{"id":"a97b5563-9722-447f-9262-cb6931657614","arxiv_id":"2509.07676","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"A feedback-triggered regeneration method combined with dynamic multipath decoding improves LLM accuracy on math and coding benchmarks compared to prompt-based self-correction.","lead":"Researchers at Tencent propose a self-correction framework for large language models that regenerates an answer only when the user signals it is wrong, and then uses a beam-search-like decoding that keeps multiple candidate reasoning paths. On math and code benchmarks, the method beats two prompt-based self-correction baselines across several open-source models from 1B to 13B parameters.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"LTM's min-PPL answer selection (Eq. 2, §3.2) is unvalidated on math/code; if PPL tracks fluency rather than correctness, the LTM contribution to FTR's gains is unsupported.","rationale":"The reader's weakest assumption identifies the same load-bearing concern: LTM's final answer selection by minimum PPL is the mechanism by which multipath decoding claims to improve over beam search and sampling, yet the paper provides no evidence that PPL correlates with correctness on math or code. This concern is specific, testable, and central to the paper's second contribution. The protocol-fairness issue in Protocol 1 is real but partially mitigated by Protocol 2, which applies the same proxy trigger to all methods; the LTM/PPL issue is not mitigated anywhere. The paper's own data (Table 5 vs Table 1) suggest the feedback trigger alone produces most of the gain over baselines, so if the PPL-based selection is invalid, the marginal contribution of LTM remains unsupported. The proposed concrete test would settle this by measuring whether min-PPL selects correct answers better than simple majority voting and whether the candidate set even contains the correct answer. The reader's conditional verdict is appropriate: the concern is serious but addressable with an additional experiment, not a fatal flaw, so the verdict should remain CONDITIONAL (no change).","tokens_in":13469,"tokens_out":12083,"duration_ms":109915,"concrete_test":"Run LTM decoding on GSM8K and MultiArith with Llama3-3B and Qwen-1.5B. For each problem, record the full set of terminal candidate sequences before the final PPL selection, then compare four selection strategies: (1) min-PPL as in the paper, (2) majority voting over parsed final answers, (3) random choice among candidates, and (4) oracle-best selection. Also compute coverage: the fraction of problems where the correct answer is present among the retained candidates. If majority-vote accuracy matches or exceeds min-PPL accuracy, the stated PPL selection is not load-bearing; if coverage is low, the pruning thresholds discard the correct path.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claimed benefit of LTM over standard decoding rests on using PPL to pick the final answer from multiple candidate paths (\"ultimately selecting the optimal answer through minimum PPL evaluation\"). On reasoning and code tasks, model likelihood is not known to track correctness, and the paper provides no evidence that lower PPL selects the correct reasoning path. Because LTM's pruning thresholds in Eq. 4-5 also operate on sequence probabilities, the entire decoding mechanism is built on this unvalidated assumption. The potential failure mode is concrete: a verbose, fluent but wrong chain of thought can receive lower PPL than a terse correct answer, so the selected output can be confidently incorrect. Table 2 compares LTM only against likelihood-based baselines (beam search, sampling) and never against a correctness-aware selection such as self-consistency majority voting. Moreover, the FTR results in Table 1 do not ablate the final PPL selection criterion, so the reader cannot tell whether the reported gains come from multipath exploration, from the PPL ranking, or from the feedback trigger alone. If PPL is miscalibrated, the small but consistent LTM improvements in Table 2 could vanish or reverse, undermining the paper's second contribution even if the feedback-trigger idea remains viable.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Feedback-Triggered Regeneration (FTR): when user feedback marks an LLM output as incorrect, the model regenerates from the original prompt using Long-Term Multipath (LTM) decoding, a tree-search decoder that prunes candidate sequences by cumulative probability and selects the final answer by minimum perplexity. Experiments on GSM8K, MultiArith, and HumanEval with six open-weight LLMs compare FTR with Critic Prompt and IoE Prompt self-correction baselines under a ground-truth-triggered protocol and a GPT-4o proxy protocol, and separately compare LTM with greedy, beam, nucleus, and adaptive decoding. The paper reports consistent gains of 10%–20% over the baselines and argues that prompt-based self-correction often degrades initial accuracy.","tokens_in":13722,"tokens_out":7444,"duration_ms":68408,"significance":"If the results hold, the feedback-triggered design makes a useful conceptual point: it replaces the model's unreliable self-assessment with an external trigger, preserving correct initial answers and focusing computation on incorrect ones. The LTM decoder is also an interesting alternative to fixed-width beam search, with an explicit mechanism for dynamically widening the beam at uncertain steps. The paper has concrete strengths: it evaluates six model sizes, includes a human-feedback proxy with reported false-positive/negative rates, and provides an inference-time analysis and a case study. However, the headline comparison in Protocol 1 is structurally unfair, the LTM hyperparameters appear to be selected on the test sets, and the min-PPL selection rule is unvalidated as a correctness criterion. These issues are load-bearing for the central claims and must be addressed before the results can be accepted.","major_comments":[{"comment":"The Protocol 1 comparison is structurally unfair and the reported gains over the initial output are guaranteed by the protocol. FTR is given the ground-truth label to decide when to regenerate, whereas the Critic and IoE baselines must self-assess every output. Since FTR by construction never regenerates a correct initial answer, its accuracy is initial-correct plus corrected-incorrect, which cannot fall below the Initial Input score; the baselines can and do degrade correct answers. For example, on Llama2-7B GSM8K, Initial Input is 0.206, Critic drops to 0.171, IoE drops to 0.136, and FTR rises to 0.360. The gain of FTR over Initial Input is therefore not an empirical discovery about the method's quality, and the 10%–20% claim in Section 5.1 does not by itself validate FTR. The paper should compare FTR against baselines that receive the same external trigger (as Protocol 2 attempts), and should report FTR with standard nucleus sampling under the same trigger to isolate the contribution of LTM.","section":"§5.1, Table 1 (Protocol 1)"},{"comment":"The LTM hyperparameters p* and k* appear to be tuned per dataset and per model on the test sets. Table 6 reports different (p*, k*) pairs for every model-dataset combination, and the same chosen values are used to produce the results in Tables 1 and 2. The paper does not describe a validation split, a tuning procedure, or a sensitivity analysis, and no variance or confidence intervals are reported. This is particularly concerning because several LTM gains are small (e.g., 0.852 vs 0.851 for Qwen-3B on GSM8K in Table 2); test-set selection of two free parameters could account for such differences. The authors should tune on held-out data, report sensitivity to p* and k*, and provide multiple-seed results with confidence intervals.","section":"§5.3, Table 6"},{"comment":"The central decoding contribution of LTM rests on an unvalidated assumption that lower perplexity selects the correct answer. The final answer is selected by minimum PPL evaluation (Section 3.2), and the pruning thresholds in Eqs. (4) and (5) operate on sequence probabilities; on mathematical reasoning and code generation, token likelihood is not established as a correctness signal. A verbose, fluent but incorrect chain of thought can easily receive a lower PPL than a terse correct answer. Table 2 compares LTM only against likelihood-based decoding baselines and never against a correctness-aware selection rule such as self-consistency majority voting over sampled paths. The authors should provide evidence that min-PPL selection recovers correct answers on these tasks (e.g., a correlation analysis between PPL and correctness, or an ablation comparing min-PPL selection with majority voting); without this, the LTM contribution to FTR's gains is unsupported.","section":"§3.2, Eq. (2)–(5), Table 2"}],"minor_comments":[{"comment":"The text says the probability is the product of the likelihoods of the first i tokens, but the formula includes y_0 through y_i, so it should refer to the first i+1 tokens.","section":"Eq. (1)"},{"comment":"The figure caption contains the undefined acronym \"LTPD\" in the illustration; this appears to be a typo for LTM and should be fixed.","section":"Figure 4"},{"comment":"The efficiency discussion should be reconciled with Table 3: FTR's measured inference time exceeds the 2× cost of the Critic/IoE baselines in most cells (e.g., Llama2-7B on HumanEval is 3.87× vs 2×), while the text claims \"net efficiency gains in low error rate scenarios\"; the condition p*n < 1 is not enough given the reported numbers and should be restated with the actual measured overheads.","section":"§A.2, Table 3"},{"comment":"The phrase \"state-of-the-art prompt-based self-correction methods\" overstates the comparison, since only two prompting baselines (Critic Prompt and IoE Prompt) are evaluated; the authors should either add more baselines or soften the wording.","section":"Abstract and §1"},{"comment":"Reference [30] contains a typo in the author name \"Amir Globersons\" and should be corrected to \"Amir Globerson\".","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The paper does not mention code release; the authors should be encouraged to provide code and hyperparameter-search logs if the manuscript is revised. The Protocol 1 unfairness and the PPL-selection validation are the two decisive issues; both are fixable within the scope of the paper, so I do not recommend rejection at this stage."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The thing worth knowing here is the feedback-triggered design: instead of asking the LLM to self-critique, treat user feedback as a binary switch and only regenerate on negative signal. That is simple, sensible, and directly usable in interactive systems. The paper's own experiment comparing feedback-as-indicator against feedback-as-prompt (Figure 5) is the cleanest result in the paper and supports this design choice. I also credit the breadth of the evaluation: six open models from 1B to 13B across three benchmarks, with a GPT-4o proxy protocol that gets closer to real feedback conditions than the ground-truth protocol does.\n\nThe soft spots are real and the biggest one is structural. In Protocol 1, FTR is told which initial answers are wrong using ground-truth labels, while the Critic and IoE baselines are given no such signal. So FTR never touches correct answers, and the baselines are free to corrupt them. The 10-20% gains in Table 1 under Protocol 1 are partly guaranteed by the protocol, not discovered. Protocol 2 is fairer, but the same table does not separate the contribution of the feedback trigger from the contribution of LTM decoding; there is no ablation of LTM inside FTR. Table 6 also shows p* and k* tuned per dataset on the benchmark itself, and no error bars are reported, so the consistency claims are weaker than they look.\n\nThe weakest technical assumption is the minimum-PPL selection rule in LTM. The paper provides no evidence that lower sequence perplexity tracks correctness on math or code. A verbose, fluent wrong chain of thought can easily have lower PPL than a terse correct one. Since LTM's pruning and final selection both rely on sequence probability, this is load-bearing for the decoding contribution. The right missing control is self-consistency majority voting, or any correctness-aware selection, and the paper does not include it. That said, LTM's gains over beam search in Table 2 are small but consistent; they may be real, but the current evidence does not show why PPL is the right criterion.\n\nNone of this kills the paper. The core feedback-triggered idea is novel enough in combination, the writing is clear, and the limitations section is honest about the redundancy issue. The flaws are addressable with a validation split, an FTR-without-LTM ablation, and a selection-aware baseline. I would send this to peer review because the idea deserves scrutiny and the experiments, once unconfounded, could support a solid conditional accept. The current version is not ready as-is; the headline claims need to be rescaled to Protocol 2 and the LTM mechanism needs a correctness-aware comparison.","headline":"A practical feedback-triggered regeneration idea with a mostly known multipath decoding twist, but the headline comparison is rigged in FTR's favor and the PPL-based answer selection is unvalidated.","tokens_in":14249,"tokens_out":1486,"would_cite":false,"duration_ms":14795,"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":"The paper argues that LLM self-correction succeeds when regeneration is triggered by external feedback rather than by the model's own critique, and that multipath decoding with minimum-perplexity selection carries the improvement.","keywords":["self-correction","feedback-triggered regeneration","long-term multipath decoding","perplexity-based selection","mathematical reasoning","code generation","decoding strategies","large language models"],"falsifier":"Run LTM over a held-out set of GSM8K or HumanEval problems, record every generated candidate with its perplexity, and compare the accuracy of the minimum-perplexity answer against a majority vote over final answers and against a random candidate; if a wrong answer with lower perplexity than the correct answer is found, or if non-PPL selection does as well or better, the paper's claimed source of gain is contradicted.","tokens_in":13252,"feed_emoji":"🎯","tokens_out":10092,"duration_ms":76639,"temperature":0.7,"pith_summary":"The paper claims that self-correction of large language models fails because the model itself is a poor judge of its own answers, and because next-token decoding cannot look far enough ahead to repair a bad reasoning path. To fix this, it proposes Feedback-Triggered Regeneration (FTR), which regenerates a response only when negative user feedback arrives and, in that second pass, decodes with Long-Term Multipath (LTM) sampling that keeps multiple candidate paths and picks the final answer by lowest perplexity. On math word problems (GSM8K, MultiArith) and code generation (HumanEval), using six open-source models from 1B to 13B parameters, FTR reports consistent 10–20% accuracy gains over the initial output and over two prompt-based self-correction baselines. The practical stake is that deployed assistants could improve their answers without the instability that self-evaluation prompts introduce.","feed_headline":"Only retry wrong answers: LLM math/code gains 10-20%","feed_subtitle":"Prompt-based correction often ruins right answers; FTR retries only on user complaints, exploring many paths.","key_machinery":"Long-Term Multipath (LTM) decoding is the load-bearing mechanism: it maintains a dynamic set of $k_i$ candidate sequences at each step, chosen as the smallest set whose cumulative probability satisfies $\\sum_{j=0}^{k_i} P(s_i^j) \\geq p^* \\sum_{j=0}^{k_{i-1}V-1} P(s_i^j)$, subject to a hard cap $k^*$, and the final answer is selected by minimum perplexity $PPL_i = P(s_i)^{-1/(i+1)}$ over the surviving paths. FTR is the other half: user feedback is a binary signal that either accepts the initial output or triggers regeneration of the original prompt with LTM, so the model never receives a corrective prompt that could bias it.","core_discovery":"The central discovery is that two-stage self-correction works better when the second stage is triggered externally rather than by the model's own judgment. Prompt-based correction with Critic prompts and If-or-Else prompts frequently turns correct initial answers into wrong ones; FTR avoids this by using user feedback as a binary 'regenerate' trigger and re-running the original input without any corrective prompt. In the regeneration stage, LTM decoding replaces greedy next-token search: at each step it keeps the smallest set of candidate sequences whose cumulative probability exceeds a threshold, caps the set at $k^*$, and after generating all candidates emits the answer with the minimum perplexity, defined as the inverse geometric mean of token probabilities. In the reported experiments, FTR improves accuracy substantially over initial outputs and over both prompt-based baselines, and LTM alone outperforms greedy, beam search, and adaptive decoding in single-turn generation.","pith_inferences":["A natural next experiment is to replace minimum-PPL selection inside LTM with a lightweight task verifier (running unit tests for code, or exact answer checking for arithmetic); the paper does not report this ablation, and the gains could be larger or smaller depending on how well perplexity tracks correctness.","The efficiency advantage over double-pass baselines depends on negative feedback being relatively rare; in applications where users reject most outputs, the multipath regeneration cost would approach or exceed the 2× cost of the prompt-based baselines.","The adaptive-width pruning in LTM can be viewed as a randomized beam search whose width reacts to the flatness of the token distribution, so the reported gains over fixed-width beam search may come substantially from that adaptivity rather than from 'long-term' evaluation per se.","FTR's trigger design suggests a direct deployment test: log real thumbs-up/thumbs-down signals over a few weeks and measure whether accuracy on accepted versus regenerated answers shifts in the direction the paper's protocols predict."],"forward_implications":["FTR consistently outperforms both the initial output and the two prompt-based correction baselines on GSM8K, MultiArith, and HumanEval across six open models from 1B to 13B parameters, with Protocol 1 gains of about 10–20%.","Using feedback only as a regeneration trigger beats embedding the feedback in a corrective prompt, indicating that the corrective prompt itself causes much of the degradation seen in prompt-based self-correction.","LTM decoding alone, without any self-correction wrapper, beats greedy decoding, fixed-width beam search, and adaptive top-k decoding on the same benchmarks.","Because regeneration is skipped for accepted answers, FTR's total compute is about $N \\times t \\times (1 + p \\times n)$ instead of the baselines' $2 \\times N \\times t$, making it cheaper than mandatory double-pass correction when the regeneration rate stays below roughly 54%.","A noisy automated proxy for human feedback (GPT-4o) is enough to reproduce most of the gains, suggesting the method does not depend on perfect or oracle error signals."],"supporting_citations":[{"why":"Supplies the Critic Prompt baseline and the finding that LLMs cannot reliably self-correct reasoning.","marker":"[11]"},{"why":"Supplies the IoE (If-or-Else) confidence-based self-correction baseline and its comparison setup.","marker":"[9]"},{"why":"GSM8K dataset used for the math reasoning evaluation.","marker":"[16]"},{"why":"MultiArith dataset used for arithmetic word problem evaluation.","marker":"[17]"},{"why":"HumanEval dataset used for code generation evaluation.","marker":"[18]"},{"why":"Defines nucleus sampling, the decoding strategy used for initial outputs and as a decoding baseline.","marker":"[22]"},{"why":"Defines adaptive decoding, an entropy-based dynamic top-k baseline that LTM is compared against.","marker":"[29]"},{"why":"GPT-4o is used as the automated proxy for human feedback in Protocol 2.","marker":"[23]"},{"why":"Defines perplexity as the quality metric used by LTM to evaluate and select candidate sequences.","marker":"[12]"}],"fun_headline_variants":["Retry only on negative feedback: LLM math/code boost","Retry on user complaints, not self-doubt, boosts LLM accuracy","Don't self-correct: let users trigger retry, explore more paths","FTR: retry on feedback, explore paths, beat self-correction"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method assumes that the candidate answer the model finds most fluent, as measured by the lowest perplexity, is also the one most likely to be correct.","fun_headline_variants_meta":{"raw":{"variants":["Retry only on negative feedback: LLM math/code boost","Retry on user complaints, not self-doubt, boosts LLM accuracy","Don't self-correct: let users trigger retry, explore more paths","FTR: retry on feedback, explore paths, beat self-correction"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001563,"raw_usage":{"total_tokens":6229,"prompt_tokens":919,"completion_tokens":5310,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":535,"completion_tokens_details":{"reasoning_tokens":5229}},"tokens_in":535,"tokens_out":5310,"duration_ms":31044,"temperature":1.0,"reasoning_tokens":5229,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T16:12:24.190943+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run LTM over a held-out set of GSM8K or HumanEval problems, record every generated candidate with its perplexity, and compare the accuracy of the minimum-perplexity answer against a majority vote over final answers and against a random candidate; if a wrong answer with lower perplexity than the correct answer is found, or if non-PPL selection does as well or better, the paper's claimed source of gain is contradicted.","supporting_citations":[{"cited_title":"Large language models cannot self-correct reasoning yet","cited_arxiv_id":null,"evidence_quote":"Supplies the Critic Prompt baseline and the finding that LLMs cannot reliably self-correct reasoning."},{"cited_title":"Solving general arithmetic word problems","cited_arxiv_id":null,"evidence_quote":"MultiArith dataset used for arithmetic word problem evaluation."},{"cited_title":"The curious case of neural text degeneration","cited_arxiv_id":null,"evidence_quote":"Defines nucleus sampling, the decoding strategy used for initial outputs and as a decoding baseline."},{"cited_title":"Improving open-ended text generation via adaptive decoding","cited_arxiv_id":null,"evidence_quote":"Defines adaptive decoding, an entropy-based dynamic top-k baseline that LTM is compared against."}],"review_version":2}