vs. DIPP, collision rate 0.0167→0.0091), with a further joint-reasoning stage improving to 0.3640 ADE and 0.0074 collision rate. The central claim is that the gains come from preserving each task's \"core skill\" in a narrow set of high-magnitude parameters while suppressing cross-task interference.","tokens_in":27235,"tokens_out":9903,"duration_ms":88585,"significance":"If the mechanism is valid, the paper makes a practically relevant contribution: it gives one of the first systematic studies of parameter-level interference in compact unified prediction-planning models and proposes a training/merging recipe that could improve edge-deployable robot navigation. The empirical work is unusually thorough: replicate the overlap phenomenon on DTPP, provide seed and schedule sensitivity analyses, extend DPT to HiP-AD/E2E-AD, and report inference cost (10.4 ms, 6.44M params), which supports the resource-constrained motivation. The main weakness is that the causal interpretation is not yet fully isolated from the magnitude-saliency proxy shared by the diagnostic, the causal experiment, and the remedy. This is a correctness-risk concern rather than a circularity in the performance claims, because the final metrics are external and the forced-overlap experiment gives some causal grounding. The requested additional controls would determine whether the mechanism is specifically \"core-skill preservation\" as opposed to a more generic sparsity/regularization effect.","major_comments":[{"comment":"The mechanistic interpretation is not validated independently of the magnitude-based selection. Algorithm 1 selects top-K coordinates by |gradient|, Eq. (10) keeps top-K task-vector entries by magnitude, and Supp. Tab. 5 forces overlap only among masks that were themselves chosen by gradient magnitude. This does not test whether magnitude identifies task-critical parameters. The observed gains could stem from generic sparse-mask regularization, from disjointness, or from DPT-specific hyperparameters rather than from preserving 'core' coordinates. Please add a random-mask control: run DPT with randomly selected masks of the same size and schedule, and merge random 1% subsets of task vectors. If magnitude selection is load-bearing, these controls should be clearly worse on the same metrics. This should also be applied to the diagnostic, since the overlap metric (Eq. 9) and the remedy share","section":"Sec. 3.4–3.5, Supp. Sec. 6.4 (Tab. 5)"},{"comment":"The text states that Sparse Merging 'consistently outperforms dense merging (K=100%) and moderately sparse configurations (e.g., K=10%)', but Table 3 stops at K=40, so the dense-merging comparison is missing. Because sparsity is the second half of the remedy, please report K=100% (and ideally intermediate values such as 60–80%) under the same DPT settings. Also, the accompanying claim that K=2 is 'best overall across all metrics' is inaccurate: in Table 3, K=1 has a lower collision rate (0.0091 vs. 0.0096), so the sentence should be rephrased to reflect the small non-monotonicity.","section":"Sec. 3.5, Table 3"},{"comment":"The interpretation of the upper rows of Table 1 is unclear. The text says Plan/Pred Finetune 'performance remains largely unchanged even with DPT', but the table shows large swings in the cross-task metrics, e.g., Plan Finetune has prediction ADE 1.9825 while Plan Finetune + DPT has 0.7368, and Pred Finetune planning ADE changes from 0.8889 to 0.8469. It is not stated whether these rows are single-task fine-tuned models evaluated with both heads, merged models, or something else. Please specify exactly which parameters are trained and what is evaluated, since this affects the claim that DPT preserves each task's inherent capability and the interpretation of DPT's compatibility with other merging methods.","section":"Sec. 4.2, Table 1"},{"comment":"The joint-reasoning stage is used to obtain the best numbers in Table 4 (DPT + SM + JR), but its description is only two sentences: a fine-tuning stage with a joint task loss on 'parameter regions that remain inactive after DPT and Sparse Merging'. The loss definition, learning rate, number of epochs, whether masks are frozen, and which specific parameter regions are updated are not given either in the main text or in the supplementary implementation details. Please provide these details, since this stage is part of the headline result.","section":"Sec. 4.2 (Effect of Adding Joint Reasoning)"}],"minor_comments":[{"comment":"The symbol K is overloaded: in Algorithm 1 it denotes the per-step mask size, while in Eq. (8)–(10) it denotes the Sparse Merging mask ratio. This creates confusion when both are discussed in the same section. Use distinct notation (e.g., K_dpt and K_sm).","section":"Algorithm 1, Eq. (8)–(10)"},{"comment":"Several figure captions contain internal version/date strings that should be removed: Fig. 4 'Final Ver_251114_1853', Fig. 6 '251113 22:01', and Supp. Fig. 1 '2511191603'.","section":"Figure captions"},{"comment":"The comparison against DIPP and DTPP does not state whether these baselines were re-trained under the same data splits and training protocol or taken from official checkpoints. Without this, differences in preprocessing or training budget may confound the architecture comparison.","section":"Sec. 4.1, Table 4"},{"comment":"The quantity D is defined as the number of parameters activated by at least one task, which is itself a function of K. This notation is easy to confuse with dataset size or embedding dimension; rename it (e.g., N_union) and state the dependence on K explicitly.","section":"Eq. (9)"},{"comment":"Minor wording issue: 'the fact that their performance remains largely unchanged even with DPT' does not match the large cross-task improvements in the table. Even if the intended meaning is 'single-task capability is preserved', the sentence should be rewritten to avoid the contradiction.","section":"Sec. 4.2, Table 1"}],"recommendation":"major_revision","confidential_remarks":"The paper is empirically broad and mostly well executed, but the central mechanistic claim rests on an unvalidated magnitude-saliency assumption. The requested random-mask controls and dense-merging comparison are straightforward and would substantially raise confidence. The novelty relative to the multi-task learning and model-merging literature is moderate, and I would advise tempering the 'first to diagnose' framing until the distinction from gradient-conflict methods is made sharper."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know. The headline number is real: DPT plus Sparse Merging gets a 6.4M-parameter unified model that beats much heavier baselines on JRDB planning (ADE 0.404 vs 0.805 for DIPP; collision 0.009 vs 0.017). The second is that the causal story—overlapping top-K gradient coordinates cause the degradation, and preserving them fixes it—is plausible but only partially supported.\n\nWhat's new: the DPT alternating disjoint-mask fine-tuning procedure and sparse top-1% task-vector merging, applied to compact prediction/planning models. I haven't seen this combination in the model-merging or multi-task robotics literature. The empirical documentation is unusually thorough: overlap-versus-mask-ratio curves, forced-overlap causal experiment in Supp Tab 5, replication on DTPP, sensitivity sweeps over schedules and seeds, efficiency numbers, and a HiP-AD autonomous-driving extension. That's more than most papers in this area ship.\n\nSoft spots, in proportion. The stress-test concern is legitimate: no random-mask control. All experiments select coordinates by gradient magnitude, including the causal overlap experiment, so we learn that overlap among these coordinates hurts, but not that these coordinates are task-critical. The gains could come from disjointness, from regularization toward pretrained weights, or from the DPT schedule and learning rates. That's an interpretation gap, not a fundamental flaw. There are also smaller issues: no released code, single-run main tables, and hyperparameters like K=1% and the 50:50 allocation ratio appear to be selected on the JRDB test set. The full-JTA numbers and HiP-AD results are less dramatic than the headline JRDB numbers, which I would have liked to see discussed.\n\nFor whom: people working on compact unified models for edge robotics and those applying model merging to small backbones. It deserves a serious referee. My recommendation: send to peer review, ask for code/configs, validation-based hyperparameter selection, error bars on baselines, and a random-mask control. The empirical method likely stands even if the causal story needs revising.","headline":"Solid, well-documented empirical method for compact prediction/planning; the 'skill conflict' story needs a random-mask control before I'd fully buy it.","tokens_in":27711,"tokens_out":6296,"would_cite":true,"duration_ms":57139,"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":"In compact unified models, prediction and planning compete for the same weights; the paper shows that training them on disjoint parameter regions and then sparsely merging the top 1% of each task's changes resolves the conflict and roughly","keywords":["motion prediction","motion planning","model merging","skill conflict","disjoint parameter training","sparse merging","crowd navigation","edge deployment"],"falsifier":"Replace the top-K gradient selection in Algorithm 1 with random selection of the same mask sizes on JRDB; if the random-mask DPT still delivers the reported ADE/collision gains, then magnitude saliency is not the cause — and if the controlled-overlap Table 5 trend disappears under random masks, Skill Conflict is not the mechanism.","tokens_in":26712,"feed_emoji":"🤖","tokens_out":8667,"duration_ms":77478,"temperature":0.7,"pith_summary":"Unified robot navigation models that share one small encoder for motion prediction and motion planning suffer from what the paper calls Skill Conflict: the two tasks' gradient updates land on the same weights, so strengthening one skill degrades the other and joint performance saturates early. To fix this, the paper proposes Disjoint Parameter Training (DPT), which alternately trains each task and permanently assigns it a distinct, non-overlapping set of high-gradient parameters, then freezes those masks so each task's core skill stays local. After that, Sparse Merging adds back only the most salient 1% of each task vector to the pretrained weights, suppressing interference between adjacent features. On crowd-navigation benchmarks, this framework cuts planning average displacement error roughly in half compared with a leading unified baseline while using only about 6.4M parameters, and an optional joint-reasoning stage improves both tasks further. A sympathetic reader would take away that parameter-level separation, not model size, is the lever that makes compact unified prediction and planning work.","feed_headline":"Weight split halves planning error in a 6.4M-parameter robot model","feed_subtitle":"Predict and plan on separate weights, merge only the top 1%, and a tiny model beats bigger unified ones.","key_machinery":"The load-bearing objects are the task vector τ_t = Θ_t − Θ_0, the binary activation mask M_t^(K) that selects top-K coordinates by magnitude, and the alternating DPT loop that grows disjoint masks from unassigned parameters. The merge formula Θ_merged = Θ_0 + M_pred^(K%) ⊙ τ_pred + M_plan^(K%) ⊙ τ_plan is the single operation that does the work: disjoint masks guarantee spatial separation of core skills, while the 1% sparsity filters out low-magnitude adjacent-feature coupling that would otherwise cause cross-task interference. The paper also uses the Overlap(K) metric (coordinate-wise AND of the two masks) as the diagnostic and the controlled-overlap experiment as the causal test.","core_discovery":"The paper's central discovery is a mechanism, not just a method: when prediction and planning are trained in a compact shared encoder, the top-magnitude parameter regions that each task needs overlap heavily (8% overlap at 10% utilization, 39% at 50%), and this overlap is causally tied to degraded joint performance. DPT prevents the overlap by construction: during the first half of fine-tuning, the algorithm alternately selects the top-K unassigned parameters by gradient magnitude, assigns them exclusively to one task via a binary mask, and updates only those; in the second half the masks freeze and each task consolidates. Sparse Merging then combines the two models as Θ_merged = Θ_0 + M_pre","pith_inferences":["If the magnitude-saliency premise is right, a testable extension follows: DPT should combine cleanly with post-training compression — pruning or quantizing the merged model's low-magnitude coordinates should cost little, because the skill was already concentrated in the top 1%.","The overlap diagnostic could be used as a cheap predictor before committing to DPT: compute the two top-K masks from a standard fine-tuned pair; if Overlap(K) is high, switching to disjoint training should help, and the expected gain should scale with the measured overlap.","The paper's framing suggests Skill Conflict is a general resource-constrained multi-task phenomenon, not just prediction-vs-planning: their own three-way split of prediction shows conflicts within a single task; one could extend DPT to more tasks, with the caveat that the 'sweet spot' tradeoff between specialization and capacity-sharing will need a principled allocation rule.","Because the merge formula is a simple arithmetic identity on masks and task vectors, it is architecture-agnostic; one could transfer it to other compact multi-task models (e.g., embedded vision-language-action policies) provided the top-K gradient premise holds there."],"forward_implications":["DPT is orthogonal to the merging rule: adding it to Task Arithmetic, Ties Merging, and T-Switch consistently improves both planning and prediction on JRDB and JTA, so merging methods designed for large models can be rescued for compact models by changing how material models are trained.","Sparse merging is not a detail: K=1–2% consistently beats dense merging (K=100%) and moderate sparsity, implying that in compact encoders each task's transferable skill lives in a narrow coordinate set.","Because the merged model is a single network with no added inference branches, it keeps edge compatibility: roughly 6.4M parameters, 10.4 ms latency, 24.6 MB model size, and 1.96 MB inference memory on an RTX 3090.","An optional joint-reasoning fine-tuning stage on the unassigned parameter regions recovers the benefit of unified reasoning (ADE 0.3640, collision 0.0074) without sacrificing the disjoint skill separation.","The same recipe transfers beyond social navigation: applying DPT and sparse merging to an end-to-end autonomous driving model improved closed-loop and open-loop planning/perception metrics on Bench2Drive."],"fun_headline_variants":["Separate weights, merge top 1%, beat larger robot models","Disjoint parameters halve planning error in tiny robot model","Skill conflict in shared encoders fixed by disjoint training","Split prediction and planning weights, then merge sparsely","6.4M-param model with disjoint training outperforms big ones"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The load-bearing premise is that a task's 'core skill' is captured by the parameters with the largest gradient magnitudes during alternating training, so that freezing 99% of the model in sparse merging discards only irrelevant weights; if a task's information is broadly distributed or lives in low-magnitude co-adapting coordinates, the disjoint-top-K recipe would throw away needed capability.","fun_headline_variants_meta":{"raw":{"variants":["Separate weights, merge top 1%, beat larger robot models","Disjoint parameters halve planning error in tiny robot model","Skill conflict in shared encoders fixed by disjoint training","Split prediction and planning weights, then merge sparsely","6.4M-param model with disjoint training outperforms big ones"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000177,"raw_usage":{"total_tokens":1149,"prompt_tokens":781,"completion_tokens":368,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":525,"completion_tokens_details":{"reasoning_tokens":284}},"tokens_in":525,"tokens_out":368,"duration_ms":4398,"temperature":1.0,"reasoning_tokens":284,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-01T11:09:13.008348+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Replace the top-K gradient selection in Algorithm 1 with random selection of the same mask sizes on JRDB; if the random-mask DPT still delivers the reported ADE/collision gains, then magnitude saliency is not the cause — and if the controlled-overlap Table 5 trend disappears under random masks, Skill Conflict is not the mechanism.","supporting_citations":[],"review_version":1}