{"id":"748136ec-71fb-4ef9-a082-c63ab51c7190","arxiv_id":"2509.09801","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"A two-stage fine-tuning method that runs LoRA before ReFT achieves 85.17% on BoolQ with 3+3 epochs, higher than 20-epoch LoRA or ReFT alone.","lead":"This paper stacks two popular methods for adapting large language models, first adjusting the model's weights, then adjusting its internal signals, and tests the combination on a question-answering benchmark. The authors report that the stack slightly beats either method alone while using less training time.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The central synergistic-efficiency claim rests on a 0.12 percentage point accuracy gap (about 4 of 3,270 examples) from a single unreported seed; no variance or matched-epoch baselines are reported, so the HEFT advantage over LoRA-only is not established.","rationale":"The paper's strongest claim is a point-estimate comparison from Table 1. The gap that carries the argument is 0.12 percentage points between HEFT(3+3) and LoRA-only(20); this is about 4 examples, while the binomial standard error for a proportion near 0.85 at n=3,270 is roughly 0.62 percentage points. With one run and no reported seed, the result is compatible with a null effect. This is not a disagreement with consensus; it is an internal evidential gap: the paper itself provides no repeated runs. The appendix's runnable script is a genuine strength, but it does not set a seed and therefore does not yet quantify replication uncertainty. The limitations section is honest about task, order, and hyperparameter scope, but missing replication is not listed. A matched-epoch control such as LoRA-only at 3 or 6 epochs is also absent, so the coarse-to-fine interpretation cannot be distinguished from simply training LoRA for fewer epochs plus a small ReFT bump. The reader's weakest_assumption identifies the same load-bearing issue, and I agree with that identification. The appropriate outcome is unchanged: conditional acceptance conditioned on seeds, variance reporting, and matched-epoch baselines. This is not a reason to reject outright, because the proposed method is cheap, reproducible from the appendix, and the point estimate, if replicated, would support the claim.","tokens_in":13251,"tokens_out":9067,"duration_ms":79063,"concrete_test":"Run the published script for HEFT(3+3), LoRA-only(20), ReFT-only(20), and also LoRA-only(3) and LoRA-only(6), using at least 5 distinct seeds for each condition, recording the seed values (the appendix script currently sets no random seed before Trainer initialization). Report per-seed validation accuracy, mean, standard deviation, and a paired 95% confidence interval or bootstrap interval for the difference HEFT(3+3) minus LoRA-only(20). If the interval includes 0, or if LoRA-only at 3 or 6 epochs already matches HEFT(3+3), the claimed synergy and efficiency advantage do not survive.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim (abstract and Section 4.1) is that HEFT with 3+3 epochs reaches 85.17%, exceeding LoRA-only at 20 epochs (85.05%) and ReFT-only at 20 epochs (83.36%), implying a synergistic accuracy/compute gain. The decisive comparison is HEFT versus LoRA-only: a gap of 0.12 percentage points corresponds to about 4 validation examples out of 3,270. Treating each validation example as an independent Bernoulli draw with p near 0.85, a rough standard error for a single run is about 0.62 percentage points, so the observed gap is smaller than one standard error. Table 1 reports one run per condition; no seeds, repeats, confidence intervals, or significance tests appear anywhere in the paper. Thus the 'profound synergistic effect' and the efficiency conclusion built on it are not empirically distinguishable from run-to-run noise. In addition, no learning-curve baselines are reported (e.g., LoRA-only at 3 or 6 epochs, ReFT-only at 3 epochs, HEFT 3+0 or 0+3), so even a real gap would not isolate the hierarchical composition as the cause. The limitations section acknowledges task, order, and hyperparameter scope, but does not address the missing replication.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes HEFT (Hierarchical Efficient Fine-Tuning), a two-stage PEFT strategy that first applies LoRA in weight space and then applies ReFT in representation space, with the rationale that LoRA provides a coarse, broad adaptation and ReFT refines task-specific representations. The method is evaluated on BoolQ using Llama-2-7B-chat, reporting that HEFT with 3+3 epochs reaches 85.17% accuracy, exceeding LoRA-only at 20 epochs (85.05%) and ReFT-only at 20 epochs (83.36%), while using less wall-clock time. A 20+20-epoch HEFT variant reaches 85.47%. The paper interprets these results as evidence of a synergistic accuracy/compute gain and frames the contribution as a principled coarse-to-fine composition of heterogeneous PEFT methods.","tokens_in":13506,"tokens_out":3389,"duration_ms":398963,"significance":"If the central claim were statistically established, the paper would make a useful empirical contribution to the growing literature on composing PEFT methods, and the public code and detailed experimental appendix would support reproducibility. The efficiency gains in wall-clock time are plausible and potentially interesting. However, the headline result currently rests on a very small accuracy difference with no estimate of run-to-run variability, so the claimed 'profound synergistic effect' is not yet supported by the evidence presented.","major_comments":[{"comment":"The central claim that HEFT outperforms LoRA-only rests on a gap of 0.12 percentage points (85.17% vs 85.05%), which is about 4 examples out of the 3,270 validation items. With a single run per condition and no reported seeds, confidence intervals, or significance tests, the observed difference is within one standard error of a Bernoulli proportion at p≈0.85 and n=3,270 (SE≈0.62 pp). The paper therefore does not currently establish that HEFT is more accurate than LoRA-only; the authors should report multiple seeds with mean±std or confidence intervals, and should apply a paired test (e.g., McNemar's test on the same validation examples) to support the accuracy advantage.","section":"Abstract, Section 4.1, Table 1"},{"comment":"The experimental design lacks matched-epoch and ablation baselines that would isolate the effect of the hierarchical composition. Without LoRA-only at 3 epochs, ReFT-only at 3 epochs, HEFT with 3+0 epochs, and HEFT with 0+3 epochs, the reported 85.17% cannot be attributed to the LoRA-then-ReFT hierarchy rather than to the total number of training epochs or to the mere addition of any second-stage fine-tuning method. The wall-clock time comparison is also hardware-dependent; matched-compute and matched-parameter comparisons would be more informative for the efficiency claim.","section":"Section 3.4, Table 1"},{"comment":"The rows labeled 'Our Zero-Shot HEFT Results' are not zero-shot: the HEFT model was fine-tuned on the BoolQ training split. Comparing a fine-tuned model against zero-shot base models conflates fine-tuning benefits with base-model capability, and the claim that a fine-tuned 7B model 'surpasses' much larger zero-shot models is therefore misleading. For a fair context, the authors should report the zero-shot accuracy of Llama-2-7B itself or compare with fine-tuned baselines from the same setup.","section":"Section 4.2, Table 2"}],"minor_comments":[{"comment":"The phrase 'three epochs' in the abstract and introduction is ambiguous; the method actually uses 3 epochs of LoRA plus 3 epochs of ReFT. Please state '3+3 epochs' consistently.","section":"Abstract and Section 3.4"},{"comment":"There is a typo: '2 hours and 19 minutes minutes' should read '2 hours and 19 minutes'.","section":"Section 4.1"},{"comment":"The code does not set a random seed. For reproducibility, please specify fixed seeds for the model, data loader, and training loop, and report the seed(s) used for each reported run.","section":"Appendix A.1"},{"comment":"The hardware description only states 'at least 32GB of VRAM'; please identify the exact GPU model(s) used, since wall-clock training times are not comparable across different accelerators.","section":"Table 1 and Section 3.2"},{"comment":"The reference for Bian et al. [2025] lists the arXiv identifier as 'abs/2504.0743', which appears incomplete; please verify the full identifier.","section":"References"},{"comment":"The scatter plot would benefit from error bars and a larger font; as printed, it is difficult to read the labels and points, which is important because the efficiency claim is one of the main results.","section":"Figure 1"}],"recommendation":"major_revision","confidential_remarks":"The paper is a single-task, single-model empirical study with a strong promotional tone in the abstract. The main technical weakness is statistical: the accuracy gap that carries the synergistic claim is about 4 out of 3,270 examples with no variance estimate. I would ask the authors to either provide repeated runs with significance testing or substantially temper the 'profound synergistic effect' wording. The public code is a positive feature, as is the explicit limitations section, but the limitations do not address the missing replication or the matched-epoch baseline gap."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nHere's my read on 2509.09801. The paper proposes HEFT: run LoRA for a few epochs, merge, then run ReFT on top. The specific composition is new in the cited literature—prior work mostly combines homogeneous adapters—and the appendix ships a runnable script, which is real credit. The writing is clear, the citations look right, and the limitations section is honest about the single task, model, and order.\n\nThe problem is the evidence for the central claim. The headline result is that HEFT 3+3 hits 85.17% on BoolQ validation versus 85.05% for LoRA-only and 83.36% for ReFT-only, both trained for 20 epochs. The decisive comparison is HEFT versus LoRA-only: a gap of 0.12 percentage points, about 4 examples out of 3,270. For a binary accuracy near 0.85, one standard error is roughly 0.6 percentage points, so the gap is well within noise. No repeated runs, no seeds, no confidence intervals, no significance test anywhere. The abstract's 'profound synergistic effect' is not supported by these numbers.\n\nThe efficiency claim has a similar hole. HEFT 3+3 took 1:23, but the only baselines are 20-epoch LoRA and 20-epoch ReFT. Without LoRA-only at 3 or 6 epochs, or ReFT-only at 3 epochs, you can't tell whether the speedup comes from the hierarchy or just from training less. ReFT-only at 3 epochs might take ~20 minutes and be faster than HEFT. The paper also omits single-stage HEFT controls (3+0 or 0+3), so even a real accuracy difference wouldn't isolate the composition.\n\nMinor but worth noting: the abstract says 'only three epochs' when HEFT is actually 3+3, six total epochs. And Table 2 compares fine-tuned 7B against zero-shot base models; that's apples-to-oranges and doesn't tell us much.\n\nOverall the idea is plausible and cheap to test, and the paper is a reasonable preliminary report. But the empirical support for the headline claim isn't there. I'd send it to peer review rather than desk reject—the script and the specific composition merit a referee's attention—but the right recommendation after review is probably major revision or reject, once the author runs multiple seeds, reports variance, and adds matched-compute baselines.\n\nIf you're curious, you can replicate the experiment on one GPU in under two hours. I'd say maybe bring it to reading group as an example of why error bars matter.\n\nBest,\n[Your name]","headline":"A plausible two-stage PEFT pipeline whose central synergistic claim rests on a 0.12 percentage point gap with no error bars or matched-compute baselines.","tokens_in":14056,"tokens_out":5315,"would_cite":false,"duration_ms":37548,"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":"Sequential LoRA-then-ReFT fine-tuning on a 7B model hits 85.17% on BoolQ after 3 epochs per stage, beating 20-epoch single-method baselines.","keywords":["parameter-efficient fine-tuning","LoRA","ReFT","hierarchical composition","BoolQ","coarse-to-fine","large language models","reasoning"],"falsifier":"Run HEFT-3+3, LoRA-20, and ReFT-20 ten times each with different seeds on the BoolQ validation split and compare mean accuracies with 95% confidence intervals; if the intervals overlap, the synergistic advantage vanishes. Alternatively, apply HEFT to a second reasoning benchmark and check whether the 3+3 config still beats 20-epoch single-method baselines.","tokens_in":13002,"feed_emoji":"⚡","tokens_out":6171,"duration_ms":46383,"temperature":0.7,"pith_summary":"This paper proposes a two-stage fine-tuning recipe called HEFT that combines two parameter-efficient methods in sequence: first LoRA, which edits the model's weights with low-rank updates, and then ReFT, which surgically edits the model's internal activations. On the BoolQ yes/no reasoning benchmark, a Llama-2-7B model trained for just three epochs of each stage reaches 85.17 percent accuracy, slightly above the 85.05 percent achieved by LoRA alone after twenty epochs and well above ReFT alone at 83.36 percent. The same two-stage recipe run for twenty epochs each reaches 85.47 percent. The claim is that this coarse-to-fine order—broad weight-space adaptation before precise representation-space intervention—produces a better accuracy-per-compute trade-off than either method alone.","feed_headline":"LoRA-then-ReFT out-tunes either method alone","feed_subtitle":"A 7B model taught in two passes reaches 85.17% on BoolQ in under 90 minutes—above 20-epoch baselines.","key_machinery":"The mechanism is the hierarchical two-stage pipeline itself. Stage 1 is LoRA: the base Llama-2-7B model's weights are frozen and a rank-8, alpha-32 low-rank adapter is trained and then merged into the weights, producing a consolidated 'coarse-adapted' base model. Stage 2 is ReFT via a LoReFT intervention: the merged model is frozen again, and a low-rank linear subspace intervention (dimension 4) is trained on the block-output representation of transformer layer 15, directly editing hidden representations during the forward pass. The argument is carried by this order: a global weight-space shift first, then a local, interpretable representation-space edit.","core_discovery":"The central discovery the paper reports is a synergy between weight-space and representation-space fine-tuning: on the BoolQ benchmark, composing LoRA first and ReFT second (HEFT) yields accuracy higher than either method achieves in isolation, and does so with far less training time. The paper's headline measurement is that a 3-epoch LoRA stage followed by a 3-epoch LoReFT stage reaches 85.17% validation accuracy, beating a 20-epoch LoRA-only run (85.05%) and a 20-epoch ReFT-only run (83.36%), while completing in 1 hour 23 minutes versus 6 hours 52 minutes and 2 hours 19 minutes. A 20+20-epoch HEFT run reaches 85.47%, sustaining the advantage. The paper interprets this as a coarse-to-fine effect: the LoRA stage moves the model into a favorable parameter region, and the ReFT stage then performs fine-grained editing of the internal representations within that region.","pith_inferences":["The reported 0.12-point gap between HEFT-3+3 and LoRA-20 corresponds to roughly four validation examples out of 3,270; with no seeds or confidence intervals reported, the synergy claim is likely within run-to-run noise until replicated.","The efficiency comparison is epoch-based, not compute-based; a LoRA-only run stopped at three epochs would be much faster (though lower-accuracy), so the headline 'fraction of the compute' claim depends on matching performance targets, not matched budgets.","The coarse-to-fine principle suggests a testable prediction: on benchmarks where ReFT alone is strong (e.g., commonsense reasoning), HEFT's gain over LoRA-only should be larger than on BoolQ; on generation-heavy tasks, the ReFT stage may add little or hurt.","The paper's modularity idea implies a concrete architectural experiment: train one LoRA base, then attach separate ReFT heads for BoolQ and, say, CommonsenseQA; if interference is low, HEFT becomes a building block for multi-skill specialization."],"forward_implications":["A 7B model can be specialized to a reasoning task to a level competitive with 70B-scale zero-shot models using about 1.4 GPU-hours of fine-tuning.","The LoRA-then-ReFT order yields a reusable LoRA-adapted 'foundation' model onto which multiple lightweight ReFT interventions could be layered for different sub-skills.","The 20+20 HEFT result (85.47%) shows the synergy holds beyond the early-training efficiency win, not merely as a fast-start artifact.","The method's success on BoolQ suggests that heterogeneous PEFT composition is a distinct algorithmic lever, separate from choosing a single best method.","Training time drops from 6:52 (LoRA-only) or 2:19 (ReFT-only) to 1:23 for the 3+3 HEFT configuration while accuracy rises above both."],"supporting_citations":[{"why":"Introduces LoRA, the weight-space low-rank adaptation used in Stage 1 of HEFT.","marker":"[Hu et al., 2021]"},{"why":"Introduces ReFT and the LoReFT intervention used in Stage 2 of HEFT.","marker":"[Wu et al., 2024]"},{"why":"Provides the BoolQ dataset and its validation split used for fine-tuning and evaluation.","marker":"[Clark et al., 2019]"},{"why":"Provides the Llama-2-7B base model and the zero-shot Llama-2-70B baseline used for comparison.","marker":"[Touvron et al., 2023b]"},{"why":"Supplies the 'intruder dimensions' critique of LoRA that motivates adding a representation-space refinement stage.","marker":"[Shuttleworth et al., 2024]"},{"why":"Documents prior work on combining PEFT modules (mostly homogeneous), the contrast for this paper's heterogeneous composition.","marker":"[Asadi et al., 2024]"}],"fun_headline_variants":["Two-pass fine-tune beats single-method marathons","Coarse-to-fine PEFT: less training, better reasoning","LoRA first, ReFT second: 7B model wins in 83 minutes","Hierarchical PEFT hits 85.17% on BoolQ in 3 epochs","Weight-space then representation-space: a winning combo"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The claimed synergy rests on the assumption that HEFT's 0.12-percentage-point edge over LoRA-only after three epochs is a real difference and not run-to-run noise, but the paper reports a single run with no seeds, repeats, or confidence intervals.","fun_headline_variants_meta":{"raw":{"variants":["Two-pass fine-tune beats single-method marathons","Coarse-to-fine PEFT: less training, better reasoning","LoRA first, ReFT second: 7B model wins in 83 minutes","Hierarchical PEFT hits 85.17% on BoolQ in 3 epochs","Weight-space then representation-space: a winning combo"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000233,"raw_usage":{"total_tokens":1547,"prompt_tokens":1053,"completion_tokens":494,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":669,"completion_tokens_details":{"reasoning_tokens":400}},"tokens_in":669,"tokens_out":494,"duration_ms":4858,"temperature":1.0,"reasoning_tokens":400,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T15:58:23.357779+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run HEFT-3+3, LoRA-20, and ReFT-20 ten times each with different seeds on the BoolQ validation split and compare mean accuracies with 95% confidence intervals; if the intervals overlap, the synergistic advantage vanishes. Alternatively, apply HEFT to a second reasoning benchmark and check whether the 3+3 config still beats 20-epoch single-method baselines.","supporting_citations":[],"review_version":2}