{"id":"4f642735-6bc9-4529-9305-e94050bb2082","arxiv_id":"2412.11499","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"DeDer distills LLM chain-of-thought reasoning into a two-tier small-language-model policy (rationale writer plus planner) and reports state-of-the-art ALFRED success rates for small-model embodied agents.","lead":"DeDer is a framework that distills embodied reasoning from large language models into two small-language-model policies, one that writes rationales and one that turns them into action plans. On the ALFRED benchmark the distilled small models beat several larger language-model planners and prior distillation methods, while running in under a second on consumer GPUs.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"DeDer's reported gains depend on an embodied KG whose triples are never derived from raw observations; if the KG comes from simulator state, the off-the-shelf claim and the sLM-versus-PaLM comparison are unsupported.","rationale":"The reader's weakest_assumption identifies exactly the KG-acquisition gap, and I agree it is the load-bearing point. Even granting the two-tier distillation design, the contrastive attention module, and the internally coherent ablations, the paper's own equations make g the only input to the reasoning-policy, so the source of g determines whether DeDer is an embodied agent or an oracle-fed lookup. The paper is incomplete at a load-bearing point, but the concern is addressable by a described perception module and by released code, so conditional acceptance remains the appropriate disposition rather than rejection. Secondary issues, such as the uncontrolled trained-versus-few-shot comparison against LLM-planner-PaLM and an internally inconsistent row in Table A.14 (t5-small with gpt2-large shows Unseen Environment SR 67.5 and GC 39.8, an impossible SR/GC ordering), reinforce the need for greater scrutiny and reproducibility but do not replace the primary concern. No ad hominem is intended; the critique concerns the argument's assumptions, not the authors.","tokens_in":22469,"tokens_out":6182,"duration_ms":61105,"concrete_test":"Require the authors to specify and implement a perception-to-triple pipeline for U and V, then rerun the Unseen Spatial and Unseen Environment splits of Table 1 with triples obtained only from the agent's raw observations (RGB, depth, and segmentation masks) through that pipeline. If the observed-KG SR drops to or below the SCoTD/SCOTT baselines, or if the pipeline cannot be constructed without accessing simulator object positions or containment state, the central claim is falsified.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Equations (4)-(7) make the embodied KG g the sole input to the reasoning-policy: Phi_R : g -> R, with g updated as U : (g_{t-1}, a_{t-1}, o_t) -> g_t. The paper never defines a function from raw ALFRED observations (RGB/depth images) to triples. Appendix D's example lists Observation o as 'Mug, Spatula, Apple, Fork ...' and retrieved facts '(Mug, in, Sink), (Agent, at, Sink), (Agent, CanSee, Mug), (Agent, PickedUp, Fork)'; these encode object classes, locations, containment, agent position, and action history, none of which are directly readable from an egocentric RGB frame. If U or V is populated from simulator/ground-truth state, then at every planning step DeDer receives privileged environment information that the text-only LLM baselines (LLM-planner-PaLM, ZSP, SayCan) do not receive. Because Phi_R (Eq. 7) consumes g, not o, all task-relevant grounding enters through this unspecified KG. The reported 81.8/52.7/40.3 SR could then reflect access to a perfect state-perception oracle rather than distillation of embodied reasoning into sLMs, and the deployment claim on capacity-limited devices is unsupported. This is a gap, not a demonstrated error: the paper may intend a vision-based triple extractor, but no such component is described, and no code/data are released to check.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces DeDer, a framework for distilling embodied chain-of-thought reasoning from a large language model into two small language models: a T5-small-based reasoning-policy that generates rationales from an embodied knowledge graph, and a GPT-2-based planning-policy that maps rationales to action plans. The rationale dataset is built by prompting an LLM with MDP-featured in-context learning and self-critic filtering over 312 expert ALFRED trajectories. The reasoning-policy is trained with a rationale reconstruction loss and a behavior-based contrastive loss, while the planning-policy is trained with a plan reconstruction loss. Evaluation on ALFRED reports success rates of 81.8% (Seen), 52.7% (Unseen Spatial), and 40.3% (Unseen Environment) for DeDer-GPT2, outperforming both LLM-based planners such as LLM-planner-PaLM and prior distillation baselines SCoTD and SCOTT, with sub-second inference on RTX 3090/3050 GPUs.","tokens_in":22768,"tokens_out":3886,"duration_ms":36455,"significance":"If the reported results are taken at face value and the missing perception component is properly specified, the paper would make a noteworthy empirical contribution: it would show that sLM-based policies, after targeted distillation, can outperform much larger LLM planners on a complex embodied benchmark while running on off-the-shelf hardware. The two-tier decomposition, the contrastively prompted attention for concurrent rationale generation, and the self-verification pipeline for rationale dataset construction are all sensible and useful ideas. On the negative side, the paper releases no code or data, the provenance of the embodied knowledge graph is unspecified, the headline Unseen Environment result rests on only 58 tasks, and there is a numerical inconsistency between Table 5 and Table A.14. These issues currently prevent the central claim from being fully verified.","major_comments":[{"comment":"The embodied KG is the sole input to the reasoning-policy Phi_R in Eq. (7), and Eq. (5) defines an update U : (g_{t-1}, a_{t-1}, o_t) -> g_t, but the paper never specifies how raw ALFRED observations (RGB-D frames) are converted into triples such as (Apple, On, Table) or (Agent, at, Sink). Appendix D lists observations as object-name lists and retrieved facts that encode agent position, containment, visibility, and action history, none of which are directly readable from an egocentric frame. If these triples come from simulator ground truth or an external perception oracle, then at every planning step DeDer receives privileged state information that the text-only LLM baselines do not receive, which would undermine both the comparison with PaLM/LLaMA2 planners and the claimed deployment on off-the-shelf devices. The authors must either describe a concrete perception-to-triple mechanism or explicitly state the assumption that such oracle information is available; without this, the central claim is unsupported.","section":"§4.2, Eqs. (5)-(7); Appendix D"},{"comment":"There is a direct numerical inconsistency in the capacity-scaling experiments. Table 5 reports Unseen success rate 46.1±1.1 for the t5-small + gpt2-large configuration, while Table A.14 reports 67.5±1.7 for SR and 39.7±2.0 for GC in the corresponding row. The GC value being lower than SR in that row is also internally inconsistent with the GC>=SR pattern in every other row. This discrepancy affects the conclusion in §5.3 that the reasoning-policy capacity is the main driver of generalization. The authors need to correct the numbers and ensure Tables 5 and A.14 are consistent.","section":"Table 5 vs. Table A.14"},{"comment":"The headline Unseen Environment result is based on only 58 tasks, and the reported confidence intervals (±0.9 SR for DeDer-GPT2) appear to reflect variance across three training seeds only, not the variance across tasks within the category. With 58 tasks, the standard error across task sampling would be much larger. The paper should report task-level bootstrap confidence intervals or otherwise justify why the tiny reported intervals are appropriate. This is important because the central claim of superiority over SCoTD and SCOTT in the Unseen Environment column rests on differences of roughly 13-16 percentage points, which may or may not be significant under a correct statistical treatment.","section":"§5.1, Table 1; Unseen Environment evaluation"}],"minor_comments":[{"comment":"The KG retriever threshold delta and the contrastive margin epsilon in Eq. (12) are never assigned values; only alpha is given (0.5 in Appendix B.4.2). Please report the actual values used for delta and epsilon in all experiments.","section":"§4.2, Eq. (6) and Appendix B.4"},{"comment":"The method name is written inconsistently as DEDER, DeDer, and DEDER-GPT2; please standardize the notation.","section":"Throughout"},{"comment":"The caption refers to 'Phi_D' but the paper defines the planning-policy as Phi_P; please correct the symbol.","section":"§5.3, Table 5 caption"},{"comment":"The x-axis labels in Figure 4 are not legible in the provided version; please add readable query indices and a clear legend.","section":"Figure 4"},{"comment":"It is unclear whether the PaLM-based baselines (SayCan-PaLM, LLM-planner-PaLM, ZSP-PaLM) in Table 1 are from the original papers or reimplemented with the hyperparameters in Table A.7; please clarify the source of each baseline number.","section":"Appendix B.1"},{"comment":"The sentence 'In evaluating in off-the-shelf devices, we adopt sLMs for these language planning baselines' is ambiguous, since Table 1 also includes PaLM and LLaMA2 variants; please state clearly which baselines use which underlying model and how the affordance/grounding inputs are provided to each.","section":"§5.1, Baselines and Table 1"}],"recommendation":"major_revision","confidential_remarks":"The embodied-KG provenance problem is the main correctness risk and lands exactly as the stress-test note suggests. I do not think this warrants rejection, because the gap might be fixable by specifying a real perception module or by explicitly weakening the deployment claim, but it is a load-bearing missing component, not a cosmetic flaw. The numerical inconsistency between Tables 5 and A.14 is also a serious editorial issue that must be corrected before the paper can be accepted. Please also ask the authors to address the small test-set statistics for the Unseen Environment category."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear Colleague,\n\nQuick take: DeDer is a reasonable distillation framework with a neat two-tier split, but the paper's headline numbers are not yet believable because the embodied knowledge graph—the sole perceptual input to the policy—is never derived from raw observations. If the triples come from ALFRED's simulator state, the comparison against text-only LLM planners is unfair and the \"off-the-shelf\" claim collapses. That's a gap, not a proven error, but it's the load-bearing one.\n\nWhat's genuinely new: the separation of a reasoning-policy (T5-small) that outputs rationales from a planning-policy (GPT-2) that outputs actions, with the reasoning-policy prompted by an embodied KG and trained with contrastive attention to produce multiple rationales in one pass. The design is sensible: keep the expensive part small and structured, let the planner be a simple conditional model. The ablations show the KG and the contrastive learning each add something, and the inference-time measurements (0.65s on an RTX 3090) are concrete.\n\nWhere it wobbles: first, the KG. Equations (5)-(7) make g the only input to the reasoning-policy, but the paper never defines how an RGB observation becomes triples like (Mug, in, Sink). The example in Figure A.8 lists object classes, containment, agent position, and action history—none of which are directly readable from an egocentric frame. If U or V uses ground-truth simulator state, then every step receives privileged information that SayCan, ZSP, and LLM-planner don't, and the 40.3% Unseen Environment SR may reflect a perfect perception oracle, not distilled reasoning. This is the key thing to fix.\n\nSecond, the comparison to LLM-planner-PaLM is few-shot vs. trained; that's an apples-to-oranges headline. The same-size distillation baselines (SCoTD, SCOTT) are more relevant, but they also don't get an oracle KG, so the gap there is confounded too.\n\nThird, Table A.14 has an impossible row: t5-small with gpt2-large reports Unseen Environment SR 67.5 and GC 39.7. GC should be at least SR if tasks are a conjunction of subgoals. Likely a typo, but it makes you wonder about the other numbers.\n\nNo code or data are released, so I can't check. These are all fixable. If the KG is actually built from vision, the authors need to describe that module and preferably release it; if it's simulator state, the claims need to be scaled way back.\n\nVerdict: worth a serious referee. The idea has legs, the write-up is coherent, and the problem is important. But the current evidence does not support the strong claims, and the KG gap is exactly what a careful reviewer should push on.\n\nBest.","headline":"DeDer has a sensible two-tier distillation idea, but the headline ALFRED numbers rest on an unspecified embodied KG that may be oracle state, so the strong claims should not be taken at face value.","tokens_in":23356,"tokens_out":3662,"would_cite":false,"duration_ms":31688,"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":"DeDer claims embodied planning can be distilled from PaLM-scale LLMs into two small models that run in under a second on off-the-shelf GPUs, outperforming the teacher-scale planner on ALFRED.","keywords":["embodied AI","chain-of-thought distillation","small language models","two-tier policy","embodied knowledge graph","ALFRED benchmark","zero-shot planning"],"falsifier":"Run DeDer on ALFRED feeding it raw RGB-D observations only, with the embodied knowledge graph built by a perception module trained on separate data instead of taken from simulator state. If success rates drop toward the end-to-end baseline that performs the same task without the graph, then the knowledge graph, not the distilled reasoning, is the source of the reported performance, and the claim that sLM policies can plan on off-the-shelf devices is falsified.","tokens_in":22220,"feed_emoji":"🤖","tokens_out":8110,"duration_ms":59050,"temperature":0.7,"pith_summary":"DeDer claims that a large language model's embodied planning ability can be distilled into two very small language models that run on ordinary hardware in under a second per decision. The framework splits policy-making into a reasoning tier, which emits rationales about goals, states, observations, and sub-goals, and a planning tier, which turns those rationales into the next action. On the ALFRED household benchmark, the two small models (a T5-small reasoner plus a GPT-2 planner, about 160 million parameters in total) report higher success rates than LLM-based planners such as PaLM-540B, and higher rates than the prior distillation methods SCoTD and SCOTT. If correct, embodied agents would no longer need to query a large model at runtime, which is what would make complex planning feasible on capacity-limited, off-the-shelf devices. The load-bearing ingredient is an embodied knowledge graph of semantic triples that the reasoning tier reads before producing its rationales.","feed_headline":"Two tiny language models beat a 540B-parameter planner","feed_subtitle":"On ALFRED household tasks, a T5+GPT-2 pair reports higher success than PaLM-540B in under a second per step.","key_machinery":"The load-bearing object is the embodied knowledge graph: a growing set of semantic triples of the form (subject, relation, object), for example (Apple, On, Table) or (Agent, Pickup, Knife), which is refreshed at every step by an update function $U$ and filtered by a retriever $V$ before being presented to the reasoning policy. This graph converts the agent's interaction history and current observation into a compact prompt that a small language model can actually read. The reasoning policy is an encoder-decoder with soft prompt pools, a causal attention module, and a gated attention module, trained jointly with a rationale-reconstruction loss and a behavior-based contrastive loss defined on pairs of graphs that lead to the same action. The planning policy is a separate decoder trained to generate the next plan conditioned on the rationales, so the whole pipeline produces a plan in one step rather than iteratively.","core_discovery":"The paper's central claim is that decomposing an LLM's decision process into a two-tier hierarchy makes it possible to distill embodied planning into small language models that match or exceed the teacher-scale LLM planners. DeDer writes the policy as $\\Phi_{\\mathrm{sLM}} = \\Phi_P \\circ \\Phi_R$, where the reasoning policy $\\Phi_R$ (a T5-small encoder-decoder with causal and gated attention) produces a set of rationales in a single forward pass, and the planning policy $\\Phi_P$ (a GPT-2 decoder) converts those rationales into an executable action plan. The authors report that DeDer-GPT2 reaches 81.8% success on seen tasks, 52.7% on tasks with randomized object layouts, and 40.3% on tasks in entirely new environments, compared with 66.8%, 33.6%, and 17.2% for LLM-planner-PaLM on the same categories, all evaluated zero-shot. They also report 0.65 seconds of inference time on an RTX 3090 and 1.16 seconds on an RTX 3050, against 9.37 seconds for an LLM-planner variant running on the same hardware class.","pith_inferences":["Editorial inference: the paper never specifies who converts raw observations into the semantic triples of the embodied knowledge graph; if the triples come from the ALFRED simulator's ground-truth state rather than from perception, the comparison against LLM planners may reward access to privileged information rather than planning skill.","Testable extension: run DeDer with the same small models but with triples produced by an off-the-shelf vision-language perception module; the drop in success rate would quantify how much of the reported performance depends on the assumption that the graph is already correct.","The contrastive loss on graph pairs is a transferable idea: forcing embeddings of graphs that lead to the same action to be close should make any compact re-planning module more stable under small environment changes, outside ALFRED as well."],"forward_implications":["Embodied planning can run entirely on small models: roughly 160 million parameters total, with sub-second inference on consumer GPUs, so no large model needs to be queried during task execution.","The two-tier split transfers better than end-to-end distillation: at the same GPT-2 capacity, DeDer reports 81.8% seen success versus 33.1% for the single-tier End2End baseline.","Reasoning capacity matters more than planning capacity: growing the T5 reasoner from small to large raises unseen-environment success from 40.3% to 47.7%, while growing the GPT-2 planner leaves performance essentially flat.","The reported zero-shot results in unseen environments imply that the distilled rationales carry knowledge that generalizes beyond the expert dataset's scenes, object layouts, and task descriptions."],"supporting_citations":[{"why":"Supplies the ALFRED benchmark and evaluation protocol on which all reported success rates are measured.","marker":"(Shridhar et al., 2020)"},{"why":"LLM-planner, the primary language-planning baseline that DeDer claims to surpass in seen and unseen settings.","marker":"(Song et al., 2023)"},{"why":"Provides PaLM, the teacher LLM used for rationale extraction and the 540B-parameter comparison point.","marker":"(Chowdhery et al., 2023)"},{"why":"SCoTD, the symbolic chain-of-thought distillation baseline whose performance DeDer compares against.","marker":"(Li et al., 2023)"},{"why":"SCOTT, the self-consistent chain-of-thought distillation baseline used as a comparison.","marker":"(Wang et al., 2023)"},{"why":"Defines chain-of-thought prompting, the reasoning mechanism that DeDer distills into the sLM policies.","marker":"(Wei et al., 2022)"},{"why":"Supplies the retrieval-augmented in-context learning mechanism used to fetch examples for rationale extraction.","marker":"(Ram et al., 2023)"},{"why":"Provides T5, the encoder-decoder backbone of the reasoning policy.","marker":"(Raffel et al., 2020)"},{"why":"Provides GPT-2, the decoder backbone of the planning policy.","marker":"(Radford et al., 2019)"},{"why":"SayCan, the LLM-plus-affordance baseline used for comparison.","marker":"(Brohan et al., 2023)"}],"fun_headline_variants":["Tiny T5+GPT-2 pair outplans PaLM-540B on ALFRED","DeDer: small models, big win over LLM planners","sLM duo outperforms 540B planner on ALFRED","Distill LLM reasoning to two tiny models, win ALFRED"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The paper assumes that an embodied knowledge graph written as semantic triples such as (Apple, On, Table) is available at every planning step, but it does not describe how raw observations are turned into those triples; if that conversion needs an external perception system or ground-truth simulator state, the claimed deployment on off-the-shelf devices is not supported.","fun_headline_variants_meta":{"raw":{"variants":["Tiny T5+GPT-2 pair outplans PaLM-540B on ALFRED","DeDer: small models, big win over LLM planners","sLM duo outperforms 540B planner on ALFRED","Distill LLM reasoning to two tiny models, win ALFRED"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00079,"raw_usage":{"total_tokens":3530,"prompt_tokens":1038,"completion_tokens":2492,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":654,"completion_tokens_details":{"reasoning_tokens":2409}},"tokens_in":654,"tokens_out":2492,"duration_ms":16878,"temperature":1.0,"reasoning_tokens":2409,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T14:51:57.182217+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run DeDer on ALFRED feeding it raw RGB-D observations only, with the embodied knowledge graph built by a perception module trained on separate data instead of taken from simulator state. If success rates drop toward the end-to-end baseline that performs the same task without the graph, then the knowledge graph, not the distilled reasoning, is the source of the reported performance, and the claim that sLM policies can plan on off-the-shelf devices is falsified.","supporting_citations":[],"review_version":1}