{"id":"0174d382-225c-4872-9b77-21d6c50a210c","arxiv_id":"2506.00189","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":3,"one_line_summary":"Control-R-32B is a fine-tuned 32B model that takes 11 control-field scores in the prompt, claiming to let users trade reasoning depth for efficiency and reaching SOTA on AIME2024 and MATH500 at that scale.","lead":"The paper trains a 32B language model to follow 11 prompt-injected 'reasoning control fields' meant to adjust how deeply the model searches before answering. It shows accuracy gains on math benchmarks, but provides little direct evidence that the control fields actually change reasoning effort rather than just improving the base model.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Controllability claim is unsupported: no direct measurement of reasoning effort, and the only proxy (accuracy) shows weak sensitivity to control values while confounded by outcome-label fields.","rationale":"The reader's weakest assumption correctly identifies that controllability is inferred only from accuracy changes without direct process measurement. I sharpen this: (1) the existing accuracy evidence in Table 3 shows only small sensitivity to control values (0/5/9), so even the proxy is weak; (2) the test-time prompt includes outcome-leaking fields (correctness: 9, knowledge_accuracy: 9), which could account for the accuracy gain independently of any effort control; and (3) the annotation supervision is self-referential because ChatGPT-4o scores traces from the same teacher models, so the control-score-to-process mapping is not independently validated. These points strengthen the reader's conditional verdict rather than overturning it. The proposed concrete test directly measures whether reasoning effort (tokens, steps, strategies) tracks the execution-control fields with quality fields held constant; this is the decisive missing experiment. No machine-checked proofs or released artifacts exist to independently support the central claim, so the conditional verdict remains appropriate pending this evidence.","tokens_in":15837,"tokens_out":4255,"duration_ms":45237,"concrete_test":"Run a controlled decoding experiment on 100 AIME2024 problems with Control-R-32B under four conditions: no control fields; all execution-control fields set to 0/5/9 while quality fields (correctness, knowledge_accuracy, etc.) remain fixed at 9; and all fields set to 9. For each condition, record average generated token count, number of distinct reasoning steps, frequency of backtracking/strategy-switch phrases, and Pass@1. If token count and step count do not increase monotonically across the 0/5/9 execution-field conditions, or if accuracy is flat while execution fields vary, the claim that RCFs control reasoning effort is falsified.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that RCFs enable controllable test-time scaling, i.e., the model adjusts reasoning effort (search depth/breadth, error handling, strategy switching) according to the 11 control fields. To support this, the paper must show that changing control values actually changes the reasoning process, not just the final answer. Table 3 provides only accuracy numbers: Control-R-32B jumps from 6.67% to 63.3% on AIME2024 when the control string is present, but varying all fields from 0 to 5 to 9 changes accuracy only from 63.3% to 63.3% to 70.0% (and 91.4/91.8/93.2 on MATH500). This is weak evidence of fine-grained control, and the large presence/absence gap suggests the model has learned a brittle format dependency rather than a graded reasoning dial. Moreover, the test-time prompt sets all fields to 9, including outcome-leaking fields 'correctness' and 'knowledge_accuracy', so any accuracy gain could come from the model being prompted to 'be correct' rather than from modulating search effort. The paper never measures token length, step count, backtracking frequency, or strategy composition under different control values, so the 'controllable Long CoT reasoning process' remains an indirect inference. The self-referential annotation (ChatGPT-4o labeling traces from the same models later used as training targets) further weakens the link between control scores and actual process attributes.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Reasoning Control Fields (RCFs), an 11-dimension set of integer-valued control signals appended to prompts, together with a Conditional Distillation Finetuning (CDF) procedure and a new dataset, Control-R-4K, to train a 32B model (Control-R-32B) that can adjust long chain-of-thought reasoning effort at test time. Control-R-32B is evaluated on AIME2024, MATH500, GPQA-Diamond, and AIME2025 Part I, reporting 70.0%, 93.2%, 61.1%, and 55.0%, respectively. The central claim is that RCFs enable 'controllable Long CoT reasoning' and state-of-the-art performance at the 32B scale.","tokens_in":16320,"tokens_out":2876,"duration_ms":30156,"significance":"If the central claim were substantiated, the paper would offer a useful, lightweight mechanism for trading reasoning depth against compute in long-CoT models, and the Control-R-4K dataset with per-trace attribute annotations would be a reusable resource for conditional reasoning distillation. The paper also gives a clear probabilistic framing of conditional generation over reasoning traces. However, the evidence presented does not currently demonstrate that the control fields modulate reasoning behavior: accuracy differences are small, unreplicated, and confounded by outcome-related prompt fields, and no direct measurement of reasoning effort under differing control values is reported. The MATH500 result is also likely in-distribution because the training data includes a random sample of MATH, of which MATH500 is a subset.","major_comments":[{"comment":"The central claim that RCFs enable controllable test-time scaling is not directly supported by the reported experiments. Table 3 reports only accuracy under all-fields-equal-to-0/5/9; it never reports token length, number of reasoning steps, backtracking frequency, strategy-switch counts, or any other process-level measure under different control values. Table 2 reports token lengths and 'Wait' occurrences only by correctness outcome, not by control-field setting. Thus the paper shows that a prompt string affects accuracy, not that it controls the reasoning process.","section":"§3.3, Table 3"},{"comment":"The sensitivity of accuracy to control-field values is weak and likely within noise. On AIME2024, all-0 and all-5 both give 63.3% and all-9 gives 70.0%; on MATH500 the numbers are 91.4%, 91.8%, and 93.2%. The paper uses Pass@1 with n=1 and provides no error bars or repeated-sampling statistics, so the 6.7- and 1.8-point differences do not establish fine-grained control. This is especially important because the test-time prompt sets 'correctness' and 'knowledge_accuracy' to 9, so part of the accuracy gain may come from directly prompting the model to be correct rather than from modulating search effort.","section":"§3.3, Table 3"},{"comment":"The RCF labels are produced by ChatGPT-4o annotating the same long-CoT traces that are later used as training targets, with no inter-annotator agreement, no validation against externally defined reasoning-effort measures, and no evidence that the scores correlate with actual process attributes. This makes the control semantics circular: the model learns to imitate the annotator's score-to-trace mapping rather than to implement an independently verified notion of search depth, breadth, or error correction.","section":"§2.2, Appendix B.3"},{"comment":"The claim of state-of-the-art MATH500 performance is undermined by training-data contamination: the main subset of Control-R-4K includes a random sample of problems from the MATH dataset, and MATH500 is a subset of MATH. Thus the 93.2% figure is in-distribution for Control-R-32B, while the comparison baselines were not trained on this subset. The paper should report results on a held-out, non-overlapping benchmark and should state explicitly which AIME years and MATH splits were excluded from training.","section":"Appendix B.1"},{"comment":"The paper defines pass@k but reports all main results as Pass@1 with n=1. This means each benchmark number is a single deterministic or single-sample run, so the reported differences in the ablation study have no sampling uncertainty. The authors should either report multiple seeds with means and confidence intervals or clearly state that the numbers are single-run point estimates and avoid drawing conclusions from small gaps.","section":"§3.1, pass@k definition"}],"minor_comments":[{"comment":"The first sentence has a grammatical error: 'This paper target in addressing' should be 'This paper targets the challenges of...'.","section":"Abstract"},{"comment":"The label 'Cotrol-R' in Figure 1 is a typo for 'Control-R'.","section":"Figure 1"},{"comment":"The caption reads 'Ablation study on the affection of reasoning control fields'; 'affection' should be 'effect'.","section":"Table 3 caption"},{"comment":"Table A2 reports the metric as Pass@1 (n=4), while §3.1 defines the metric as Pass@1 and Table 1 uses n=1. This inconsistency should be reconciled.","section":"Table A2 caption and §3.1"},{"comment":"The references list Lightman et al. 2023a and 2023b as separate entries for the same paper; this should be a single citation.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The core idea is timely and the dataset could be valuable, but the current manuscript's headline claims outrun the evidence. The most pressing fixes are process-level measurements under different control values, contamination-free evaluation, and validation of the annotation procedure. These are substantial but feasible additions, so I recommend major revision rather than rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper takes SteerLM's attribute-conditioned generation, ports it to long CoT reasoning with 11 hand-defined control fields, and trains a 32B model that hits competitive numbers on AIME2024 and MATH500. That application is new, and the Control-R-4K dataset, with its per-trace control-field annotations, could be useful to people working on reasoning distillation. The method is simple and clearly described, and the ablation shows the control string matters: without it, accuracy collapses, and raising all fields from 0 to 9 improves AIME2024 from 63.3% to 70.0%. That is real evidence of some sensitivity, and I don't want to minimize it.\n\nThe soft spot is the central claim. The paper says RCFs let users dial reasoning effort, but it never measures reasoning effort. No token lengths, step counts, backtracking frequency, or strategy composition under different control values. Table 3 only shows accuracy, and accuracy is a weak proxy — especially when the test-time prompt sets correctness and knowledge_accuracy to 9, which leaks the outcome. The presence/absence gap (6.7% vs 70%) also suggests the model has learned a format dependency rather than a graded dial; the graded response between 0 and 9 is modest, and on MATH500 it's almost flat. The labeling loop is also self-referential: ChatGPT-4o judges traces from DeepSeek-R1/QwQ, and the model trains on those judgments with no external validation that the scores track real process attributes.\n\nTwo smaller issues: MATH500 is sampled from MATH, which appears in the training data, so the headline 93.2% is in-distribution. And the AIME2025 results use Pass@1 with n=4 while other tables use n=1, so the comparison is apples-to-oranges. No artifacts are released, which makes the dataset hard to reuse or verify.\n\nOverall, this is a serious paper about a real problem, but the load-bearing claim — controllable test-time scaling — is not yet supported. The fix is straightforward: measure the reasoning process under different control settings, drop or ablate the outcome-leaking fields, and report error bars. The authors even acknowledge the calibration issue in their limitations, so the gap is recognized rather than hidden.\n\nI'd send this to peer review. The idea is timely, the dataset has value, and the concerns are addressable with additional experiments. But the reviewers should push hard on the controllability evidence before this is accepted.","headline":"Competitive 32B accuracy with SteerLM-style control fields, but the controllability claim rests on accuracy alone — no direct evidence that reasoning effort changes.","tokens_in":16737,"tokens_out":2084,"would_cite":false,"duration_ms":21729,"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":"Control scores in the prompt let a 32B model tune its reasoning depth","keywords":["reasoning control fields","controllable test-time scaling","long chain-of-thought","conditional distillation finetuning","underthinking","overthinking","tree search reasoning","math reasoning benchmarks"],"falsifier":"Run Control-R-32B on a fixed question set with all control scores set to 0 versus 9 and measure actual token counts, number of distinct strategies, and backtracking events in the sampled traces; if the all-9 condition does not produce longer or structurally deeper reasoning while improving accuracy, the central controllability claim is not supported.","tokens_in":1752,"feed_emoji":"🎛️","tokens_out":3043,"duration_ms":91552,"temperature":0.7,"pith_summary":"This paper tries to establish that long chain-of-thought reasoning can be made controllable at test time by conditioning the model on a short string of eleven integer scores, which the authors call Reasoning Control Fields. The scores regulate how deep and broad the search should be, how aggressively errors should be detected and corrected, how often strategies should switch, and how the process should score on correctness, efficiency, completeness, coherence, knowledge accuracy, and clarity. If correct, a user could dial reasoning effort up or down per query instead of accepting whatever depth the model chooses, directly addressing the twin failures of underthinking and overthinking. To support this, the paper builds a curated dataset, proposes a conditional distillation fine-tuning objective, and reports that a 32B model trained this way outperforms comparable 32B reasoning models on AIME2024 and MATH500.","feed_headline":"Eleven prompt scores let a 32B reasoning model dial its thinking depth","feed_subtitle":"With all eleven set to 9, the tuned 32B model hits 70% on AIME2024 and 93.2% on MATH500.","key_machinery":"The Reasoning Control Field (RCF) string is a prompt suffix such as \"<control> search_depth: 8; search_breadth: 7; error_detection: 8; error_correction: 7; strategy_switching: 6; correctness: 9; efficiency: 7; completeness: 8; coherence: 8; knowledge_accuracy: 9; clarity_of_steps: 8 <control/>\", prefixed by a line break. Each of the eleven integer scores, ranging 0 to 9, is intended as a regulator for a specific aspect of the tree-search process, and Conditional Distillation Finetuning forces the model's next-step probabilities to depend on the entire string, so changing any score should change how the model searches.","core_discovery":"On its own terms, the paper establishes that long CoT reasoning can be viewed as an in-context tree search and that a fixed textual vector of eleven control scores can modulate that search. It formalizes generation as $P(R|q,C)$, where $C$ is the control string, and trains with the conditional distillation objective $\\mathcal{L}_{CDF}=\\mathbb{E}[-\\sum_i \\log P(r_i|r_{<i},q,C;\\theta)]$. The resulting Control-R-32B scores 70.0% on AIME2024 and 93.2% on MATH500, which the paper reports as state-of-the-art among 32B models, and 61.1% on GPQA-Diamond and 55.0% on AIME2025 Part I. The ablation shows the control string is load-bearing: removing it collapses AIME2024 accuracy from 70.0% to 6.7% and MATH500 from 93.2% to 3.2%, while setting all scores to 0 or 5 gives intermediate results.","pith_inferences":["A natural extension the paper does not build is a difficulty-estimating router that chooses control scores automatically per query, turning the all-9 setting into one point on a cost-accuracy frontier rather than a fixed recommendation.","Because the control labels were produced by an automated annotator reading the same traces used for training, the eleven scores may reflect that annotator's stylistic preferences more than objective reasoning properties; an independent re-annotation study would clarify this.","The tree-search framing suggests a direct test the paper leaves implicit: token counts, branch counts, and backtracking events should move monotonically with stated search_depth and related scores, which would turn the accuracy-based proxy into a mechanism-level verification."],"forward_implications":["If the control fields truly modulate reasoning effort, users can trade accuracy against compute per query by adjusting the eleven scores rather than switching to a different model.","The control string is not decorative: removing it degrades accuracy dramatically, so deployment of Control-R-32B must always supply the field.","The conditional objective applies to any next-token language model, suggesting the same recipe could be transferred to other base models and other long-reasoning tasks.","Task sensitivity varies: AIME2024 moves substantially with control settings while MATH500 moves only slightly, so the optimal control configuration is task-dependent rather than universal.","Framing long CoT as tree search with execution-control and process-quality scores gives an interpretable vocabulary for reasoning behavior, which could be reused for analysis and monitoring."],"supporting_citations":[{"why":"Supplies long-CoT teacher traces used to build the training data and the forced <think> generation prompt used during evaluation.","marker":"(DeepSeek-AI et al., 2025)"},{"why":"SteerLM provides the attributed conditional generation idea that the Reasoning Control Fields directly adapt.","marker":"(Dong et al., 2023)"},{"why":"Provides the answer-grading tool used to judge correctness on the main benchmarks.","marker":"(Lightman et al., 2023a)"},{"why":"Defines the AIME2024 benchmark, one of the two central evaluation sets.","marker":"(Patel et al., 2024)"},{"why":"Source of the comparison numbers for the 32B baselines in the main results table.","marker":"(Bespoke_Labs, 2025)"},{"why":"Source of the AIME2025 Part I comparison results reported in the appendix.","marker":"(SRI_Lab, 2025)"},{"why":"Frames overthinking and test-time scaling, the problem this paper positions itself against.","marker":"(Muennighoff et al., 2025)"},{"why":"Provides the in-context tree search perspective the paper uses to define the eleven control fields.","marker":"(Rubin et al., 2021)"}],"fun_headline_variants":["Eleven control scores let a 32B model tune reasoning depth","Control-R dials thinking effort on 32B for AIME 70%","Tunable reasoning: 32B model hits 70% on AIME2024","Steer long CoT with control fields — Control-R-32B","Reasoning control fields give 32B models a dial for effort"],"cache_read_input_tokens":18816,"weakest_assumption_plain":"The paper assumes that accuracy differences under different control scores prove the model adjusts its reasoning effort, but it never measures reasoning length, step count, or strategy use, and the control labels were assigned by an automated annotator to the same traces used for training, making the supervision signal partly circular.","fun_headline_variants_meta":{"raw":{"variants":["Eleven control scores let a 32B model tune reasoning depth","Control-R dials thinking effort on 32B for AIME 70%","Tunable reasoning: 32B model hits 70% on AIME2024","Steer long CoT with control fields — Control-R-32B","Reasoning control fields give 32B models a dial for effort"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000161,"raw_usage":{"total_tokens":1238,"prompt_tokens":948,"completion_tokens":290,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":564,"completion_tokens_details":{"reasoning_tokens":190}},"tokens_in":564,"tokens_out":290,"duration_ms":3462,"temperature":1.0,"reasoning_tokens":190,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T12:10:09.279860+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run Control-R-32B on a fixed question set with all control scores set to 0 versus 9 and measure actual token counts, number of distinct strategies, and backtracking events in the sampled traces; if the all-9 condition does not produce longer or structurally deeper reasoning while improving accuracy, the central controllability claim is not supported.","supporting_citations":[],"review_version":1}