{"id":"ae2652d1-179f-4505-a993-d385757c2bb5","arxiv_id":"2608.11498","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A driving Q-learning agent trained on language-structured scenarios develops threat-focused attention but its actions stay close to a constant baseline, a failure formalized as the recognition-control gap.","lead":"This paper tests whether training a driving policy on language-generated traffic scenarios makes it truly react to threats. It finds the policy learns where the danger is, but not reliably what to do about it, a pattern it calls the recognition-control gap.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Language-structured claim is confounded: D_rand omits the scripted adversarial interaction, so success/attention gains may reflect adversarial training content rather than language; a programmatic-control experiment is required.","rationale":"The paper is an honest empirical study with a robust negative result: trained policies plateau near the best constant action, and the decision-frequency ablation shows real sensitivity to control frequency. However, the positive claim that 'language-structured training' improves success and threat attention rests on the comparison between D_lang and D_rand (Section 3.1). D_rand differs from D_lang not only by lacking language-derived structure but also by omitting the scripted adversarial interaction program entirely. Thus the observed 6-point success gain and 2.1× vs 1.2× attention ratio could be reproduced by any distribution with coherent adversarial events, regardless of whether an LLM was involved. The paper's own statement that the parser is replaceable (Section 4.1) supports this concern. The reader's verdict (CONDITIONAL) already identifies this as the weakest assumption; I agree and recommend no change to the verdict, but the condition should explicitly include a programmatic-interaction control experiment. The policy-collapse and recognition-control gap findings are less affected by this confound and remain valuable.","tokens_in":10289,"tokens_out":8984,"duration_ms":83897,"concrete_test":"Train ERQ-Net on D_prog: the same four interaction programs and parameter ranges as D_lang, but generated by random sampling without any LLM. Evaluate on the same 500 language-structured test scenarios and report success and Athreat. If D_prog matches D_lang (55–58%, ~2.1× attention), the gains are due to adversarial interaction structure, not language. If D_prog matches D_rand (49–52%, 1.2×), the LLM's language-to-parameter mapping is causal. This isolates the confound in Section 3.1.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The most load-bearing assumption is that D_rand (Section 3.1) is a valid counterfactual for isolating language-induced structure. It is not. D_rand matches marginal ranges of vehicle count, position, lane, and speed, but omits the scripted adversarial interaction program (cut-in, sudden braking, overtaking, tailgating). D_lang therefore differs from D_rand in two conflated ways: (i) scenarios are instantiated from LLM-parsed language descriptions, and (ii) they contain a coherent adversarial interaction, whereas D_rand has none. The reported gains — success 55–58% vs 49–52% and attention 2.1× vs 1.2× (Tables 2–3) — can be fully explained by training on any distribution containing structured adversarial events, irrespective of language. The paper even notes that the parser is 'replaceable by any model that emits schema-valid JSON' (Section 4.1), implying the causal content lives in the interaction program, not in the language itself. Without a third training distribution that uses the same interaction programs and parameter ranges but is generated programmatically (no LLM), the 'language-structured' attribution is underdetermined. The recognition–control gap and policy-collapse results are less affected by this confound, but the paper's central positive claim is not established.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Language-Structured Relational Q-Learning, instantiated through an Ego-Centric Relational Q-Network (ERQ-Net), to study whether training on language-structured traffic scenarios produces threat-aware and adaptive control policies. Natural-language descriptions are parsed into schema-valid configurations and executable interaction programs (cut-in, sudden braking, overtaking, tailgating), but the policy observes only kinematic graph features, not language or semantic roles. The authors compare ERQ-Net trained on language-structured scenarios (D_lang) against a random-control distribution (D_rand), report a six-point success improvement and an increase in adversary-focused attention, and then show that trained policies remain comparable to the best constant action while a portfolio of simple policies solves 76% of scenarios. They formalize this as a recognition-control gap and investigate reward shaping, decision frequency, encoder ablations, scenario quality, and state-interface transfer to CARLA.","tokens_in":10638,"tokens_out":4072,"duration_ms":36306,"significance":"If the central attribution were established, the paper would make a useful contribution to language-conditioned scenario generation and to evaluating whether representation-level threat awareness translates into control. The manuscript is unusually honest: it reports the policy-collapse result as a first-class finding, includes constant-action and policy-portfolio baselines, and provides a decision-frequency ablation that gives a concrete, testable mechanism for the recognition-control gap. The CARLA transfer study also draws a careful and appropriate distinction between state-interface transfer and perception-level sim-to-real transfer. However, the positive claim that language structure, rather than the presence of coherent adversarial interaction programs, drives the observed gains is underdetermined by the current experimental design. The recognition-control-gap analysis is more robust than the headline success improvement and should be framed as the main contribution.","major_comments":[{"comment":"The random-control distribution D_rand is not a valid counterfactual for isolating language-induced structure. D_lang and D_rand differ by two factors at once: scenarios are instantiated from LLM-parsed language descriptions, and they contain a coherent adversarial interaction program, whereas D_rand samples vehicle count, position, lane, and speed independently and omits the structured interaction. The reported 55-58% versus 49-52% success gain and the 2.1x versus 1.2x attention ratio could therefore be produced by any training distribution that contains structured adversarial events, regardless of language. Since Section 4.1 states that the parser is replaceable by any model that emits schema-valid JSON, the causal content may reside in the interaction program rather than in the language. A third training condition using the same interaction programs and parameter ranges generated programmatically (no LLM) is needed to attribute the gains to language structure; without it, the paper's central positive claim is underdetermined.","section":"Section 3.1 and Tables 2-3"},{"comment":"The threat-attention ratio Athreat is computed from the model's own final-layer attention coefficients, and the paper interprets the increase from 1.2x to 2.1x as evidence of 'emergent threat awareness.' This interpretation is partially circular: attention is jointly optimized for Q-value estimation, so preferential weighting of the adversarial actor can be a byproduct of value computation rather than evidence of a semantically meaningful threat representation. The claim needs external validation, for example by comparing attention to ground-truth relevance masks derived from the interaction program, or by an intervention showing that high attention to the adversarial node is causally necessary for the observed success (e.g., clamping or ablating that attention). As written, the awareness claim is not independently supported.","section":"Eq. (4) and Table 3"},{"comment":"The headline six-point improvement is reported as ranges over three seeds without confidence intervals or significance tests. Although the reported min-max ranges do not overlap, three seeds is a small sample, and the same 500-scenario test set is evaluated across seeds, so a paired test or a bootstrap over scenarios should accompany the central comparison. This matters because the paper's main positive claim rests on this difference; the current presentation does not distinguish a stable effect from seed noise.","section":"Section 5.1 and Tables 2-3"},{"comment":"The recognition-control gap at each decision frequency is computed as the shortfall from the 76% portfolio union reported in Table 2, which was obtained at 1 Hz. If the simple policies in the portfolio also benefit from higher decision frequency, the 76% ceiling is not constant across frequencies, so the reported shrinkage of the gap from 18 to 8 points may overstate the improvement attributable to decision frequency. The portfolio union should be recomputed at 2, 5, and 10 Hz before concluding that temporal resolution closes the gap.","section":"Table 6"}],"minor_comments":[{"comment":"The margin-shaping term r_margin(dfront) is not defined precisely; its functional form should be stated for reproducibility.","section":"Section 3.3, Eq. (8)"},{"comment":"The 'Best constant action (SLOWER)' row reports a single value (57%) while neighboring rows report ranges over seeds; please report the seed range for this baseline as well.","section":"Table 2"},{"comment":"The manuscript names the Gemini API but does not give the model version, temperature, or other decoding parameters; these details are needed for reproducibility of the parser results.","section":"Section 4.1"},{"comment":"The figure caption mentions night settings, but the table reports only clear, heavy rain, and dense fog; please clarify whether a night condition was actually evaluated.","section":"Figure 4 caption"},{"comment":"The indicator function is rendered as a boxed symbol in the text; use \\mathbb{1} consistently to avoid confusion.","section":"Eq. (5)"},{"comment":"Please state explicitly that the reported success values are min-max ranges over three seeds, not confidence intervals, so that readers do not overinterpret the six-point gap.","section":"Table 3 caption"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is careful and honest about its negative results, and the policy-collapse analysis is solid. The main obstacle is the confounded control condition for the headline language-structure claim; this is fixable with an additional programmatic-control experiment, so I do not view it as an unresolvable error. I would encourage the editor to treat the recognition-control gap and the decision-frequency analysis as the paper's strongest contributions."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know. The most durable contribution here is the recognition-control gap, not the language result. And the language-structured training claim, as presented, is underdetermined. I'd send the paper to a serious referee, but the positive headline should be downgraded until the missing control is run.\n\nThe gap is real and well argued. Training on these scenarios gets the policy to attend to the adversarial actor (attention ratio 2.1x) yet the trained policy barely matches the best constant action (55-58% vs 57%), while a portfolio of simple policies solves 76%. That exposes a failure mode representation-quality improving while control stagnates that is worth the field's attention. The decision-frequency ablation strengthens it: success climbs from 55-58% at 1 Hz to 66-69% at 10 Hz, and the gap shrinks to 7 points. That is a clean, reproducible-sounding result, and the authors resist overclaiming it. The paper is also honest about the limits of the CARLA transfer.\n\nNow the soft spots. The strongest positive claim, that language structure causes the 6-point success gain and the 2.1x attention gain, is not established. D_rand matches marginal ranges of vehicle count, position, lane, speed, but omits the scripted adversarial interaction. So D_lang differs from D_rand in two ways at once: it has language-instantiated scenarios and it has coherent adversarial events. The gains could come entirely from training on any distribution with structured adversaries. The authors themselves say the parser is replaceable by any model emitting schema-valid JSON, which pretty much concedes that language per se is not the causal ingredient. A programmatic control, same interaction programs and parameter ranges but no LLM, is needed before the language attribution holds. Minor: the threat-attention metric is internal attention, but the target adversary is externally defined by the scenario spec, so it is not fully circular. Still, the attention numbers lack confidence intervals and are reported as ranges over three seeds. No code, data, or exact prompt templates are supplied, which limits reproducibility, despite the paper using the word \"reproducible\" for the policy-collapse analysis.\n\nThe negative results stand up; the positive language claim does not. I would recommend peer review with a request for the programmatic control, per-seed scatter or CIs, and public artifacts. The recognition-control gap alone is worth a serious referee's time.","headline":"The recognition-control gap is a genuinely useful diagnostic, but the language-attribution claim is confounded and needs a programmatic control before it is taken at face value.","tokens_in":11064,"tokens_out":2278,"would_cite":false,"duration_ms":23390,"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":"Language-structured training gives a driving policy threat awareness, but not the ability to choose the right maneuver.","keywords":["autonomous driving","safety-critical scenarios","relational Q-learning","graph attention","language-conditioned scenario generation","recognition-control gap","threat awareness","policy collapse"],"falsifier":"Measure the criticality score of the random-control training distribution under the same non-reactive reference ego used for the language-structured set—the collision rate when the ego drives at constant velocity with no avoidance. If the random-control criticality is substantially below the 0.75 reported for the language-structured set, the six-point success gain and 2.1x attention gain could reflect scenario difficulty rather than language-induced structure, and the counterfactual does not isolate semantics.","tokens_in":10049,"feed_emoji":"🚗","tokens_out":7372,"duration_ms":59492,"temperature":0.7,"pith_summary":"This paper asks what an autonomous-driving policy actually learns when its training scenarios are generated from natural-language descriptions—cut-ins, sudden braking, overtaking, tailgating—while the language itself is hidden at test time. It proposes an ego-centric relational Q-network (ERQ-Net) that reads a dynamic traffic graph and reports that language-structured training improves test success from 49–52% to 55–58% and boosts attention to the adversarial vehicle from 1.2x to 2.1x. The central finding, however, is a failure mode the paper names the recognition–control gap: the trained policy performs about as well as always choosing 'SLOWER', while a portfolio of twelve simple policies solves 76% of the same scenarios. The authors argue this shows emergent threat awareness does not automatically translate into adaptive control, and they show that reward reweighting and margin shaping do not close the gap, while raising the control frequency from 1Hz to 10Hz shrinks it from 18 to 7 percentage points.","feed_headline":"Threat-aware driving policies still collapse to a constant action","feed_subtitle":"Language-structured training lifts threat focus to 2.1x, yet a fixed action matches it; 12 simple policies solve 76%.","key_machinery":"The load-bearing object is the Ego-Centric Relational Q-Network (ERQ-Net), a graph-attention value network. At each timestep the scene becomes a k-nearest-neighbour traffic graph (k=5, 100m radius) whose nodes are vehicles described by kinematic features; two multi-head graph-attention layers aggregate neighbours into an ego-node embedding, and a two-layer MLP maps that embedding to Q-values for five maneuvers: LEFT, IDLE, RIGHT, FASTER, SLOWER. The graph encoder and Q-head are trained jointly by the temporal-difference objective, so attention is shaped by its utility for action selection rather than by a separate representation loss. Natural-language descriptions are parsed by a schema-constrained language model into actor configurations and interaction programs, and those descriptions and semantic roles are withheld from the policy, which must infer threat relevance from observable kinematics alone. The second key object is the recognition–control gap measure, the difference between the success of a union of policies and the best single policy, which quantifies solvable headroom that no single policy exploits.","core_discovery":"On the paper's own terms, the central discovery is that a relational Q-network trained on language-structured safety-critical scenarios learns to attend to the adversarial vehicle, raising its threat-attention ratio from 1.2x to 2.1x and improving success by up to six points over a matched random control, but the policy then plateaus at roughly the success of the best constant action, 'SLOWER' at 57%, and does not exploit the 76% headroom available to a union of twelve simple policies. The authors formalize this as a recognition–control gap of 18 percentage points, show that reward reweighting and dense margin shaping actually reduce success, and find that increasing the decision frequency from 1Hz to 10Hz narrows the gap to 7 points. The same policy also transfers zero-shot from a highway driving simulator to a second urban simulator through a shared graph-state interface with about 73% average success, which the paper interprets as geometric portability of the state interface rather than visual robustness.","pith_inferences":["A cleaner counterfactual would train on the same coherent interaction programs with randomized or shuffled language labels; if performance and attention match the language-structured condition, the gains come from structured adversarial content, not from natural-language semantics.","The same recognition–control gap may appear in other language-generated training domains, such as robotic manipulation or game playing, where the policy must infer intent from observed behavior; decision-frequency and policy-union diagnostics are cheap to port.","The 7-point residual gap at 10Hz suggests that beyond temporal resolution, the discrete low-level action space (five maneuvers) or lack of memory may limit scenario-adaptive control; continuous or hierarchical actions are a testable next step.","If language-conditioned generation is to serve safety validation, the practical takeaway is to treat the generated scenes as a probe for attention and failure modes, not as a direct recipe for a deployable controller."],"forward_implications":["If the recognition–control gap is real, then reporting mean success over random controls can hide the fact that a learned policy is no better than a constant action; benchmarks should include constant-action baselines and policy-union headroom.","Raising the decision frequency from 1Hz to 10Hz cuts the gap from 18 to 7 points, so temporal resolution within the critical interaction window is a concrete lever for turning threat recognition into action.","Language-structured training alone is enough to shape attention toward relevant actors, which suggests language can act as a curriculum signal even when it is never seen at execution time.","Reward reweighting and margin shaping do not prevent policy collapse, implying the failure is not a simple reward-shaping issue and that objective-level diversity preservation is needed."],"supporting_citations":[{"why":"Defines the multi-head graph attention operator used in ERQ-Net's encoder to infer inter-vehicle relevance.","marker":"[17]"},{"why":"Supplies the deep Q-learning update with target networks that ERQ-Net's joint objective builds on.","marker":"[9]"},{"why":"Provides the highway driving simulator used for training and the policy-collapse experiments.","marker":"[8]"},{"why":"Provides the urban driving simulator used for zero-shot state-interface transfer.","marker":"[4]"},{"why":"Representative language-conditioned traffic generation method that motivates the hidden-language training setup.","marker":"[14]"},{"why":"Language-guided traffic simulation approach used as context for the language-to-scenario pipeline.","marker":"[22]"},{"why":"Language-to-scenario generation method that the paper contrasts with its focus on downstream policy learning.","marker":"[20]"},{"why":"Closed-loop adversarial training approach used as a contextual comparison for criticality and scenario generation cost.","marker":"[21]"}],"fun_headline_variants":["Language-structured driving RL boosts threat attention, not control","Recognition-control gap: threat-aware RL still matches constant action","Threat-aware driving policies fail to beat constant action","2.1x threat attention but policy still collapses to constant action","Language-structured driving RL: recognition without control"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The random-control training set is treated as a valid counterfactual: it matches the marginal ranges of vehicle count, position, lane, and speed but omits the coherent adversarial interaction, so the six-point success gain and 2.1x attention gain are attributed to language-induced structure rather than to the presence of structured adversarial content.","fun_headline_variants_meta":{"raw":{"variants":["Language-structured driving RL boosts threat attention, not control","Recognition-control gap: threat-aware RL still matches constant action","Threat-aware driving policies fail to beat constant action","2.1x threat attention but policy still collapses to constant action","Language-structured driving RL: recognition without control"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000985,"raw_usage":{"total_tokens":4209,"prompt_tokens":1005,"completion_tokens":3204,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":621,"completion_tokens_details":{"reasoning_tokens":3125}},"tokens_in":621,"tokens_out":3204,"duration_ms":17795,"temperature":1.0,"reasoning_tokens":3125,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T14:09:59.756167+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Measure the criticality score of the random-control training distribution under the same non-reactive reference ego used for the language-structured set—the collision rate when the ego drives at constant velocity with no avoidance. If the random-control criticality is substantially below the 0.75 reported for the language-structured set, the six-point success gain and 2.1x attention gain could reflect scenario difficulty rather than language-induced structure, and the counterfactual does not isolate semantics.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the highway driving simulator used for training and the policy-collapse experiments."},{"cited_title":"In: CoRL","cited_arxiv_id":null,"evidence_quote":"Provides the urban driving simulator used for zero-shot state-interface transfer."},{"cited_title":"In: CoRL","cited_arxiv_id":null,"evidence_quote":"Representative language-conditioned traffic generation method that motivates the hidden-language training setup."},{"cited_title":"In: CVPR (2024)","cited_arxiv_id":null,"evidence_quote":"Language-to-scenario generation method that the paper contrasts with its focus on downstream policy learning."},{"cited_title":"In: CoRL","cited_arxiv_id":null,"evidence_quote":"Closed-loop adversarial training approach used as a contextual comparison for criticality and scenario generation cost."}],"review_version":1}