acceptodds
Under review as a conference paper at ICLR 2027

Recur, Switch, Recur, Commit: Function-Adaptive Hybrid Reasoning

Abstract

Large language models can reason through explicit tokens or recurrent latent states, but hybrid systems still lack a principled rule for deciding which representation to use at each step. We show that this choice is governed more by local reasoning function than by position: latent recurrence better supports search and planning by retaining alternative candidates, whereas explicit computation better supports calculation and verification by preserving exact values and entities. Controlled interventions reverse these preferences as candidate multiplicity and exact-value demand change, and frozen hidden states predict them before router training. We therefore introduce Function-Adaptive Hybrid Reasoning, combining capability-preserving multi-path training, function-grounded switching supervision, and budgeted router-only reinforcement learning for bidirectional, state-conditioned routing. Across eight model configurations and twelve core tasks, the full method improves the twelve-task macro over explicit CoT-RL on every backbone, by 0.8–2.2 percentage points. On Qwen3-8B, it gains 1.9 points while using 0.72 normalized compute and a 0.75 estimated normalized latency. These results suggest that effective hybrid reasoning is less about when to compress thought than about matching each local computation to the representation that best supports it.

Primary area:
interpretability and explainable AI
Keywords:
latent-reasoning+chaint-of-thought

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_6m7COpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper is about hybrid reasoning in LLMs. The authors use intermediate reasoning steps as explicit text tokens or latent hidden states and investigate which type of reasoning step works better with each representation and why. Experiments reported show that search and planning steps work better in latent representation and that calculation and verification steps work better as explicit text. They explain this by noting that latent states can keep several candidate paths at once and explicit tokens keep exact values more reliably. Based on these results, the authors propose Function-Adaptive Hybrid Reasoning. The training has three stages. First, they fine-tune the model to support both modes. Second, they train a router supervised by human labels of step types. Finally, they refine the router with RL under a compute budget. At inference time, the router chooses at each step whether to continue, switch mode, or give the final answer. The authors evaluate this method on 8 models and 12 benchmarks.

Critical strengths
  • The question the paper is addressing is clear and useful. Instead of just considering the moment of switching between latent and explicit representations, this paper asks which content and steps benefit more from either latent or explicit modes. This is a more meaningful approach for hybrid reasoning.
  • A good analysis with well-defined controls is provided. The authors compare the outcomes of different modes on reasoning steps. They support the main idea of the paper by showing that the type of step is more informative than its position. They also show that this preference flips when the number of candidates or exact values in the tasks are changed.
  • The evaluation is broad covering different models and benchmarks, and the proposed method improves over CoT-RL on all of the settings.
Critical weaknesses
  • Several works with similar ideas have not been cited and compared such as SWITCH (arXiv 2606.13106), LaTER (arXiv 2605.07315), and SwiReasoning (arXiv 2510.05069). The authors should explain the novelty of their work compared to these similar and recent works and add direct comparisons.
  • The function-based supervision as one of the main contributions of the work adds little to the final performance. The supervised router beats the no-label router by a small margin. Also, a simple MLP trained on hidden states, and without fucntion labels is performing at the same level.
  • Speed and efficiency claims are not measured and backed up with real wall-clock timings. The claims are based on the cost model and should be supported by real timing numbers at realistic settings. The authors’ own serving estimates show that mixing modes in a batch increases latency and lowers throughput.
  • The pipeline has many separate steps: latent-capability training, preference measurement, human annotation, control-token grounding, router training and RL. It also relies on up to 8,000 function labels from student annotators and a fixed set of 8 functions, three of which show no clear mode preference. This makes the method expensive to apply to new domains, especially for a small gain.
Questions for authors
  • How often does the router switch per problem on average, and what share of steps is latent vs explicit on different tasks? A few example routes showing where switches happen would help readers understand the learned policy and its impact in action.
  • Search and planning steps become latent and cannot be read by users. Have the authors considered decoding or inspecting these latent steps, and how do they see this trade-off for settings where reasoning traces need to be checked?
8 comments

6m7C · W1 — Relation to SWITCH LaTER and SwiReasoning

与 SWITCH LaTER 和 SwiReasoning 的关系

**Original concern **

Several works with similar ideas have not been cited and compared such as SWITCH (arXiv 2606.13106), LaTER (arXiv 2605.07315), and SwiReasoning (arXiv 2510.05069). The authors should explain the novelty of their work compared to these similar and recent works and add direct comparisons.

Thank you for pointing us to these closely related methods. SWITCH already learns repeated transitions, which makes your question especially valuable: what does Hybrid explain about choosing a boundary? We are pleased to share two findings that motivate our design. With four viable candidates, the effective candidate count was 2.74 in latent states and 1.31 in explicit states. Four steps after an exact value was established, explicit execution recovered it in 81.4% of cases and latent execution in 42.8%. This reversal motivates using the information required by the next operation to initialize routing. The comparison with native SWITCH below tests the practical value of that principle.

  1. We provide representational evidence and a theoretical account of why latent CoT helps during exploration and why combining it with explicit CoT is useful. A latent decision state can keep several viable approaches available. Writing an intermediate result can make its exact value easier to reuse later.
  2. We derive an empirically grounded, interpretable principle for choosing where to switch. The principle asks what the next operation needs to preserve. It gives us a reason for each proposed boundary and a way to test whether it places that boundary more accurately than a confidence rule.

1. We identify why latent CoT can outperform explicit CoT during search and planning

A latent step feeds a projection of the current hidden state into the next Transformer update. An explicit step feeds back the embedding of one selected token. Both retain a continuous cache. The distinction matters before a problem has been resolved. Committing to one written approach can make another viable approach harder to recover. With four verifiable solution paths, the effective candidate count at a decision state is 2.74 in latent mode, 1.31 in explicit mode and 1.08 after shuffling latent states (Fig. 3a, p. 4). Candidate information is denser in this diagnostic in a specific sense. More distinct continuations remain accessible from the same decision state.

We then tested whether that information actually helped solve the problem. When later evidence required a particular path, removing its aligned component reduced success by 25.1 points. Adding the same component back restored success to 71.2%, compared with 79.0% under sham (Fig. 7e–f, p. 9). The route-by-erasure interaction is −8.4 points (reported interval [−11.2, −5.7], App. I, p. 26), further connecting the damage to the choice of reasoning mode. Conversely, when the task required exact reuse of a numerical value four steps later, explicit CoT succeeded 81.4% of the time and latent CoT 42.8% (Fig. 3b). Together, these findings explain why alternating the two modes can help within a single solution. The model can defer commitment while several approaches remain useful, then write down the value on which subsequent steps depend.

2. The interventions and ablations locate useful switches where the information demand changes

Retrieval/search, planning/decomposition and hypothesis enumeration favor latent reasoning while the next step still depends on exploring alternatives. Exact calculation, entity binding and constraint checking give the model a reason to switch to explicit reasoning. Fig. 2d reports latent preferences for the first three operations and explicit preferences for calculation and verification. Summarization, comparison and revision vary by instance. This also explains why a calculation near the start can favor explicit CoT and a new search near the end can favor latent CoT. The order in which those operations happen can change within the same problem (§7.4).

Two observations connect this principle to the representation. First, neighboring pre-decision states agree on the sign of the mode advantage 79% of the time, versus 51% after label permutation (§4.2). Second, increasing the number of viable candidates from 1 to 8 moves latent-minus-explicit quality from +0.1 to +3.5 points, while increasing the number of exact fields from 0 to 8 moves it from +2.4 to −3.0 (Fig. 3c–d). On calculation steps, forcing the router’s chosen action at the same state scores 10.6 points above forcing the opposite action (§7.6, Fig. 7b). These results motivate the local model ΔQ≈αB−βDexact\Delta Q\approx\alpha B-\beta D_{\rm exact}, where BB is the demand to retain viable candidates and DexactD_{\rm exact} the demand for exact reuse. The useful boundary is the point at which the relative value of these demands changes enough to justify switching.

3. Function-informed initialization gives each proposed boundary a testable rationale

We use the observed preferences to initialize the router. On one frozen capability checkpoint, we force latent and explicit continuations to estimate the relative benefit of each mode for each function. Stage 2 combines those estimates with the segment’s functional labels, downweights uncertain differences and accounts for the cost of switching. At inference the router reads the current state, mode and remaining budget (§5.2, Eq. 2). This gives the initialization a concrete interpretation. A latent span is proposed because keeping an alternative available is expected to improve the continuation.

The published rules give uncertainty different roles in choosing a reasoning mode. SWITCH uses high next-token Shannon entropy to label contiguous spans of written CoT during initial training. At inference the model emits learned boundary markers. SwiReasoning uses a relative signal at inference. Once the current mode has lasted long enough, entropy rising above the block reference can trigger explicit → latent, and entropy falling below it can trigger latent → explicit. ThinkRouter takes a different action when confidence is low. If the maximum next-token probability falls below the threshold, it samples a discrete token to avoid mixing incompatible continuations. At or above the threshold it feeds back a probability-weighted embedding. SwiReasoning uses a soft embedding to keep alternatives open during exploration. ThinkRouter points out that mixing mutually exclusive moves can introduce noise. Both concerns can arise at different points in one problem. These choices give us reason to question whether a single direction of uncertainty can reliably tell us when latent CoT will help. Entropy measures how dispersed the next-token distribution is, its change is relative to a block's history, and maximum probability measures only its peak. Each leaves open a consequential question about the alternatives receiving that probability mass. Some should remain available until later evidence arrives. Others prescribe incompatible actions for a decision that the evidence already settles. Thank you for prompting us to make that distinction concrete.

sample

One broadcast archive lets us vary the reasoning demand while keeping the facts fixed. An archivist must identify the cut of a documentary that aired at 20:00. The playout log names reel R17, which the catalogue assigns to the Harbor cut. An independent recording opens with a bell and ends with rain. Moor shares the bell and Vale shares the rain. Only Harbor has both. The same file contains three signed notices. The web copy WEB-71 was cleared at 19:30 and placed on hold at 19:45. The archive master ARC-17 was cleared at 19:50. For each copy, its own latest signed notice governs.

Table W2.1. Two questions about the same archive, at higher and lower pre-action vocabulary entropy

Pre-action vocabulary entropyKeep two valid ways to identify the cut. Predicted L / E successCarry forward one binding decision. Predicted L / E success
High, if observedThe log and the recording offer equally plausible starting points. Which cut aired, and can the model still justify it if the editor later accepts only one source? Either the reel lookup or the two sound cues establish Harbor. 74% / 71%The web clearance, the later web hold and the master clearance are interleaved. Should the model clear or hold the web copy? Both actions have textual support, but only WEB-71 · HOLD · 19:45 satisfies the rule. 70% / 73%
Low, if observedThe playout entry sits beside the reel index, making a reel lookup an obvious first move. Can the model still identify Harbor from the bell and the rain if the editor later requests audio evidence alone? 78% / 76%The two web notices now appear together in time order. HOLD is easy to anticipate. Four catalogue entries later the model must still give WEB-71 · HOLD · 19:45. Substituting the master's ID or time fails. 75% / 78%

L / E denotes success under a forced latent / explicit reasoning window. The percentages are illustrative targets for held-out groups of source files. Each example describes a candidate test item. Its entropy stratum and the better mode both await measurement. Example questions and answer keys

The low-entropy candidate case is particularly informative. A model can be confident about a substantive first move, looking up R17, while having good reason to keep the acoustic proof available. Conversely, in the high-entropy exact case, hesitation between two well-supported words cannot give two conflicting web decisions equal authority. A scalar uncertainty rule must assign the same action probabilities at identical signal values. If latent mode gains a>0a>0 points on a fraction qq of such states and explicit mode gains b>0b>0 on the remainder, its smallest expected loss against a policy that recognizes the demand is min⁡(qa,(1−q)b)\min(qa,(1-q)b).

The submitted comparisons show a routing gain and identify the tasks that benefit

<a id="w2-diagnostic-en"></a>

Table W2.2. Hybrid and routing baselines. Reported in the submission, Qwen3-8B.

MethodRouting quality (%)Hybrid-Full minus row (pp)
Entropy64.5+4.8
Confidence65.2+4.1
Router without function labels66.3+3.0
TARPO*67.5+1.8
ThinkRouter*65.2+4.1
SLT*65.5+3.8
System-1.5*66.5+2.8
Hybrid-SFT67.1+2.2
Hybrid-Full69.30.0

On this routing diagnostic, Hybrid-Full scores 69.3%, which is 1.8 points above TARPO*, the strongest named adaptation, and 4.8 points above the entropy rule. Estimated compute is 0.70 for Hybrid-Full and 0.76 for TARPO*. We agree that the asterisks in D.6 gave readers too little information to judge implementation fidelity.

Table W2.3. Two pure-mode baselines and Hybrid before and after router RL on every core task. Reported, Qwen3-8B.

Core taskExplicit CoT-RLPure latentHybrid-SFT (pre-RL)Hybrid-Full (post-RL)Post − pre (pp)
GSM8K93.587.193.293.8+0.6
MATH-50085.877.688.190.7+2.6
AIME 202562.243.362.263.3+1.1
AIME 202655.638.955.655.60.0
HumanEval+76.468.177.077.2+0.2
MBPP+79.972.381.282.8+1.6
LiveCodeBench v651.240.651.053.2+2.2
GPQA Diamond54.549.056.658.6+2.0
MMLU-Pro66.661.966.767.0+0.3
BBEH31.530.034.236.1+1.9
IFEval76.873.676.576.0−0.5
MuSiQue46.941.347.650.1+2.5
Equal-task macro65.157.065.867.0+1.2

App. D.7 (p. 18) reports these scores in the three-seed Qwen3-8B setting. Scores are percentages of each task’s official metric, including exact match on MuSiQue and a within-task balanced macro on BBEH. The bottom row equally weights the 12 displayed task scores and is shown to one decimal. These point differences do not supply paired significance tests.

Reviewer_ZDKiOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

The paper considers the question of whether a language model should output its reasoning and at what point it should continue to reason internally without producing any words. The authors suggest that internal reasoning allows the model to keep track of multiple possible ways of solving a problem, while producing written reasoning enables it to reuse specific numbers and other details later on. In order to test this explanation, they alter the information that a problem requires the model to keep track of, and then train a controller to decide when to reason internally, when to write a step, and when to provide the answer. The study includes eight different models and twelve benchmarks; for each model, the approach results in an improvement of between 0.8 and 2.2 percentage points over a model that has been trained to use written reasoning with reinforcement learning.

Critical strengths
  1. The experiments investigate whether internal reasoning is more helpful and when writing out the steps is more helpful. The authors find that internal reasoning is more useful when a problem requires keeping a greater number of possible solutions available, while written reasoning is more useful when more exact values have to be remembered and reused (Section 4.3, lines 195–215; Figure 7c–d). It is a nice and novel finding.

  2. The method has been assessed over a number of models and on a wide variety of tasks. For all eight model configurations, the method results in an improvement in the average scores on these benchmarks, and the appendix contains the results from multiple training runs (Table 1).

Critical weaknesses
  1. The paper gives the results of its main experiments without stating how to carry them out. This introduces a lot of readability issues.

    First, the authors design problems that involve considering a greater number of possible paths or recalling more exact values. Yet they do not provide the problem templates or the rules according to which they were constructed (Section 4.3, lines 212–215). As a result, readers are unable to determine whether increasing the number of possible paths keeps the need to remember exact values the same, or whether increasing the demand for exact values has the same effect on the number of possible paths. This is important since the conclusion rests on separating out the effects of these two requirements.

    Secondly, the authors somehow remove information about one possible path from the model's internal state and then check the accuracy difference. However, the paper does not explain how they identify the information associated with that path, which layer they alter, or how much they modify it.

    In order to assess the paper's explanation, the first experiment should indicate what changes were made to the problems, and the second experiment should show what changes took place within the model. With these strong readability issues, I am more aligned with rejecting the paper at this stage.

  2. The paper does not indicate what advantages it has over previous switching methods. SWITCH (Yang et al., 2026) has already learned when to switch between internal and written reasoning. In order to determine what improvements result from these training decisions a comparison with SWITCH is required. The comparisons with TARPO, ThinkRouter, SLT and System-1.5 also need to be made clearer: Appendix D.6 states that those methods were modified, but does not explain the changes made. Hence the readers are unable to assess how closely the versions that have been tested resemble the original methods.

Questions for authors
  1. Could you provide some examples of the test problems? Also, the text should be made more readable by clarifying the way in which you increase the number of possible solution paths.

  2. What modifications do you make in the model as part of the path-removal experiment? Could you please explain how you determine the information linked to one solution path and how you remove it, including which layer you alter and the extent of the changes?

  3. What is the difference between your method and SWITCH? You should make the comparison under the same training and evaluation conditions. Moreover, please explain the modifications you made to the other baseline methods.

  4. Can you give us the complete details regarding the training and evaluation settings? Please state the datasets used, the training schedule, the learning rates, the computation limits, and the evaluation settings required in order to reproduce the results (Section 5.2).

11 comments

Response to Reviewer ZDKi /

We sincerely thank the reviewer for the careful and constructive feedback. We are greatly encouraged by the reviewer’s recognition of the key finding from our controlled experiments—that the relative advantages of latent and explicit reasoning reverse as the demands for candidate preservation and exact-value reuse change. We address each concern in turn below.

ZDKi · W1 — Task construction and candidate intervention

W1 (1/2). How we construct and verify multiple paths

Original concern

First, the authors design problems that involve considering a greater number of possible paths or recalling more exact values. Yet they do not provide the problem templates or the rules according to which they were constructed (Section 4.3, lines 212–215). As a result, readers are unable to determine whether increasing the number of possible paths keeps the need to remember exact values the same, or whether increasing the demand for exact values has the same effect on the number of possible paths.

Thank you for asking whether the two axes of Fig. 3 can actually be varied independently. That is the condition on which our interpretation rests. We are pleased to show exactly how a problem branches and what must stay correct when the branches rejoin.

candidate count K and exact-value count D. In §4.3 and Fig. 3c–d (p. 4), K is the candidate count and the exact-value count DcountD_{\mathrm{count}} gives how many exact values need to be retained or reused. The plotted levels are K = 1, 2, 4, 8 and D = 0, 2, 4, 8. In our checkable template, K is the number of admissible candidates at the specified point before the later cue. D counts required exact fields, such as carton count and charge, so that example has D = 2.

Construction rules. Each item has its own given facts, task goal and exact reference answer. We can resample quantities, rates or side lengths when generating a variant, and we verify the paths and answer using that variant's inputs.The specific construction method is as follows:

  1. Generate legal orders. We list each operation, its prerequisites and the information it reads or changes. Two operations can swap when both prerequisites are satisfied and their execution order preserves the state used by later steps. We enumerate schedules that respect the dependencies. For example, two checks can use the same verified input and store their results independently before a later operation needs both.
  2. Generate different solutions. For a given item, we construct complete derivations that use different representations or methods and reach its reference answer. Each derivation is checked against the source facts and the exact reference answer. We record operation orders and mathematical methods as distinct units, so wording changes and reordered steps do not inflate the number of methods.
  3. Verify each path. We check prerequisites at each step and require every accepted trace to reach the same verified task condition and exact answer. Integers, exact fractions or integer monetary units carry the calculations. Any requested rounding is applied to the reference value at the end.

Question template. “Given [this item's facts and numerical inputs], complete [task goal]. Follow [prerequisites and allowed or not allowed methods]. Return [named values] at [requested precision].” We record the source ID, rules, trace, intermediate checks and reference values. In our controlled comparisons of path count, we hold the item fixed and change a process rule or add a verified method. Using the above template, we can make the model generate different rollouts under constraints for the same question. We use the following method to identify and compare rollouts that are equivalent before key nodes as multi paths.

When paths are equivalent at a key node. Our method represents tool histories using acquired information atoms and auxiliary operations. Its information sets allow different orders to meet at the same main node. Its projection to success-reachable anchors can also align a state that has acquired a superset of the anchor's information. We test a specific downstream operation. Both paths must establish its verified prerequisites and match relevant values and side effects. We then replay the same continuation at a common remaining budget. This also lets us compare different methods, whose intermediate steps can differ.

Example 1. One shipment workflow opens a second branch

A school needs 288 textbooks. The warehouse has 360 and keeps 48 in reserve. Each carton holds 12 and weighs 9 kg when packed. The truck fits 26 cartons and 240 kg. The carrier charges 120 yuan plus 18 yuan per carton, with a 600-yuan budget. The 08:40 departure, 75-minute journey and 10-minute unload must finish by 10:15. Decide whether to release the shipment and return the exact carton count and charge.

  • Shared prefix. Confirm 360 − 48 = 312 available books, then pack 288 ÷ 12 = 24 cartons.
  • First path. Check the charge (552 yuan) and then the load (216 kg). Allowing the two checks to swap opens a second path that checks 216 kg first and then 552 yuan.
  • Shared completion. The routes meet after both checks. The independent timing check gives 10:05. With all checks passed, generate the manifest and release 24 cartons at 552 yuan.

Holding the timing check after the two branches makes the new fork explicit. The timing check can also move, with five positions among the four other steps before the manifest. Requiring “check charge before load” gives 5 orders. Allowing either order gives 10, all with the same exact result. Requiring the two additional exact fields, available stock and weight, keeps those counts at 5 and 10.

Example 2. Different mathematical methods with the same inputs

For a right triangle with legs 6 and 8, area decomposition gives radius 2. The area is 24, the hypotenuse is 10, and the three triangles formed by the incenter have combined area r(6+8+10)/2=12rr(6+8+10)/2=12r. A coordinate derivation instead puts the circle center at (r,r)(r,r) and the hypotenuse on 3x+4y=243x+4y=24. Tangency gives (24−7r)/5=r(24-7r)/5=r, so this route also gives 2. Both methods can report the radius, hypotenuse and area exactly.

Requesting two or four decimal places changes the output format. Requiring the hypotenuse and semiperimeter to be computed first and then using r=A/sr=A/s admits only the area method among the two shown. This is a process constraint at the same numerical precision, just like requiring the charge check first.

What the available data show. The following are path enumerations and exact reference checks.

MaterialChange the precision demand with rules fixedChange the path options with exact fields fixed
Seven-step shipment2 → 4 exact fields. The charge-first rule allows 5 → 5 orders, and the flexible rule allows 10 → 10All 5/5 and 10/10 orders return the exact reference values, including 24 cartons and 552 yuan
Incircle1 → 3 exact fields. 2 → 2 verified methodsBoth methods give radius 2, and both give 10 and 24 when those fields are requested
Controlled generator E01, 32 sourcesFrom 0 to 2, 4 or 8 fields, 384/384 per level preserve the count and selected route. Per-value digit changes pass 768/768 comparisonsFrom 1 to 2, 4 or 8 candidates, 288/288 per level preserve the exact requested answers, 4,032/4,032 paired fields
Earlier controlled generator, 20 sourcesFrom 0 to 2, 4 or 8 fields, 240/240 per level preserve the count and route. Digit changes pass 480/480180/180 pairs per higher K preserve the exact answers, 2,520/2,520 paired fields
Earlier worked examples E02 and E031 → 2 reported fields in E02 and 1 → 3 in E03. 2 → 2 orders or methods, respectivelyBoth E02 orders give 2,449.00 yuan. Both E03 methods give 335.00 yuan

Thus, the verified traces have the same numerical accuracy by construction. For model outputs, we will measure the change in exact-field accuracy on same-source pairs with the requested values and precision fixed, and will compare route accuracy at fixed K as reporting demand changes.

Our actions

We will add the construction template and both worked examples beside §4.3, with separate columns for legal schedules, distinct methods, required fields and decimal precision. Thank you for your suggestions, which have improved the readability of our paper.

W1 (2/2). How we identify a path-specific component and test its causal use

Original concern

Secondly, the authors somehow remove information about one possible path from the model's internal state and then check the accuracy difference. However, the paper does not explain how they identify the information associated with that path, which layer they alter, or how much they modify it.

We appreciate the distinction you draw between changing a problem and changing information inside the model. Your question leads to a sharper identity test. How and can we extract a coordinate-method intermediate from independent coordinate derivations, weaken its use in a recipient and strengthen it again while the area proof stays available? We describe that test and the observations that would distinguish residual information from a later rederivation.

1. How we identify a path-specific component, intervene and measure the change

Thank you for asking how we identify information specific to one reasoning path. Our proposed test has three steps.

  1. Identify the path information. We choose a verifiable intermediate result that distinguishes the target path from other valid paths. On a separate set of problems, we use verified target-path traces to learn which directions in the hidden state predict that result. We freeze those directions. For each test problem, we then extract the component along those directions from an independent run that follows the target path, after subtracting a matched reference state. We call this vector bb.
  2. Suppress or strengthen it. At the same reasoning step, we subtract one copy of bb from a state in which both paths’ intermediate results can be read out. In a separate run progressing along an alternative path, we add one copy of the same bb. For the proposed Qwen3-8B test, both edits act on the output of Transformer block 18. We record the size of each edit relative to the state being edited.
  3. Check what changed. We immediately test whether the target intermediate result is still readable, then let the model continue. A selective effect would mean that subtraction makes the target path harder to complete, addition makes the model more likely to produce a valid target-path solution, and the alternative path remains usable when requested. We compare both edits with random edits of the same size.
2. Subtraction and enhancement statistics and a test of component identity

In this chapter, we verify that the vectors we extract are indeed information related to the target path.

Subtraction and delayed restoration · planning values. The final-answer column uses the B-request copy. All comparisons within a delay share the proof budget.

Wait before proofArmValid B proofValid A proofCorrect final answer
0 updatesAB sham80.0%84.0%95.0%
0 updatesEqual-norm random78.0%84.0%95.0%
0 updatesUnrelated component78.5%84.0%95.0%
0 updatesSubtract B donor72.0%84.0%94.8%
0 updatesZero the fitted projection71.5%84.0%94.8%
0 updatesA-only reference72.0%84.0%95.0%
1 updateSham, then sham80.0%84.0%95.0%
1 updateSubtract B, then sham72.0%84.0%94.8%
1 updateSubtract B, then add it back77.0%84.0%95.0%

B loss is 8 pp versus sham, 6 pp beyond random and 6.5 pp beyond the unrelated edit. A is stable.

Enhancement from an A-progress state · planning values. Free B choice and valid free B proof both use all free responses. Requested B is scored on a separate copy.

ArmFree B choiceValid free B proofRequested B proofFree final answer correct
A-progress sham28.0%24.0%72.0%95.0%
Equal-norm random addition29.0%24.0%73.0%95.0%
Unrelated donor30.0%25.0%73.0%94.9%
Add the same B donor40.0%35.0%79.0%95.0%

This pattern gives +12 pp in choice, +11 pp in valid free B proofs versus sham, and +7 pp in requested B proofs. The valid-free gain is +11 pp over random and +10 pp over the unrelated donor.

The above experiment shows that the vectors we extracted are indeed related to the target path and play a key role in the calculation of the target path.

3. A sample:One triangle under subtraction, restoration and enhancement
Reviewer_mZruOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes Function-Adaptive Hybrid Reasoning, which switches between latent and explicit computation according to local reasoning demands. It combines capability-preserving training, function-based routing supervision, and router-only reinforcement learning, motivated by observations that latent reasoning favors candidate exploration while explicit reasoning better preserves precise information.

Critical strengths
  1. The paper studies a meaningful question: when each reasoning representation is useful.

  2. The evaluation covers multiple backbones and tasks, with paired interventions and erasure/add-back experiments providing useful evidence of complementary capabilities.

Critical weaknesses
  1. Important details remain unclear, including the training data and teacher, curriculum schedules, loss weights, estimation of function-level preference probabilities, and budget-constrained RL implementation.

  2. Recent adaptive-routing baselines are evaluated as controlled adaptations, but their modifications and tuning are insufficiently explained. Their comparative performance on the full core suite is also not established.

  3. Compute and latency improvements rely on estimates without sufficiently specified cost coefficients or hardware calibration, leaving practical inference benefits uncertain.

  4. Limited demonstrated novelty. The framework largely combines established latent reasoning, adaptive routing, distillation, and RL components. Its distinctive function-based supervision does not show a clear advantage over direct state-based supervision: Appendix D.3 reports identical downstream accuracy for the function rule and frozen-state MLP, with slightly better predictive metrics for the MLP.

Questions for authors

Please see weaknesses.

4 comments

Response to Reviewer mZru Verified official review · Author-visible forum checked on 19 November 2026. · Summary letter

Thank you for asking us to connect novelty to verifiable implementation and comparative evidence. We separate the information-demand hypothesis, the supervision design and its demonstrated benefit. Your concern about the D.3 tie is retained explicitly, rather than treated as resolved by an architectural description.

mZru · W1 — An executable training and evaluation specification Original concern Important details remain unclear, including the training data and teacher, curriculum schedules, loss weights, estimation of function-level preference probabilities, and budget-constrained RL implementation.

Core concern. What exact data, teachers, curriculum, losses, preference estimator and budget-constrained RL updates are used?

Response to the reviewer

Thank you for identifying the five choices another group would need to train Hybrid. Your question connects the preference estimate to the update that actually receives its gradient. Section 5.2 gives the three objectives and the RL freeze policy. The three-seed replay ablation provides a practical clue: removing explicit replay reduced forced-explicit quality from 67.0% to 59.3%, and removing the latent loss/replay reduced forced-latent quality from 60.9% to 51.4% (App. D.4, p. 16). We give an independent recipe that preserves both paths and works through the preference and budget calculations. The archived supplement contains DEEPEN. We apologize for that packaging error. This independent recipe does not by itself identify the original run configuration.

Data and teachers. The independent recipe uses 20,000 training and 2,000 development source problems. Training quotas are GSM8K 3,000, NuminaMath-CoT 4,000, CodeContests 4,000, HotpotQA 3,000, OpenBookQA 2,000, RLVR-IFeval 2,000 and new multipath families 2,000. Each development quota is one tenth as large. We split by source/family, remove overlap with the 12 test sets and retain failures in the coverage report. The Stage 1 teacher is frozen explicit Qwen3-8B (revision b968826d). We keep the first verified trace in at most two attempts per source. The Stage 2 teacher is the frozen Stage 1 model following the same route with natural-language controls. Dataset revisions and splits Training stage What learns and from which losses Schedule per seed 1 · Preserve both modes Latent projection and rank-16 attention LoRA. Text cross-entropy / endpoint distillation (temperature 2) / explicit replay weights 1 / 0.5 / 1. AdamW 1e−4, batch 64, ≤313 updates. Updates 1–104 / 105–208 / 209–313 use 10 / 25 / 50% latent coverage on mixed traces, at most 2 / 4 / 8 consecutive latent steps. 2 · Ground legal actions Independent two-layer 256-wide GELU router at 3e−4 and new control embeddings at 5e−5. Freeze content. Text / endpoint / control / route / pure-mode replay weights 1 / 0.5 / 1 / 1 / 0.5, each averaged over its supervised elements. 2,000 training sources × 4 ordered segments, batch 128, ≤3 epochs or 189 updates. 3 · Refine the route Freeze backbone, vocabulary head, projection and control embeddings. Optimize only the router, with a frozen Stage 2 reference. AdamW 1e−5, 256 prompts × 4 rollouts, one pass per batch, ≤100 updates or 102,400 rollouts per seed.

Estimating a preference. From the same frozen prefix and cache we force each mode for the same segment under a shared budget and continuation policy. Two paired draws per mode yield at most 32,000 continuations. Each draw gives latent 1 for a verifier win, ½ for a tie, 0 for a loss. Average the two credits to obtain uᵢ. In the other four of five source folds, each source contributes total segment weight one. With four segments per source and annotator probability qᵢ(f), estimate p_f(L) = [1 + Σ qᵢ(f)uᵢ/4] / [2 + Σ qᵢ(f)/4]. The held-out target is Σ_f qᵢ(f)p_f(L) (Eq. 2). For 60 equally weighted single-function sources with 36 wins, 12 ties and 12 losses, this gives 43/62 = 0.694. We shrink the training weight for weak or rare preferences, retain ambiguous targets and charge 0.05 log-loss units per switch when smoothing legal routes. Weighting and uncertainty

Enforcing the budget. Set B to the matched explicit CoT-RL development mean active GPU time and c=C/B. For each rollout use r=verifier−0.005×mode changes−λ(c−1). Center and scale r within its four-rollout prompt group. Clip the selected legal action's probability ratio to [0.8,1.2] in the clipped surrogate and use route-only KL weight 0.02. After each batch set λ←clip(λ+0.01(mean c−1),0,1). Starting at 0.03, a 20% overrun gives 0.032. A failed answer scores zero and retains its cost. We check development quality and realized cost every five updates. This recipe uses the independent head in Table 2 (69.3% diagnostic). Its vocabulary-tag variant scored 69.0%. We log each loss term's gradient support. A pure replay example that touches only frozen tensors measures retention without updating the router.

Your question gives readers a continuous path from one verified training example to one optimizer update. Schedule and caps **Evidence navigation

mZru · W2 — Native baselines and controlled adaptations Original concern Recent adaptive-routing baselines are evaluated as controlled adaptations, but their modifications and tuning are insufficiently explained. Their comparative performance on the full core suite is also not established.

Core concern. Which baseline components and tuning settings are retained, and how do native systems compare on the full core suite?

Response to the reviewer

Thank you for asking how we adapted each baseline and how far the comparative evidence extends. We checked the two relevant tables separately. On the Qwen3-8B routing diagnostic, adapted TARPO scored 67.5% and Hybrid 69.3%. On the full 12-task core suite, recomputing the displayed task scores gives 67.0% for Hybrid and 65.1% for explicit CoT-RL. These numbers answer different questions. We give the adaptation details and the exact full-suite test needed to establish performance against native SWITCH and the other adaptive systems. Submitted population Comparator scores (%) Hybrid score (%) Routing diagnostic, App. D.6, pp. 16–17 Entropy 64.5, confidence 65.2, no-label router 66.3, Hybrid-SFT 67.1 69.3 Same diagnostic, controlled adaptations TARPO* 67.5, ThinkRouter* 65.2, SLT* 65.5, System-1.5* 66.5 69.3 Full core, App. D.7, pp. 18, 21 Original 62.3, explicit SFT 63.6, explicit RL 65.1, pure latent 57.0, CODI* 60.2, Hybrid-SFT 65.8 67.0

The core row is an equal-task mean recomputed from the 12 printed task scores, rounded to one decimal. The submission separately reports the three-seed macro for Full and CoT-RL. The four starred routing gaps are +1.8, +4.1, +3.8 and +2.8 points. App. A identifies the shared evaluation conditions and distinguishes TARPO's native joint update from router-only training. The starred scores concern controlled adaptations and therefore cannot be attributed to the complete native systems. Task scores · Reported comparisons Native system Operation that must survive the comparison SWITCH High-entropy training spans initialize repeated entry and exit. Its curriculum and Switch-GRPO use hidden-state recurrence. Paper TARPO A learned hard/soft meta-action and its content branches are trained jointly with the backbone. Paper ThinkRouter Low maximum token probability can select a discrete token. Confident steps can use a sparse soft embedding. Retain Cold Stop. Paper SLT A drafter proposes text locally. A gate accepts a prefix for compression. Charge the drafter, compressor and fallback. Paper System-1.5 Retain both depth and reasoning-step shortcuts and their individual costs. Paper LaTER / SwiReasoning LaTER learns a latent intuition prefix with an exit to explicit derivation. SwiReasoning can switch repeatedly using relative block entropy and directly feeds back a weighted embedding. LaTER · SwiReasoning

Why the proposed boundary could help. With four candidate routes, the effective candidate count was 2.74 in latent states and 1.31 in explicit states. Four steps later, exact-value reuse was 81.4% in explicit and 42.8% in latent mode (§4.3, p. 4). This predicts an advantage when an episode alternates between keeping alternatives available and reusing a precise value. Repeated routing scored 69.3% versus 67.5% for the stronger single-switch policy at the same estimated compute of 0.70 (Table 2). These observations motivate a comparison at equal pre-action confidence, with the executor and stopping fixed.

Full-suite comparison protocol. The comparison uses identical source IDs and scorers on 12 tasks, matched access to data and teachers, up to four development configurations per method including its native default, and a common tuning cap. We freeze the selection before testing five operating budgets. At matched measured GPU time (within 5%), the primary decision criterion over native SWITCH is ≥2 points, with an adjusted paired lower bound above zero. 65.0% versus 67.1% illustrates a passing point estimate. At the secondary common-quality target of 66.0%, the time-ratio criteria are ≤0.90 against SWITCH and ≤0.95 against each other system, conditional on overlapping attainable quality.

Thank you for asking us to show precisely which opponents and tasks support a comparative claim. Native tuning and decision rules **Evidence navigation <a id="h-m-w3"></a> mZru · W3 — Hardware calibration and end-to-end service time Original concern Compute and latency improvements rely on estimates without sufficiently specified cost coefficients or hardware calibration, leaving practical inference benefits uncertain.

Core concern. How are operation costs calibrated on hardware, and do they translate into measured end-to-end gains?

Response to the reviewer

Thank you for asking how we assign a cost to a latent step and how that translates into elapsed time. Calibration is central here. Table 1 reports 67.0% for Hybrid and 65.1% for explicit CoT-RL with normalized estimates of 0.72 compute and 0.75 latency. App. C illustrates why a device trace matters: its modeled change from homogeneous to mixed batching takes p95 from 1.62 to 1.96 and throughput from 5.62 to 4.92, in normalized units. We welcome the opportunity to specify a measurement that can identify the cost coefficients and test user-visible latency directly.

Calibrating Eq. 4. The calibration protocol uses one NVIDIA B200 in BF16, tensor parallelism 1, with identical Qwen3-8B tokenizer, backend build, scheduler, KV policy and output caps. In separate development traces we vary cache length (256/1,024/4,096/8,192 positions) and batch size (1/8/16). We time prefill, the decode backbone at the actual cache length, vocabulary projection and sampling, latent projection, route decisions and control boundaries. A latent step still incurs a backbone forward. We log calls to the vocabulary head even in latent mode if the implementation uses it for routing. Within each shape, matched operator toggles and kernel traces identify incremental costs in milliseconds per executed batch event. We report inseparable fused terms together. On untouched mixed batches we compare predicted with measured busy time. The calibration criteria are median absolute percentage error ≤10% and 90th percentile ≤20%. If they fail, practical comparisons use direct timings and we report where the cost model failed. Coefficient and validation protocol

Measurement definitions for the public suite and retail. The public workload uses all 12 core tasks. The second replays held-out real retail customer-service conversations with their tool environment. App. G (p. 24) reports success of 68.6% for explicit CoT-RL, 69.2% for Hybrid-SFT and 68.9% for Hybrid-Full. These are offline outcomes. For timing we replay matched arrivals at 60% of common reference capacity in ten randomized paired blocks per arm, each with 100 warm-up and 1,000 timed requests. We include retries, failures, queueing and tool waits. Capacity has a separate arrival-rate sweep at a fixed SLA. Conditional example on one B200, explicit CoT-RL → Hybrid-Full Public core mix Held-out retail replay Mean busy GPU time per request 1,000 → 900 ms 500 → 450 ms Mean arrival-to-final-response time at common load 4.00 → 3.80 s 2.50 → 2.40 s p95 arrival-to-final-response time at common load 5.00 → 5.10 s 4.00 → 4.08 s Terminal responses/s at sustainable capacity 0.94 → 1.04 1.80 → 2.00 Correct responses within SLA/s at that capacity 0.60 → 0.66 1.15 → 1.27

The table illustrates the target if the assumed explicit reference times are observed. In the example, GPU work falls 10% in both workloads while p95 rises 2%. A shorter average would therefore support a different claim from a shorter tail. Busy time is the union of device-active intervals divided by all timed requests. End-to-end time runs from arrival through terminal status. We report timeout and error rates alongside raw capacity and correct on-time goodput.

We require a GPU-time ratio ≤0.90 with paired upper bound <1, a quality-difference lower bound >−1 point, and a p95 ratio upper bound ≤1.05 at common load. A capacity claim also requires ≥10% more correct on-time responses/s with a lower ratio bound >1.

Thank you for identifying what a credible conversion from forward passes to seconds requires. **Evidence navigation <a id="h-m-w4"></a> mZru · W4 — Mechanistic contribution and the role of functional supervision Original concern Limited demonstrated novelty. The framework largely combines established latent reasoning, adaptive routing, distillation, and RL components. Its distinctive function-based supervision does not show a clear advantage over direct state-based supervision: Appendix D.3 reports identical downstream accuracy for the function rule and frozen-state MLP, with slightly better predictive metrics for the MLP.

Core concern. What is the distinct scientific contribution, and does functional supervision improve over an equally supervised state MLP?

Response to the reviewer

Thank you for pressing us to identify the scientific contribution and for pointing out the MLP result in App. D.3. We are pleased to share the observation from which the method was designed. In Fig. 3, increasing candidate count moves latent minus explicit accuracy from +0.1 to +3.5 points. Increasing exact-value demand moves the difference from +2.4 to −3.0 points. This reversal suggests a testable distinction between keeping alternatives available and making a value easy to reuse. The MLP and functional rule tie at 67.1% downstream. Whether the functional signal also improves a learned controller therefore deserves its own equally supervised comparison.

Tables R1–R4 separate the information-demand finding from the incremental value of functional supervision. The primary Qwen3-8B decision thresholds are interaction +4 percentage points and gain +1 point over the equally supervised MLP. The manuscript’s D.3 tie is retained. The worked comparison uses candidate-mode difference +3 points, exact-reuse difference −3 points, and same-label MLP/structured diagnostic scores 67.1/68.1%. The settings in §8 specify the comparison; Table R12a defines separate cross-model tests. Stochastic latent-content sampling is outside the present system.

The submitted Fig. 3 (p. 4) provides the motivating observation: the latent-minus-explicit difference grows from +0.1 to +3.5 points as the number of candidates increases, whereas increasing exact-value demand changes it from +2.4 to −3.0 points. This suggests that uncertainty alone may not identify the useful mode. Similar probability dispersion can arise from several viable routes, incompatible operations, uncertain wording or an exact commitment that will be needed later. Our hypothesis is that the task value of the information being preserved matters, not only its uncertainty.

This is a conditional explanation. Both modes retain continuous caches, so explicit execution does not necessarily erase alternatives, and latent execution does not automatically preserve them. Under an explicitly defined semantic partition, the entropy chain rule distinguishes uncertainty between operations from variation within an operation; ordinary next-token entropy generally does not identify that partition. The evaluation tests the explanation with confidence-balanced paired continuations, independently manipulated candidate/exact-value demands, and selective interventions with random, unrelated and restoration controls. The analysis retains neutral and uncertain cases.

You are also correct about Appendix D.3 (p. 15): the rule and MLP have identical accuracy of 67.1 and estimated cost of 0.69, and the MLP has slightly better predictive metrics. This result does not establish superiority. The comparison includes function-based supervision with direct paired-utility supervision, full-state prediction and an MLP receiving the same auxiliary labels. Resource accounting includes label, teacher and rollout costs, and policy evaluation holds the capability model and stopping rule fixed. This tests whether functional structure improves learning or control rather than merely yielding an alternative explanation.

The existing deterministic latent update is also compatible with RL through stochastic routing or output actions. A trainable latent-content sampler is a distinct mechanism for within-mode exploration and is not part of the present system or novelty claim. Our current contribution is tied to the information-demand analysis and the evidence for functional supervision.

E02–E05 add an important falsification condition: order diversity, valid strategy diversity and backup-route execution are different measurements. A function-based account becomes stronger only when it predicts and controls the latter behaviors beyond confidence, importance and a same-label state model. The low-rank sampler is a distinct mechanism with explicit stochastic-policy precedents and is outside the submitted deterministic system.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.