{"id":"a5622b34-e404-449d-89f1-11d708fb0d8f","arxiv_id":"2608.06819","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"A receiver-conditioned token reranker, trained on a fixed answer-verified continuation, improves collaborative math decoding over LLM-local-preference baselines.","lead":"FutureBridge trains a small reranker to pick the next token a small language model should use when a large model offers alternatives. On five math benchmarks it beats methods that follow the large model's own preferences, improving average accuracy by 35.1% over greedy decoding.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Single-future supervision may measure fidelity to one LLM solution path, not ability to reach any correct solution; multi-future rank stability is untested.","rationale":"The reader's weakest_assumption correctly identifies the teacher-forced B_H as a surrogate for free-running SLM outcomes. My concern sharpens this: the surrogate is not only teacher-forced, it is conditioned on a single arbitrary answer-correct LLM trajectory. This introduces a reference-path bias that is distinct from the standard exposure-bias gap. The paper's existing rollout diagnostics are reassuring but do not address alternative verified futures, so the concern is genuinely untested. However, the reader's verdict is already CONDITIONAL, and the missing evidence is a concrete experiment rather than a demonstrated contradiction. Therefore I do not change the verdict; I recommend running the multi-future stability test before accepting the central claim without reservation.","tokens_in":14116,"tokens_out":10437,"duration_ms":116156,"concrete_test":"For a sample of admitted states, generate K=8 answer-correct LLM futures per state via temperature sampling with the same verifier filter used in Eq. (4). For each state, compute B_H for every candidate under each of the K futures and measure Kendall tau between candidate rankings across futures. If the average tau is below ~0.7 or the top candidate changes in more than 20% of states, the single-future surrogate is path-dependent and the reranker supervision is not a stable estimate of Eq. (1). Optionally, retrain the reranker on all K futures per state and check whether rollout agreement and Math Avg. improve.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Equations (4)-(7) define B_H as the SLM's teacher-forced likelihood of one answer-verified LLM greedy future f_t, fixed per state. The ideal objective in Eq. (1) rewards any candidate whose free-running SLM continuation reaches a correct final answer. If an admitted state has several correct solution paths, B_H systematically favors candidates that make the SLM imitate the LLM's single path and penalizes candidates that would support a different but equally valid path. This is not merely the acknowledged teacher-forcing/free-running gap (Sec. 4.3); it is a reference-trajectory bias. The Figure 2(a) and Table 9 rollout diagnostics compare B_H with complete rollouts, but they test one reference future per state and do not show that the candidate ranking is stable across alternative verified futures. The horizon sweep (Fig. 2b) varies only prefixes of the same future, so it cannot detect this bias. Consequently, the central claim that the reranker models whether the receiving SLM can continue reasoning is not yet established for states with diverse correct solutions.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"FutureBridge proposes a token-level collaboration method in which, at admitted decoding states, a joint pool of SLM and LLM next-token candidates is reranked by a lightweight scorer trained to predict how well each candidate supports the receiving SLM's subsequent reasoning. During training, an answer-verified greedy LLM trajectory supplies a fixed shared future; for every candidate, the frozen SLM's average teacher-forced log-likelihood of that future, after inserting the candidate, becomes the supervision target. This target is distilled into a LoRA-based reranker that sees only the current state and candidate token. At inference, the LLM contributes only next-token candidates, the reranker selects one token, and generation returns to the SLM. On five mathematical reasoning benchmarks, FutureBridge improves Qwen3-1.7B Math Avg. from 37.30 (greedy SLM) to 50.38, and from 49.14 for the S2T baseline, under the same request schedule and intervention budget; similar gains are reported for Qwen3-0.6B. The paper includes controlled ablations on candidate-pool composition, supervision target, future horizon, and training-future quality, plus candidate-level rollout diagnostics.","tokens_in":14335,"tokens_out":6953,"duration_ms":68340,"significance":"If the central claim holds, the paper makes a useful conceptual contribution: it replaces LLM-local preference with an SLM-conditioned compatibility signal for token selection, and it shows through matched ablations that this signal, rather than merely the enlarged candidate pool, drives the end-to-end gains. The manuscript is also commendably concrete: it fixes the request policy, intervention budget, candidate pools, architecture, and optimization budget across comparisons; it reports deployment cost (4.8 LLM calls, one appended token per intervention); and it explicitly acknowledges that the compatibility score is a teacher-forced surrogate, not a causal outcome, providing rollout diagnostics to probe the connection. The quality ablations (unverified, incorrect, cross-state, and token-shuffled futures) are a good falsifiable check that the supervision signal is not vacuous. The main unresolved risk is that the score is defined against a single answer-verified LLM future per state, so the measured \"future compatibility\" may partly reflect fidelity to one particular correct path rather than general ability of the receiver to continue reasoning.","major_comments":[{"comment":"The compatibility score B_H is defined against one answer-verified greedy LLM future f_t fixed per state. If a state admits several correct solution paths, B_H rewards candidates that make the SLM imitate that particular LLM continuation and penalizes candidates that support an equally valid alternative path. The diagnostics in Fig. 2(a) and Table 9 evaluate only one reference future per state, and the horizon sweep in Fig. 2(b) varies only prefixes of the same future, so neither can detect this bias. As a result, the central claim that the reranker models \"whether the receiving SLM can continue reasoning\" is not yet fully established: the current supervision cannot distinguish future-compatibility from single-trajectory imitation. Please add a multi-future analysis, e.g., sampling several verified correct futures per state and reporting top-1 agreement or Kendall tau of the B_H rankings across futures, or training with a multi-future aggregated target; this would demonstrate that the score is stable across correct solution paths.","section":"§4.2, Eqs. (4)-(7), and Fig. 2(a)/Table 9"},{"comment":"The teacher-forced surrogate B_H is acknowledged in the text, and the rollout agreement in Table 9 is a useful check, but the agreement is only 69.7% pairwise and top-1 success on OlympiadBench is 40.1% for Direct-B_H. Because the shared future is the LLM's own greedy continuation, the score may mechanically favor the LLM token a_T^t and candidates that imitate the LLM, rather than candidates that genuinely support free-running SLM success. Table 9 does not report agreement separately for the singleton a_T^t, for SLM-only versus LLM-only candidates, or by rollout success. Please add these subgroup diagnostics; without them, the improvement over \"LLM local score\" in Table 3 cannot be attributed unambiguously to receiver-conditioning rather than to a proxy for LLM-path fidelity.","section":"§4.3 and Table 9"},{"comment":"The logged training states are collected by running the frozen SLM without any intervention tokens, whereas at deployment the reranker is applied to states that may contain tokens selected by the reranker at earlier admitted events. The paper does not quantify the distribution shift between these logged states and the states actually encountered at inference. If earlier intervention tokens change the SLM's subsequent state distribution, the distillation targets may be unrepresentative of deployment. Please measure this drift (for example, report a state-overlap or KL metric between logged and deployment states, or retrain on rollouts generated with the deployed reranker) and report whether end-to-end accuracy changes.","section":"§4.2 and Supplement A"}],"minor_comments":[{"comment":"The label \"Δ accuracy\" in the figure is not defined in the main text; please state explicitly in the caption or text that values are accuracy-point differences relative to the H=16 configuration.","section":"Figure 2(b)"},{"comment":"The term \"Direct-B_H\" is used in Figure 2 and Table 9 without a definition in the main text; please define it at first use as the privileged offline compatibility score computed from the shared future.","section":"§5.2 and Table 9"},{"comment":"The supplement reports that Table 1 gives the seed mean without per-cell standard deviations; please include standard deviations for the main head-to-head comparisons, at least for FutureBridge and S2T, since the reported differences are only 1-2 accuracy points in some rows.","section":"§5.1 and Table 1"},{"comment":"The text says MATH validation is used to select the request threshold, reranker checkpoint, and H; please clarify whether the threshold and H are selected jointly or independently, and confirm that no held-out benchmark is used in any selection step.","section":"Section 5.1, collaboration protocol"}],"recommendation":"major_revision","confidential_remarks":"The paper's strengths are real: the controlled ablations are well designed, the deployment protocol is clearly separated from privileged training information, and the quality ablations are a good falsifiable check. The main risk is the single-future reference bias, which is addressable in a revision with multi-future rank-stability experiments and finer-grained rollout diagnostics. I would support a major revision rather than rejection, because the central claim is defensible and the requested experiments are within the scope of the paper's methodology."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Rong, this one is worth a real look. The core idea is genuinely new: instead of ranking candidate tokens by the LLM's next-token likelihood, they score each candidate by how well the frozen SLM can continue from it under a fixed answer-verified future, then distill that into a token-only reranker. The experiments are well controlled: they compare against S2T, R2R, and Takeover under the same request schedule and candidate budget, and the ablations show the gain comes from the future-compatibility target, not just candidate expansion. The candidate-level rollout diagnostics are a nice touch—they actually check that the privileged score agrees with complete rollouts. The quality ablations, where the future is corrupted, degrade monotonically, which is the right behavior. That is real evidence.\n\nThe soft spot is the one-future-per-state design. They evaluate every candidate against a single greedy LLM trajectory that is answer-verified. That means the compatibility score measures fidelity to one solution path, not the ability to reach any correct solution. The paper acknowledges the teacher-forcing/free-running gap in Sec. 4.3, but the concern runs deeper: if a state has several legitimate continuations, a candidate that supports a different but equally correct path will be scored down. The rollout diagnostics test one reference future per state, so they cannot catch this bias. The horizon sweep only lengthens prefixes of the same future, so it does not test rank stability across futures. This is not fatal—the gains are consistent and the method is a clear improvement over local scoring—but the central claim, that the reranker models whether the receiving SLM can continue reasoning, is not fully established until they show the ranking is stable across multiple verified futures. That is the experiment I would ask for.\n\nMinor issues: code and data are not released, and the exact verifier/prompt are not specified, so independent replication is harder than it should be. Also, some related methods (AlphaRouter, Local Branch Routing) are only cited, not compared, though the main comparison set is reasonable.\n\nBottom line: this is a solid paper with a genuine contribution and a clear, testable weakness. I would send it to review. A revision that adds multi-future stability analysis and releases artifacts would make the claim much stronger.","headline":"Genuinely new receiver-conditioned token reranking with strong controlled ablations, but the single-verified-future supervision leaves a testable gap around multi-path correctness that a revision should close.","tokens_in":14899,"tokens_out":2235,"would_cite":true,"duration_ms":22090,"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":"FutureBridge reranks joint SLM-LLM token candidates by the frozen SLM's likelihood of a verified shared future, improving Qwen3-1.7B math accuracy by 35.1% relative to greedy decoding.","keywords":["token-level collaboration","small-large model collaboration","token reranking","future compatibility","teacher forcing","mathematical reasoning","knowledge distillation","decoding"],"falsifier":"Take a held-out set of admitted states, compute B_H for each candidate, and also run complete free-running SLM rollouts from each candidate. If there exists a statistically significant group of states where the B_H-selected token yields a lower final-answer accuracy than the LLM-local-preference token, or where the B_H ranking has worse pairwise agreement with rollout success than the LLM local score, the surrogate would be falsified for those candidates.","tokens_in":13933,"feed_emoji":"🧮","tokens_out":5913,"duration_ms":50351,"temperature":0.7,"pith_summary":"When a large language model helps a small one decide the next token, the usual signal is the LLM's own token probability. FutureBridge argues that this local preference can pick tokens the small model cannot build on, and that the right signal is whether the receiving small model can keep reasoning toward a correct answer from the candidate. To get that signal cheaply, the paper freezes the small model, fixes an answer-verified LLM-generated future suffix as a common reference, and scores every candidate by the small model's average teacher-forced log-likelihood of that suffix. This score supervises a lightweight token reranker that at deployment sees only the current state and candidate token, while the LLM only proposes candidates. Across five math benchmarks with Qwen3-1.7B as the receiver and Qwen3-32B as the teacher, the method lifts Math Avg. by 35.1% relative to greedy decoding and by 1.24 points over the S2T reranker under the same request schedule and candidate budget.","feed_headline":"Choosing tokens the small model can use lifts math accuracy 35%","feed_subtitle":"A small model's likelihood of a verified shared future picks better tokens than the large model's own probabilities.","key_machinery":"The machinery is the student-conditioned future compatibility score B_H(c_k | s_t, f_t^(H)): the length-normalized average of the frozen SLM's conditional log-probs for each token of a shared, answer-verified LLM future, computed under teacher forcing after inserting candidate c_k. This score converts a fixed downstream context into per-candidate labels, and the group-level soft targets in Eq. (9) distill these labels into a LoRA-based reranker r_theta that observes only (s_t, c_k). The joint candidate pool C_t = C^S_t ∪ C^T_t separates candidate coverage from selection: the LLM expands the action space, while the compatibility score picks the token.","core_discovery":"The central discovery is that the LLM's local next-token probability is a poor selector for collaborative decoding, because an LLM-preferred token can be hard for the smaller model to continue from. FutureBridge instead defines the value of a candidate token c_k at state s_t as the average teacher-forced log-likelihood B_H(c_k) that the frozen SLM assigns to a fixed, answer-verified shared future f_t after inserting c_k. Splitting the candidate pool into a joint SLM-LLM set and scoring all candidates against the same future gives directly comparable, receiver-conditioned supervision, which is then distilled into a token-only reranker. The results show this future-compatibility signal agrees substantially better with complete candidate rollouts than the LLM's local score, and that the end-to-end gains hold at two SLM scales.","pith_inferences":["We infer that the same counterfactual-compatibility recipe could transfer to non-math domains (e.g., code or multi-step instruction following) whenever a verifier can certify one correct continuation; the paper only tests math.","The paper's own Sec. 4.3 caveat suggests a stress test: if a candidate yields high teacher-forced likelihood for the shared future but derails free-running generation (e.g., by triggering repetition or a different reasoning style), the reranker would misrank it; future work could augment the surrogate with a short free-running rollout for a small subset of candidates.","Because the shared future comes from greedy LLM decoding, the method inherits any bias in that decoding; using a diverse set of verified futures as supervision could provide a denoised target, which the paper does not explore."],"forward_implications":["If the compatibility signal is right, token-level collaboration can be improved without changing the request policy or the candidate budget; the gains come purely from selection.","The reranker needs no LLM suffix at inference, so deployment cost stays token-level and the SLM context is never polluted by unverified LLM continuations.","The method generalizes to a weaker SLM (Qwen3-0.6B), gaining more over baselines than with the 1.7B receiver, suggesting that receiver-conditioned selection matters more when the receiver is less capable.","The ablation on future quality (unverified, incorrect, cross-state, token-shuffled) shows that the shared future must be state-specific and answer-correct; the supervision signal degrades monotonically as the future is corrupted.","The horizon sweep shows most of the benefit is captured by H=64, and extending to 128 adds little, meaning a modest amount of downstream context suffices."],"supporting_citations":[{"why":"The token-reranking baseline whose distillation setup FutureBridge extends and against which it reports the 1.24-point gain.","marker":"(Ye et al. 2026)"},{"why":"The divergent-path token routing baseline that FutureBridge compares with and improves upon.","marker":"(Fu et al. 2025)"},{"why":"Defines the Qwen3 model family used as the SLM and LLM in all experiments.","marker":"(Yang et al. 2025)"},{"why":"Source of the MATH training and validation splits used to construct supervision and select hyperparameters.","marker":"(Hendrycks et al. 2021)"},{"why":"One of the five evaluation benchmarks that make up the reported Math Avg.","marker":"(Cobbe et al. 2021)"}],"fun_headline_variants":["Rank tokens by the small model's future, not the LLM's: +35% math","FutureBridge: choose tokens by SLM's reasoning future, +35% math accuracy","LLM's local pick isn't best; score candidates by the SLM's future, +35% math","Instead of LLM's best token, pick the one the SLM can build on: 35% boost","Token selection via future SLM support: math accuracy up 35% vs greedy"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the average teacher-forced likelihood B_H that the frozen small model assigns to a fixed answer-verified future is a reliable proxy for the outcome the small model would actually produce if allowed to generate freely from the candidate; the paper states this is a tractable surrogate, not a causal outcome.","fun_headline_variants_meta":{"raw":{"variants":["Rank tokens by the small model's future, not the LLM's: +35% math","FutureBridge: choose tokens by SLM's reasoning future, +35% math accuracy","LLM's local pick isn't best; score candidates by the SLM's future, +35% math","Instead of LLM's best token, pick the one the SLM can build on: 35% boost","Token selection via future SLM support: math accuracy up 35% vs greedy"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001191,"raw_usage":{"total_tokens":4915,"prompt_tokens":946,"completion_tokens":3969,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":562,"completion_tokens_details":{"reasoning_tokens":3847}},"tokens_in":562,"tokens_out":3969,"duration_ms":28054,"temperature":1.0,"reasoning_tokens":3847,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T20:15:51.966997+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a held-out set of admitted states, compute B_H for each candidate, and also run complete free-running SLM rollouts from each candidate. If there exists a statistically significant group of states where the B_H-selected token yields a lower final-answer accuracy than the LLM-local-preference token, or where the B_H ranking has worse pairwise agreement with rollout success than the LLM local score, the surrogate would be falsified for those candidates.","supporting_citations":[],"review_version":1}