{"id":"2845b196-f4ac-45d5-bf73-e8e04fcf2bf7","arxiv_id":"2608.10447","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"REAM merges slow- and fast-thinking recommender models at the per-attention-head level, cutting reasoning length by up to 24.3% while preserving rating accuracy.","lead":"This paper introduces REAM, a model-merging method that shortens the reasoning traces of slow-thinking LLM recommenders by selectively blending in parameters from a fast, direct-prediction model. It reports up to 24.3% shorter outputs with maintained or improved rating accuracy on three benchmark datasets.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The fast model's task vector is ill-posed: θ_F is fine-tuned from Qwen2.5-3B-non-instruct while the declared shared base θ_B is Qwen2.5-3B-Instruct, so Δ_F = θ_F − θ_B includes a base-checkpoint difference that may drive the reported compression.","rationale":"I agree with the reader that the paper is a well-executed empirical study with credible ablations, causal validation of head-level signals, and stable calibration estimates; the main results are not obviously fraudulent or circular. However, the reader's weakest assumption concerns what happens after merging (whether protected heads stay critical), whereas the more fundamental problem is what is being merged: the fast model's difference from the declared shared base. Because θ_F was fine-tuned from a different pretrained checkpoint (Qwen2.5-3B-non-instruct) than the shared base θ_B (Qwen2.5-3B-Instruct), Δ_F = θ_F − θ_B is not a task vector in the sense of Section 3.2. This invalidates the interpretation of Eq. 8 as 'selectively injecting concise behavior.' It also means the reasoning-aware coefficients are weighting an update that includes an uncontrolled checkpoint shift, so the reported token reduction could be partly an artifact of the base-model difference rather than of REAM's allocation. The proposed decomposition test settles this directly. If the Δ_ckpt-only merge is inert, the concern does not land and the reader's CONDITIONAL verdict can remain on the original grounds. If it is active, the central claim needs reframing or the fast model must be fine-tuned from the declared base. I therefore keep the verdict CONDITIONAL, but for a different, more foundational reason than the reader's.","tokens_in":30353,"tokens_out":6071,"duration_ms":60147,"concrete_test":"Decompose the injected update and rerun the full REAM pipeline (same 500-example calibration, same coefficients from Eqs. 1–8, same FFN exclusion window, same ρ and γ) on at least Amazon Book and Yelp with Δ_F replaced by Δ_ckpt = θ_{B,non-instruct} − θ_B only. If the Δ_ckpt-only merge produces a token reduction comparable to REAM's reported 17.9–24.3%, or materially shifts MAE/RMSE, then the published gains are substantially attributable to base-checkpoint differences rather than to the fast-thinking task vector. A complementary check is to fully fine-tune TALLRec from θ_B = Qwen2.5-3B-Instruct and rerun REAM; if the accuracy and token-reduction gains disappear, the different-base configuration is the cause.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The most load-bearing concern is that Δ_F is not a task vector relative to the declared shared base. Section 3.2 (Eq. 1) defines task vectors as Δ_m = θ_m − θ_B and says both models are 'represented relative to the same reference θ_B.' But Section 5.1 and Appendix A.1 state that the fast model θ_F (adapted TALLRec) was fully fine-tuned from Qwen2.5-3B-non-instruct, while θ_B = Qwen2.5-3B-Instruct. Consequently Δ_F = θ_F − θ_B = (θ_F − θ_{B,non-instruct}) + (θ_{B,non-instruct} − θ_B). The first term is the genuine TALLRec fine-tuning task vector; the second is the checkpoint difference between the instruct and non-instruct bases. REAM's retrieval-criticality, faithfulness, and Fisher-sensitivity signals are computed on the slow model, but the coefficients are applied to this contaminated Δ_F. Some or all of the measured 17.9–24.3% token reduction could therefore come from injecting the non-instruct/instruct checkpoint delta rather than from selectively transferring the fast model's concise reasoning behavior. The central claim—that head-level, reasoning-aware coefficients enable a grounded transfer of conciseness into the slow model—presupposes that Δ_F is a behavior-specific update from θ_B. That presupposition is false as implemented, so the mechanism story and the interpretation of Table 1 are not currently supported. This is upstream of the reader's weaker-assumption concern about post-merge head roles: even if head roles were stable after merging, the update being allocated is not the fast model's task vector.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes REAM, a model merging method for compressing chain-of-thought reasoning in LLM-based recommender systems without additional training. REAM merges a slow-thinking RecZero model with a fast-thinking TALLRec model at the granularity of attention heads, assigning each head a merge coefficient based on retrieval criticality, decision faithfulness, and Fisher-weighted update sensitivity. A constrained optimization allocates the fast-thinking update while limiting perturbation risk, and late-layer FFN updates are excluded based on validation. Experiments on Amazon Book, Yelp, and Amazon Music report token reductions of 17.9% to 24.3% with maintained or improved MAE/RMSE relative to the slow model, and the paper includes ablations, causal validation of the head-level signals, hyperparameter analyses, and a generalisability study across model scales and backbones.","tokens_in":30598,"tokens_out":6216,"duration_ms":60921,"significance":"If the central mechanism is established, REAM would be a useful contribution: it is, to the authors' knowledge, the first merging framework for reasoning compression in recommender systems, it operates at a finer granularity than most merging baselines, and it ships with detailed derivations, a stability analysis, causal head ablations, and public code. The empirical pattern is consistent across three datasets, and the paper is honest about the training-free caveats. However, the significance is currently conditional on resolving a load-bearing issue with the definition of the fast task vector, which affects the interpretation of all the main results.","major_comments":[{"comment":"The fast task vector Δ_F is not defined relative to the declared shared base θ_B. The paper states that θ_F (adapted TALLRec) is fully fine-tuned from Qwen2.5-3B-non-instruct, while θ_B is Qwen2.5-3B-Instruct and is used as the shared base for both task vectors. Consequently Δ_F = θ_F − θ_B = (θ_F − θ_{B,non-instruct}) + (θ_{B,non-instruct} − θ_B). The first term is the genuine TALLRec adaptation vector; the second is a checkpoint difference unrelated to fast reasoning behavior. Since Eq. (8) applies this Δ_F to θ_S, the reported token reduction and accuracy effects could be driven by injecting the non-instruct/instruct checkpoint delta rather than by selectively transferring concise reasoning. This contradicts the statement in §3.2 that both task vectors are represented relative to the same reference, and it undermines the central claim that REAM's reasoning-aware coefficients enable a grounded transfer of conciseness. I ask the authors to re-run the experiments with a fast model fine-tuned from Qwen2.5-3B-Instruct, and to report a control experiment that merges only the pure checkpoint delta (θ_{B,non-instruct} − θ_B) using REAM's coefficients.","section":"§5.1, Appendix A.1, and §3.2"},{"comment":"The merge coefficients are computed entirely from signals measured on the pre-merge slow model θ_S, and the causal validation in §5.4.2 ablates heads of θ_S. The method assumes that the heads that are reasoning-critical before merging remain the heads that support reasoning after the fast update is applied, but the paper does not test post-merge head roles. This assumption is load-bearing because the entire design protects certain heads from the fast update; if the merge shifts which heads are reasoning-critical, the coefficients stop serving their stated purpose. Please provide evidence on the merged model itself, for example by re-computing retrieval criticality or decision faithfulness on θ★, or by ablating the previously protected heads in θ★ and comparing the effect with random ablations.","section":"§4.3 and §5.4.2"},{"comment":"The accuracy gains over the slow model θ_S are within overlapping 95% confidence intervals on all three datasets. For example, on Book, REAM's MAE is 0.6338±0.0181 versus θ_S's 0.6650±0.0180; on Yelp, 0.7564±0.0237 versus 0.7769±0.0250; on Music, 0.5348±0.0351 versus 0.5433±0.0360. The RMSE intervals overlap similarly. The claim that REAM 'maintains' accuracy is therefore supported only in the weak sense that the observed degradation is not statistically significant; the stronger claim that REAM improves accuracy over θ_S is not established. The paper should state this explicitly and, if the stronger claim is intended, provide paired significance tests or a power analysis.","section":"Table 1"}],"minor_comments":[{"comment":"The FFN exclusion window (layers 30–35) and the budget fraction ρ=0.3 are selected on validation subsets and then used as fixed choices for the reported results; Appendix B.9 itself states that the exact window should be validated once for each new backbone. This should be acknowledged as validation-based hyperparameter selection rather than a parameter-free aspect of the method, and the authors should clarify how this selection affects the generality claims in Table 3.","section":"§5.4.3 and Appendix B.9"},{"comment":"The use of the diagonal empirical Fisher as a curvature surrogate is justified by reference to the generalised Gauss–Newton decomposition, but the paper should explicitly note the known limitations of the empirical Fisher approximation (e.g., reference [23]) and state why the diagonal assumption is acceptable at head granularity.","section":"§4.2 and Appendix B.4"},{"comment":"The method is described as training-free, which is reasonable in the sense that no model parameters are optimized, but the Fisher-sensitivity estimation requires backward passes to compute gradients. The authors should clarify this terminology, since readers may understand 'training-free' to mean 'no gradient computation at all.'","section":"Appendix A.4"},{"comment":"The ordering of rows in the Music panel differs from the Yelp panel (Mass-based, √n-faith, Density-faith versus Mass-based, Density-faith, √n-faith); please make the ordering consistent.","section":"Table 12"},{"comment":"The paper notes that fully unstructured traces are left to future work; this is a fair limitation, but it should also be mentioned in the main text's discussion of applicability, not only in the appendix.","section":"A.7"}],"recommendation":"major_revision","confidential_remarks":"The base-checkpoint mismatch is the key risk. If the authors cannot re-run the fast model from the shared instruct base, the mechanism story should be substantially softened to an empirical observation about merging a slow instruct-tuned model with a fast non-instruct-tuned model. I would not reject on this basis alone, since the issue is fixable with additional control experiments, but the current version does not support the central claim as written. The statistical overlap in Table 1 is a secondary concern that should be addressed in the revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Hi [Colleague],\n\nThe paper you want to know about: REAM, a model-merging method that compresses reasoning traces in LLM-based recommenders by merging a slow-thinking RecZero with a fast-thinking TALLRec. The empirical story is clean: on three datasets they cut tokens by 17.9–24.3% while holding or slightly improving MAE/RMSE, and they beat eight training-free baselines on most metrics. The head-level allocation, driven by retrieval-criticality and decision-faithfulness plus a Fisher-weighted risk budget, is a genuine extension of L2S-Merge and ACM into structured-segment recommendation. The ablation and causal validation are careful; the KKT derivation is proper; the stability analysis of the calibration set is a nice touch.\n\nBut there is a load-bearing flaw. Section 3.2 defines both task vectors relative to a shared base θ_B = Qwen2.5-3B-Instruct, but Section 5.1 says θ_F (TALLRec) was fully fine-tuned from Qwen2.5-3B-non-instruct. So Δ_F = θ_F − θ_B is not the fast model's task vector; it's the TALLRec update plus a large non-instruct-versus-instruct checkpoint delta. The importance and sensitivity signals are computed on θ_S, but the coefficients are applied to this contaminated Δ_F. Nothing in the paper rules out that much of the measured token reduction and accuracy preservation comes from the checkpoint difference rather than from selectively importing TALLRec's concise behavior. This is not a minor caveat; it breaks the mechanism story, which is the paper's central contribution.\n\nThe other soft spots are minor by comparison. The generalization section says 'without re-tuning' while Appendix B.9 re-selects the FFN exclusion window per backbone via validation sweep. Hyperparameters γ, ρ, and the FFN window are validation-tuned, so part of the gain is selected, not predicted. Several accuracy gains fall within overlapping 95% CIs. The calibration set is 500 well-formed traces, which may not represent hard cases.\n\nIf I were refereeing this, I'd ask for a control: merge the pure checkpoint delta (non-instruct minus instruct, no TALLRec adaptation) and measure tokens and accuracy. That would tell you how much of the effect is actually the fast model. Alternatively, re-run with θ_F fine-tuned from the same Instruct base, even if that makes the model less 'fast-thinking' in behavior. The empirical phenomenon might survive, but the current paper doesn't demonstrate it.\n\nThis is worth a serious referee, because the idea is new and the engineering is strong, but the version I read needs substantive revision. I wouldn't cite it as-is, though I might cite a corrected version.\n\nBest.","headline":"Useful empirical study of head-level model merging for reasoning compression, but the fast model's task vector is built from a different base checkpoint, so the mechanism claim does not hold as stated.","tokens_in":31263,"tokens_out":3888,"would_cite":false,"duration_ms":34967,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"REAM merges fast- and slow-thinking recommenders at attention-head granularity, cutting reasoning length by up to 24.3 percent while preserving rating accuracy.","keywords":["model merging","reasoning compression","LLM-based recommender systems","attention heads","task vectors","retrieval criticality","decision faithfulness","rating prediction"],"falsifier":"Recompute retrieval criticality and decision faithfulness on the merged model and compare the top head sets with the pre-merge calibration; if the sets diverge substantially and ablating the newly critical heads changes ratings, the fixed-signal assumption fails.","tokens_in":30044,"feed_emoji":"🧠","tokens_out":5449,"duration_ms":48656,"temperature":0.7,"pith_summary":"This paper tries to establish that you can make a slow-thinking LLM recommender cheaper without retraining by merging its weights with those of a fast-thinking counterpart that predicts ratings directly. The proposed method, REAM, does this at the level of individual attention heads: each head gets its own merge coefficient, chosen so that heads important for retrieving user–item evidence and for grounding the final rating are protected, while less critical heads absorb more of the concise model's update. If the claim holds, reasoning traces shrink by up to 24.3 percent while rating accuracy stays intact or improves across Amazon Book, Yelp, and Amazon Music. This matters because reasoning traces dominate inference cost in LLM recommenders, and existing compression routes require either extra training or brittle decoding-time tricks.","feed_headline":"Head-aware merge cuts LLM recommender reasoning by 24 percent","feed_subtitle":"Merging concise and verbose recommenders per attention head keeps accuracy while trimming generated tokens.","key_machinery":"The central object is the head-level task-vector allocation. REAM anchors on the slow-thinking model and adds a fraction of the fast-thinking model's task vector to each attention-head parameter group. The fraction comes from a constrained optimisation whose per-head cost combines three signals: retrieval criticality (how often a head retrieves segment-relevant user–item evidence), decision faithfulness (how much of the rating step's attention goes to the compatibility-match segment), and Fisher-weighted update sensitivity (loss curvature along the fast update). The optimisation is solved by a water-filling KKT solution that allocates the update budget across heads, and the merge applies head-specific coefficients while excluding the final feed-forward layers.","core_discovery":"REAM's central claim is that reasoning compression in slow-thinking recommenders is a selective merging problem: import conciseness from a direct-prediction model, but allocate the task-vector update per attention head according to how much that head supports the reasoning behind the rating. On three datasets, the merged model reduces mean generated tokens by 17.9–24.3 percent compared with the slow model, and produces lower MAE and RMSE than the slow model on all three while beating eight training-free merging baselines on the accuracy–efficiency trade-off. The method needs no training and no decoding changes.","pith_inferences":["If the head roles are stable across inputs, the same coefficient schedule could be computed once per domain and applied to new users without recalibration; the paper's cross-domain Jaccard stability hints at this.","The two near-disjoint head sets suggest a natural extension: compress by steering updates through faithfulness-neutral heads while keeping retrieval heads frozen, which could be tested by comparing REAM against a frozen-retrieval variant.","The perturbation-cost formulation could be reused to predict per-input token budgets, choosing how much fast behaviour to inject based on trace difficulty rather than using one merged model for all inputs.","Because REAM only needs a verbose model and a concise model sharing a base, the same recipe might compress reasoning in other generation tasks with structured traces, such as tool-use or planning, not just rating prediction."],"forward_implications":["On all three datasets, mean generated tokens drop by 17.9% to 24.3% relative to the slow model, and MAE and RMSE are lower than the slow model's.","REAM beats all eight training-free baselines on the accuracy–efficiency trade-off, and head-level allocation is necessary: layer-wise activation baselines leave longer traces and worse MAE.","The retrieval-critical and decision-faithful head sets are causally important and nearly disjoint, so both signals pull their weight; ablating either set raises held-out NLL and divergence far more than random head ablation.","The pipeline is training-free and cheap: about 0.69 GPU-hours per dataset, roughly 4% of the combined cost of training the two source models.","The approach transfers across model scale and family, improving on the slow model on all metrics under the RecOne reasoning backbone, though on Qwen2.5-7B it shortens traces without fully matching the slow model's accuracy."],"supporting_citations":[{"why":"Supplies the slow-thinking model whose structured reasoning traces are compressed.","marker":"[22]"},{"why":"Supplies the fast-thinking model whose concise output behaviour is imported.","marker":"[4]"},{"why":"Defines task vectors, the parameter-space arithmetic that REAM's head-level merge operates on.","marker":"[17]"},{"why":"Establishes the long-to-short reasoning merging setting that REAM extends to head-level, reasoning-aware coefficients.","marker":"[42]"},{"why":"Provides the retrieval-head criterion used to define retrieval criticality.","marker":"[44]"},{"why":"Supplies the Fisher-weighted sensitivity principle behind REAM's update-sensitivity measure.","marker":"[30]"},{"why":"Provides the rating-prediction splits used for all evaluations.","marker":"[7]"},{"why":"Layer-wise activation-guided merging baseline that REAM compares against and refines to head granularity.","marker":"[50]"}],"fun_headline_variants":["Head-level merge trims LLM recommender reasoning up to 24%","Per-head merging cuts reasoning tokens in recommenders","Attention-head merge compresses LLM reasoning for recs","Selective head merging speeds up LLM recommenders","Model merging slashes verbose LLM recommender thinking"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the importance and sensitivity signals measured on the slow model before merging still describe which heads matter after merging; if the merge reshuffles head roles, the protected heads may no longer be the ones protecting accuracy.","fun_headline_variants_meta":{"raw":{"variants":["Head-level merge trims LLM recommender reasoning up to 24%","Per-head merging cuts reasoning tokens in recommenders","Attention-head merge compresses LLM reasoning for recs","Selective head merging speeds up LLM recommenders","Model merging slashes verbose LLM recommender thinking"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000687,"raw_usage":{"total_tokens":3101,"prompt_tokens":919,"completion_tokens":2182,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":535,"completion_tokens_details":{"reasoning_tokens":2115}},"tokens_in":535,"tokens_out":2182,"duration_ms":14932,"temperature":1.0,"reasoning_tokens":2115,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T14:20:13.391736+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Recompute retrieval criticality and decision faithfulness on the merged model and compare the top head sets with the pre-merge calibration; if the sets diverge substantially and ablating the newly critical heads changes ratings, the fixed-signal assumption fails.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the retrieval-head criterion used to define retrieval criticality."}],"review_version":1}