{"id":"6985269c-c6d1-477f-bc1d-7facbc98d762","arxiv_id":"2411.18201","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"ABIL uses abductive reasoning with a supplied knowledge base to learn symbolic perception from raw observations and then trains task-specific behavior policies, yielding better data efficiency and generalization in long-horizon planning tasks.","lead":"The paper presents ABIL, a method that teaches agents to imitate demonstrations while using symbolic logic to guide long sequences of actions. It claims that this combination needs fewer demonstrations and generalizes better to new environments and composed tasks than standard imitation or model-based neuro-symbolic planning.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The symbolic knowledge base is a per-instance solution plan: ABIL's gains over BC/DT may come from privileged task structure, and the 'zero-shot' throw-2/3 result appears pre-encoded in G.","rationale":"The reader's weakest assumption correctly identifies the reliance on an accurate and sufficient knowledge base. My stress-test agrees with that direction but sharpens it: the danger is not merely that G might be wrong or incomplete; the supplied G in Appendix D is already an instantiated, object-specific plan. That makes the comparison to BC/DT a comparison between a method given a high-level symbolic solution and methods that are not. The zero-shot throw-away-leftovers setup is the most concrete symptom: training with one hamburger but a graph containing hamburger_1/2/3 means the ability to handle more objects is written into the knowledge base before any learning occurs. If the graph in Appendix D is not the exact input, the paper needs to explain how a general G is mapped onto variable object counts; the description as a finite-state machine over ground atoms does not make this clear. This does not falsify the central claim, but it means the empirical support overstates what is learned from demonstrations. The paper openly acknowledges the KB limitation in Section 5, so a conditional verdict remains appropriate; the missing piece is an ablation that isolates the contribution of abductive learning from the contribution of the hand-supplied symbolic plan.","tokens_in":19027,"tokens_out":6170,"duration_ms":63266,"concrete_test":"Inspect the released GitHub repository for the exact G used in the Throwing away leftovers zero-shot task; then rerun ABIL with a reduced G that contains a single hamburger node and no repeated-chain edges. If success on the 2- and 3-hamburger evaluation drops to near zero, the zero-shot result is an artifact of G's pre-encoded object list; if performance persists, the concern is settled and the method genuinely composes the learned behavior.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim is that ABIL achieves better data efficiency and generalization through abductive pseudo-labeling and symbolic-grounded imitation. The load-bearing premise is the supplied knowledge base G, but the issue is sharper than 'accurate and sufficient': in the supplied experiments G is an object-specific solution graph, not a general task theory. Appendix D instantiates nodes as Pick(hamburger_1, countertop), Place(hamburger_2, ashcan), Put(soap, bucket), Move-into(star, box). Equation 1's constraint {z_t^i} |= G then does little more than fit perception to a human-authored plan; Eq. 2's operator selection is read off the same graph. BC and DT receive no equivalent privileged structure, so the comparisons in Tables 1-4 are confounded: the symbolic rule may be doing the work, not the abduction or the imitation objective. A concrete symptom is the zero-shot experiment in Section 4.5: training uses one hamburger, yet Appendix D.2's Throwing away leftovers graph explicitly contains hamburger_1, hamburger_2, and hamburger_3 nodes and edges. If that graph is the actual G, the generalization to two or three hamburgers is pre-encoded in the knowledge base, and the claim of zero-shot compositional generalization is not supported. If the graph is only illustrative, the paper must specify how G is instantiated for varying object counts. Section 5's limitation (2) concedes dependence on an accurate and sufficient KB, but the paper never tests robustness to KB errors or to KBs that are abstract rather than per-instance; Table 6 perturbs perception accuracy, not G.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes ABIL, a neuro-symbolic imitation learning framework for long-horizon goal-based planning. The method assumes a symbolic knowledge base G (a finite-state machine over sub-task nodes) for each task, plus expert demonstrations without predicate annotations. ABIL operates in two stages: (1) abductive pseudo-labeling, where a perception network f is trained by iteratively proposing symbolic state labels z for each demonstration frame that satisfy G while staying close to f's current predictions (Eq. 1), using 'sequential consistency' rules for sequential, alternative, and conjunctive sub-task structures (Section 3.2); and (2) symbolic-grounded imitation, where a separate behavior module is trained per logical operator and the operator is selected at inference by matching the perceived state to plan nodes (Eq. 2), yielding a gated, reasoning-aware policy (Section 3.3). The paper evaluates ABIL on BabyAI, Mini-BEHAVIOR, and CLIPort against BC, DT, and PDSketch, reporting success rates, neuro-symbolic grounding accuracy under varying data budgets, data-efficiency curves, and zero-shot transfer experiments for composed or repeated tasks. The main claims are significantly improved data efficiency and generalization relative to the baselines, achieved without predicate annotations.","tokens_in":19307,"tokens_out":18977,"duration_ms":153419,"significance":"The core idea—replacing predicate-level annotation with an abduction loop constrained by a task-level state machine, and gating per-operator behavior modules by symbolic state—is well motivated and timely. If the results hold, the practical significance is real: per-operator decomposition plus symbolic gating yields large gains on genuinely long-horizon tasks (Mini-BEHAVIOR tasks of 45-106 steps; CLIPort assembly tasks, e.g., Assembling-kits from 0.187 to 0.829), and the grounding-accuracy curves (Figures 4, 6) show a data-efficiency advantage over PDSketch that is the kind of falsifiable claim this community values. The manuscript also ships code (Appendix C), uses 100 evaluations and three seeds throughout, compares against the strong neuro-symbolic baseline PDSketch, includes an explicit robustness probe with imperfect grounding (Table 6), and states its limitations honestly (Section 5).","major_comments":[{"comment":"The zero-shot transfer results in Section 4.5 do not support the claimed compositional generalization because the target composition is already present in the supplied knowledge base. In the Throwing-away-leftovers experiment, training uses one hamburger, but the knowledge base in Appendix D.2 explicitly enumerates Pick(hamburger_1, countertop), Pick(hamburger_2, countertop), and Pick(hamburger_3, countertop) as well as Place(hamburger_1, ashcan), Place(hamburger_2, ashcan), and Place(hamburger_3, ashcan) as nodes and edges of G. If this graph is the G used at test time, Eq. (2) reads the repeated-task structure directly off G, so the zero-shot margins in Table 4 (ABIL at 0.49-0.70 vs. 0.05-0.18 for BC/DT on Throw 2 and Throw 3) reflect privileged input structure rather than learned generalization. The manuscript must state whether G is a template over parametric object variables (in which case the instantiation procedure for 1, 2, and 3 objects must be described) or an object-specific plan (in which case the zero-shot claim in Table 4 is confounded and should be withdrawn or re-run with a G that does not enumerate the test objects).","section":"Section 4.5; Appendix D.2"},{"comment":"The headline comparisons against BC and DT are confounded by the privileged per-task knowledge base G. In BabyAI, Appendix D.1 gives each task a state machine that already encodes the subgoal order (e.g., the Unlock graph contains Pick(red_key) before Goto(red_door) and Open(red_door)), and ABIL's operator selection (Eq. 2) and per-operator behavior modules (Eq. 4) consume exactly this structure, while BC and DT receive only raw observations and actions. Consequently, the large margins in Tables 1-4 could be driven by the operator decomposition itself rather than by the abductive pseudo-labeling that the paper emphasizes. To support the attribution in the abstract, the paper should add ablations that (i) provide the same operator decomposition to the baselines (e.g., per-operator BC/DT trained with ground-truth operator labels and an oracle selector derived from G) and (ii) replace the abduced labels with labels read off G directly, isolating the perception component. The comparison with PDSketch is less affected because PDSketch's full mode also consumes the symbolic transition structure; the confound specifically concerns the BC/DT baselines, which are the primary comparators for the data-efficiency and generalization claims.","section":"Section 4, Tables 1-4; Sections 3.2-3.3"},{"comment":"The formal abduction objective and the implemented procedure are not the same object. Eq. (1) defines the pseudo-labels as the argmin of a constrained optimization over the whole trajectory, but the bullet list that follows gives four informal consistency rules ('task a before task b', 'either a or b', and so on), and the paper never states the algorithm that solves Eq. (1), nor shows that the sequential-abduction rules are equivalent to (or an approximation of) that argmin. Since the perception network f is trained against these pseudo-labels, the correctness of the entire perceptual component depends on this step; the paper should specify the exact inference procedure (e.g., a dynamic program over the state-machine graph) and state any optimality gaps explicitly.","section":"Section 3.2, Eq. (1)"},{"comment":"The paper claims 'significantly improved data efficiency and generalization across various long-horizon tasks' (abstract) and describes 'a policy ensemble' (Section 3.3), but neither claim is fully supported by the reported experiments. No significance tests are reported in Section 4, and several table cells contradict the blanket claim: in Table 2, ABIL-DT is below BC on Cleaning A Car Basic (0.340 vs. 0.417) and below BC on Collect misplaced items Gen (0.279 vs. 0.325), with overlapping error bars in several other rows (e.g., Cleaning A Car Gen). Regarding the ensemble, the experiments evaluate only ABIL-BC and ABIL-DT as two separate single-backbone configurations; there is no experiment in which multiple behavior modules are combined, so the 'ensemble' terminology in the abstract and Section 3.3 overstates the method as evaluated. The authors should either add significance tests and temper the per-task claims, or add an actual ensemble evaluation.","section":"Abstract; Section 3.3; Section 4"},{"comment":"Table 2 is internally inconsistent as presented. For several tasks (Making tea, Moving boxes to storage, Cleaning A Car, Throwing away leftovers, Putting away dishes, Sorting books, Laying wood floors, Watering houseplants, Cleaning shoes, Collect misplaced items, Organizing file cabinet), the Basic evaluation lists PDSketch as '> 5 minutes', yet the corresponding Gen rows either print numeric PDSketch values despite the Basic timeout (Making tea Gen shows 0.074±0.053) or appear to omit the PDSketch column entirely (Moving boxes Gen, Cleaning A Car Gen, and others list only four values for five methods). The authors should fix the table alignment and state explicitly whether a timed-out PDSketch run is counted as a failure (0%) or excluded, since the comparison in Section 4.2 hinges on this protocol.","section":"Table 2"}],"minor_comments":[{"comment":"The abstract and introduction state that ABIL 'generates predicate candidates' for perception, but Section 3.1 takes the predicate set P as given and Section 3.2 generates only pseudo-labels (abduced symbolic states) over a fixed predicate vocabulary; please reword to avoid implying predicate invention, or describe the candidate-generation mechanism if one exists.","section":"Abstract; Section 1"},{"comment":"The operator selection rule is defined by an existential condition ('exists k in [0,K)'), so if the perceived state satisfies the node conditions of multiple plan nodes simultaneously, op_t is not uniquely determined; the paper should state the deterministic selection rule used (e.g., first matching node in plan order) and confirm that it was followed in the experiments.","section":"Section 3.3, Eq. (2)"},{"comment":"The grounding-accuracy curves are plotted without error bars or standard deviations even though the text reports three seeds; adding them would let readers assess the 'less than 20% of their data' claim in Section 4.1 against the spread.","section":"Figures 4 and 6"},{"comment":"The paper does not state whether the grounding accuracy is measured on held-out demonstrations or on the same demonstrations used to generate the pseudo-labels; since the pseudo-labels are derived from the training demonstrations themselves, please specify the evaluation protocol to rule out a self-consistency measure.","section":"Figures 4 and 6"},{"comment":"The quantitative claim 'requiring less than 20% of their data' is not tied to a specific curve or threshold in Figure 4; please state the exact demonstration budgets and accuracy levels that support the 20% figure.","section":"Section 4.1"},{"comment":"In Table 6, ABIL-DT Pickup Basic at 90% grounding accuracy (0.860±0.017) is higher than at 100% (0.845±0.035), and at 50% accuracy two cells fall below the data-driven baseline (ABIL-BC Pickup Basic 0.717 vs. 0.723; ABIL-DT Putting-blocks-in-bowls 0.462 vs. 0.539); the robustness conclusion in B.1 should be qualified to the accuracy regime where it actually holds.","section":"Appendix B.1, Table 6"},{"comment":"The limitation statement (2) concedes dependence on an accurate and sufficient knowledge base, but the paper never tests robustness to KB errors (as opposed to perception errors in Table 6); a short experiment with corrupted or incomplete G would substantially strengthen the honest framing given in the conclusion.","section":"Section 5"},{"comment":"References [21] and [28] lack venue information (a CoRR preprint and an unpublished manuscript, respectively); please complete the entries before camera-ready.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"I would encourage the editor to ask the authors to release the exact knowledge-base files (the G graphs) used in each experiment, not just the code, because the central empirical contrast (ABIL vs. BC/DT, and especially the zero-shot tables) depends on precisely what G contains; the concern raised in Appendix D.2 is concrete enough that verification will require the actual graphs. The venue fit is appropriate, and the work has a solid empirical core, but the attribution and zero-shot issues should be resolved in the revision rather than deferred to supplemental material."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"ABIL is a serious neuro-symbolic imitation paper. The new bit is sequential-consistency abduction: instead of labeling each frame by hand, they solve a constraint over the trajectory given a symbolic state machine, producing pseudo-labels for the perception net. Then they train per-operator policies and select among them with the same symbolic plan. That's a clean way to get predicate grounding without annotations, and the empirical wins over BC/DT on long-horizon tasks (Put, Unlock, many Mini-BEHAVIOR tasks, CLIPort shape packing) are large and consistent. The PDSketch comparison is weakened by timeouts, but where it runs, ABIL is usually better. Credit where earned: the idea is clear, and the grounding-accuracy ablation (Table 6) shows the method degrades gracefully rather than collapsing. But there is a load-bearing problem the paper does not confront. The supplied knowledge base G is not a general task theory; it is an object-specific solution graph. Appendix D shows Throwing away leftovers with nodes Pick(hamburger_1,...), Pick(hamburger_2,...), Pick(hamburger_3,...) and edges chaining them. Training is on one hamburger, yet the KB already contains the three-hamburger plan. So the 'zero-shot' throw-2/3 result is written into the KB, not produced by the learner. Equation 1 then largely fits perception to a human-provided plan, and Eq. 2 reads the operator selection off the same graph. BC and DT receive no such privileged structure, so Tables 1-4 may be measuring the value of the KB rather than the abduction or imitation objective. The paper must state whether G is per-instance; if it is, the zero-shot compositional claim is unsupported, and if it is meant to be abstract, the appendix needs to show the abstraction and how it is instantiated for varying object counts. The other soft spots are minor by comparison: the 'policy ensemble' is never evaluated as an ensemble—they report ABIL-BC and ABIL-DT separately; data-efficiency curves have no error bars or significance tests; PDSketch success rates are missing for timed-out tasks. Section 5's limitation (2) concedes dependence on an accurate and sufficient KB, but no experiment tests robustness to KB errors or to an abstract KB. Table 6 perturbs perception accuracy, not G. The paper deserves a serious referee: the abductive pseudo-labeling step is a real contribution that people working on neuro-symbolic learning will want to build on. But the revision needs to fix the zero-shot design, clarify the KB scope, and either drop the stronger generalization claims or support them with abstract KB experiments. Send it to peer review, with the expectation of major revision.","headline":"The core idea is real and the gains are big, but the zero-shot result is likely pre-encoded in the supplied KB, so read the strong claims with caution.","tokens_in":734,"tokens_out":1700,"would_cite":true,"duration_ms":34681,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"ABIL uses abductive reasoning over a symbolic knowledge base to generate predicate labels from unannotated demonstrations, then selects per-operator imitation policies through symbolic planning, yielding better data efficiency and…","keywords":["Imitation Learning","Abductive Learning","Neuro-Symbolic Learning","Embodied Artificial Intelligence","Long-Horizon Planning","Symbolic Planning","Data Efficiency"],"falsifier":"Take a BabyAI task family and construct two versions of the knowledge base, one correct and one with the sequence of Pick and Open reversed, then train ABIL on identical demonstrations. If success rates and predicate accuracy stay essentially unchanged, the abductive sequential-consistency constraints are not doing the claimed labeling work; if they collapse, the knowledge-base dependence is confirmed.","tokens_in":18807,"feed_emoji":"🤖","tokens_out":6776,"duration_ms":56971,"temperature":0.7,"pith_summary":"The paper proposes ABIL, a framework that blends imitation learning with symbolic abductive reasoning so agents can solve long-horizon tasks from raw observations without any predicate-level annotations. ABIL starts from a human-supplied knowledge base—a state machine whose nodes are sub-task conditions and whose edges are logical operators—and uses abductive reasoning to infer which symbolic states each demonstration must have passed through, turning those inferences into pseudo-labels that train a perception module. The learned perception maps observations to symbols, a symbolic planner picks the next logical operator, and an ensemble of behavior policies, each trained for one operator, executes the low-level actions. In experiments on BabyAI, Mini-BEHAVIOR, and CLIPort, ABIL reports substantially higher data efficiency, stronger generalization, and zero-shot transfer to composed tasks than behavior cloning, Decision Transformer, and PDSketch. If the claim holds, imitation learning would no longer need ground-truth predicate annotations to benefit from symbolic planning's long-horizon robustness.","feed_headline":"Abduction lets imitation learning plan over long horizons","feed_subtitle":"A symbolic knowledge base generates labels and picks policies, so agents need fewer demos and generalize further.","key_machinery":"The load-bearing object is the knowledge base $G$, a directed graph state machine whose vertices are sets of ground atoms (sub-task conditions) and whose edges carry a logical operator $op$ with add/delete effects. From this graph ABIL derives 'sequential consistency' abduction rules—task completed, $a$ before $b$, either $a$ or $b$, and both $a$ and $b$ in any order—which turn an unlabeled trajectory into a constrained symbolic sequence. That sequence solves the joint optimization in Eq. 1, providing pseudo-labels for predicate classifiers $f_p$ at the object level. The same graph supports a symbolic plan, and Eq. 2 maps each perceived symbolic state to the current operator, selecting which imitation policy $h_{op}$ executes next. These four pieces—KB as state machine, sequential-consistency abduction, object-level predicate grounding, and operator-ensemble selection—together carry the argument for data-efficient and generalizable long-horizon imitation.","core_discovery":"ABIL's central claim is that the hidden symbolic structure of expert demonstrations can be recovered by abduction and then used to make imitation learning compositional. Given a knowledge base $G$ and a trajectory of raw observations, the framework alternates between a perception network $f$ and a symbolic sequence $\\{z_t\\}$: Eq. 1 chooses the $z_t$ that are simultaneously close to $f(s_t)$ and consistent with $G$ under the 'sequential consistency' principle, and those $z_t$ are the pseudo-labels that train $f$. Each predicate is grounded by an object-level classifier, so the perception module learns to detect task-relevant symbols without human annotations. Once observations map to symbolic states, Eq. 2 selects the logical operator $op_t$ from the symbolic plan for the current state, and Eq. 4 trains one behavior policy per operator by imitation, with the operator's arguments coming from Eq. 3. The authors argue that this operator-level decomposition is what lets the agent handle long horizons, novel object counts, unseen distractors, and composed tasks, because reasoning and imitation happen at the level of abstract operators rather than raw action sequences.","pith_inferences":["Beyond the paper, a natural extension is to let a vision-language model propose predicate candidates instead of relying on the hand-built knowledge base, which would make ABIL's abductive labeling loop semi-automatic.","ABIL's dependence on $G$ suggests a diagnostic use: demonstrations that cannot be mapped onto $G$ could be flagged automatically, turning knowledge-base incompleteness into a detectable signal for acquiring new operators.","The paper evaluates at most a handful of operators per task; whether the one-policy-per-operator scheme and the graph planner scale to hundreds of operators is an untested question that a larger benchmark could answer.","The paper's determinism and full-observability assumptions point to a belief-space version of ABIL; if that extension is attempted, the abduction constraints would need to be reformulated for distributional symbolic states."],"forward_implications":["Without any predicate annotations, an agent can be trained from raw demonstrations and still ground the task-relevant symbols that symbolic planning needs.","Much less data is needed for reliable neuro-symbolic grounding: in BabyAI, ABIL reaches near-perfect predicate accuracy with less than 20% of the demonstrations PDSketch uses.","Long-horizon tasks that break model-based planners become tractable, including Mini-BEHAVIOR tasks of roughly 45 to 106 decision steps where PDSketch times out.","Agents can zero-shot compose learned operators, as when policies trained on pickup/open or on throwing one leftover transfer to the unlock task or to throwing two or three leftovers.","Inference stays close to learning-method speed while gaining symbolic-planning generality, because search happens at the operator level rather than in raw observation-action space."],"supporting_citations":[{"why":"Supplies the Behavior Cloning baseline against which ABIL's data efficiency and generalization are measured.","marker":"[1]"},{"why":"Supplies the Decision Transformer baseline used both directly and as an ABIL backbone.","marker":"[4]"},{"why":"Provides the BabyAI benchmark and its tasks, grounding the long-horizon and generalization experiments.","marker":"[5]"},{"why":"Provides the abductive learning framework that ABIL borrows for pseudo-label generation by reasoning back from constraints.","marker":"[6]"},{"why":"Provides the Neural Logic Machine encoder shared by all compared methods for fair evaluation.","marker":"[7]"},{"why":"Provides Mini-BEHAVIOR, the long-horizon household benchmark where PDSketch times out and ABIL generalizes.","marker":"[21]"},{"why":"Represents prior neuro-symbolic planning that needs feasibility annotations and transition data, motivating ABIL's annotation-free abduction.","marker":"[22]"},{"why":"Supplies PDSketch, the main neuro-symbolic baseline and the model-based planning approach ABIL outperforms on long horizons.","marker":"[24]"},{"why":"Provides CLIPort, the robotic manipulation benchmark with continuous actions and oracle-collected demonstrations.","marker":"[29]"},{"why":"Represents the prior regression planning network that requires predicate-level annotations, the gap ABIL removes.","marker":"[36]"}],"fun_headline_variants":["ABIL: Abductive imitation recovers symbols for long-horizon plans","Abductive imitation breaks long tasks into symbolic steps","Symbolic abduction makes imitation learning long-horizon","Neuro-symbolic abduction makes imitation learn composable plans","Abduction enables long-horizon imitation from fewer demos"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the expert-supplied knowledge base is accurate and sufficient for every demonstration in the task family; if the knowledge base is wrong, incomplete, or inconsistent with what the expert actually does, the abduced pseudo-labels and the operator selections built on them are wrong too.","fun_headline_variants_meta":{"raw":{"variants":["ABIL: Abductive imitation recovers symbols for long-horizon plans","Abductive imitation breaks long tasks into symbolic steps","Symbolic abduction makes imitation learning long-horizon","Neuro-symbolic abduction makes imitation learn composable plans","Abduction enables long-horizon imitation from fewer demos"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000815,"raw_usage":{"total_tokens":3626,"prompt_tokens":1051,"completion_tokens":2575,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":667,"completion_tokens_details":{"reasoning_tokens":2492}},"tokens_in":667,"tokens_out":2575,"duration_ms":16792,"temperature":1.0,"reasoning_tokens":2492,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T11:24:24.413392+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a BabyAI task family and construct two versions of the knowledge base, one correct and one with the sequence of Pick and Open reversed, then train ABIL on identical demonstrations. If success rates and predicate accuracy stay essentially unchanged, the abductive sequential-consistency constraints are not doing the claimed labeling work; if they collapse, the knowledge-base dependence is confirmed.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the Behavior Cloning baseline against which ABIL's data efficiency and generalization are measured."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the Decision Transformer baseline used both directly and as an ABIL backbone."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the BabyAI benchmark and its tasks, grounding the long-horizon and generalization experiments."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the abductive learning framework that ABIL borrows for pseudo-label generation by reasoning back from constraints."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the Neural Logic Machine encoder shared by all compared methods for fair evaluation."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Represents prior neuro-symbolic planning that needs feasibility annotations and transition data, motivating ABIL's annotation-free abduction."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies PDSketch, the main neuro-symbolic baseline and the model-based planning approach ABIL outperforms on long horizons."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides CLIPort, the robotic manipulation benchmark with continuous actions and oracle-collected demonstrations."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Represents the prior regression planning network that requires predicate-level annotations, the gap ABIL removes."}],"review_version":1}