Machine Learning · arXiv:2610.35801 v2 · submitted 26 Oct 2026
FOCUS: Active Critic Learning for Long-Horizon Language Model Agents
The paper says
“Sparse outcome rewards make it difficult for long-horizon language-model agents to assign credit to earlier actions.”
more of the abstract
Learned critics provide value estimates for intermediate states, but these estimates can be inaccurate, while estimating intermediate values with repetitive continuation samples from points within a trajectory incurs substantial costs. To mitigate these issues, we introduce FOCUS, an RL method that uses selective branching to help the critic network learn its value estimator. The rollout controller facilitates active learning for the critic by sampling continuations from states that are expected to most reduce the critic's estimation error. These refined value estimates then score the segment between consecutive states with a TD(λ) advantage, while the sampled evidence supplies training targets for the critic. FOCUS supports long-horizon agentic reinforcement learning through a modular context-management interface that decouples context management from rollout allocation. Our experiments show that using compaction-induced segment boundaries for value estimation and credit assignment, FOCUS outperforms GRPO and VinePPO with matched per-step generation budgets on legal reasoning, deep search, and code localization benchmarks, improving held-out accuracy at Qwen3.5-9B scale by 7.2 points over the strongest baseline on BrowseComp-Plus (55.4 vs. 48.2) and by 4.1 points on DeonticBench (56.5 vs. 52.4).
from the abstract
Pith says
Machine review · deepseek-v4-flash · 27 Oct 2026 · of version 1 of the paper; version 2 is newer · what was read was not recorded · not checked by a person
It can be wrong. Challenge this review or reply to it.
Holds up, with conditions. A coherent, unusually honest method paper whose active-critic-learning mechanism checks out, but whose headline comparison rests on best-checkpoint selection over single runs and needs fixed checkpoints and multiple seeds before 'outperforms on all three benchmarks' is credible. This line is wrong
4 major objections · 5 minor
How often objections from reviews like this one are right: not yet measured
Not checked: figures, code, what the references say.
The paper says, at Table 1, Section 3.3, Appendix B
The review did not quote this passage.
A classifier read this sentence of the paper as acknowledging the concern (probability 0.85):
“They do not account for held-out checkpoint selection, training-seed variation, or repeated stochastic decoding.”
Pith says: Objection 1 of 4 · major
Table 1 reports each method's best held-out checkpoint from a single training run, and Appendix B states explicitly that the 90% bootstrap intervals are 'conditional on the selected checkpoints' and 'do not account for held-out checkpoint selection, training-seed variation, or repeated stochastic decoding' (b358-b359).
more of objection 1
Best-of-training checkpoint selection inflates each method's reported score by an amount that need not be equal across methods, so neither the magnitudes nor the significance of the headline deltas are established: several intervals contain zero (LocBench File@1 and File@5, Deontic Hard and SARA-N at 9B, and the 27B Airline, Hard, and Overall rows), and the sub-metric picture is mixed, with SARA-N at -4.0 points at 9B. The MATH preliminary itself shows that checkpoint choice can change a ranking: FOCUS leads at updates 8, 16, and 24 but trails GRPO at update 32 (b84). I request a fixed-checkpoint comparison (e.g., last checkpoint, or an average over a defined checkpoint interval) with multiple training seeds, or a selection-aware inference procedure, before 'FOCUS outperforms GRPO and VinePPO' is stated categorically.
The paper says, at Appendix B
The review did not quote this passage.
Pith says: Objection 2 of 4 · major
Several of the reported intervals in Table 1 do not correspond to the reported point estimates. Appendix B states that the DeonticBench-9B intervals are 'preserved rather than recomputed' from an export that 'does not represent these corrected results'; the BrowseComp-Plus interval is 'supplied' because per-question outcomes are not in the local analysis export; and the LocBench columns pair FOCUS checkpoint 35 with GRPO checkpoint 1 for File@1 but checkpoint 89 for File@5.
more of objection 2
These are not paired comparisons of the two methods under a common checkpoint-selection rule, so the sub- and superscript intervals cannot be read as valid uncertainty statements about method-level differences. The corrected per-case exports and per-question outcomes should be released, and all intervals recomputed from the same data as the point estimates.
The paper says, at Appendix C, Section 3.3
The review did not quote this passage.
Pith says: Objection 3 of 4 · major
Appendix C describes a Qwen3.5-27B-supervised imitation warm-start for BrowseComp-Plus 'before the matched RL runs,' but the manuscript does not state whether GRPO and VinePPO are trained from the same warm-started actor. Because the imitation stage trains both compaction and continued search, a FOCUS-only warm-start would confound the headline +7.2 point BrowseComp-Plus result with initialization differences. Please state explicitly whether the warm-start is shared by all three methods and which checkpoint all RL runs start from.
The paper says, at Appendix A, Algorithm 2, Section 3.3
The review did not quote this passage.
Pith says: Objection 4 of 4 · major
The 'matched per-step generation budget' premise is load-bearing for the fairness of the comparison, but the numbers needed to verify it are not in the manuscript. The Appendix A tables report B_seed but not the refinement budget B_ref used in Algorithm 2, and no aggregate per-update generation (tokens, turns, or FLOPs) is reported for the three main benchmarks, in contrast to the MATH preliminary, which reports tokens in Figure 2b.
more of objection 4
With GRPO at group size 8 and 48 root trajectories, VinePPO at group size 4 with one value continuation per branching state, and FOCUS using adaptive search, seeding, and refinement budgets, the 'matched' claim rests on an accounting that is never shown. A per-method per-update token/turn budget table is needed.
5 minor objections
The paper says, at Appendix E, Eq. (10)
Pith says: minor objection 1 of 5
The unbiasedness statement in Eq. (10) concerns the value of the resumed state, and the paper itself defers the compaction-quality/training-stability link to future work (Section 5); a sentence in Section 3.3 noting that the mechanistic reading presumes compaction preserves the value-relevant information would help readers calibrate the interpretation of the gains.
The paper says, at Table 1
Pith says: minor objection 2 of 5
As rendered, the numeric columns merge into concatenated digit strings (e.g., '14.056.092.828.252.4') and the delta rows pack interval endpoints as inline sub/superscripts; the table should be reformatted with one column per metric and bracketed intervals.
show 3 more
The paper says, at Figure 3
Pith says: minor objection 3 of 5
The middle-row held-out panels are evaluated only at a subset of checkpoints, and the BrowseComp-Plus panel caption says the result is the equivalence judgement of Qwen3-32B; please state whether the same evaluator and the same checkpoint schedule underlie the Table 1 numbers.
The paper says, at Section 2.3
Pith says: minor objection 4 of 5
The distinction between N(c-bar) (completed continuation outcomes, used for selection) and n(c-bar) (direct children, used for training targets) is central but easy to miss; a short worked example contrasting the two counters would improve readability.
The paper says, at Section 2.3, Appendix E
Pith says: minor objection 5 of 5
The paper states in Appendix E that the acquisition score is a heuristic and that kappa=1 is a regularization choice whose optimality in the simplified model requires e^2 = sigma^2; consider stating both points in the main text immediately after Eq. (7).
What would settle it
Run GRPO, VinePPO, and FOCUS on DeonticBench and BrowseComp-Plus at Qwen3.5-9B with three independent training seeds each, evaluate held-out accuracy at a fixed pre-registered checkpoint (the final update) instead of the best checkpoint over training, and recompute paired 90% CIs from per-case outcomes, which Appendix B currently cannot do for BrowseComp-Plus (per-case outcomes not included in the export). If the fixed-checkpoint mean deltas remain at or above the current intervals' lower endpoints (+0.59 for DeonticBench Overall, +2.41 for BrowseComp-Plus), the headline claim survives selection-bias scrutiny; if they fall below those endpoints or change sign, the reported gains are an artifact of best-checkpoint selection over single runs.
from the skeptic's reading
How the review went
4 readings by deepseek-v4-flash · 27 Oct 2026
- Reader holds up, with conditions
- Skeptic agrees
- Editor wrote the verdict above
- Referee major revision: 4 major objections, 5 minor
Not run on this paper: the simulated authors' defence.
Pith did not check this review's objections against the paper.
The review in full
The review in brief, machine-written
Two things to know up front. First, the mechanism is coherent and the derivations hold up — this is a real method paper. Second, the headline claim ("outperforms GRPO and VinePPO on all three benchmarks") is stronger than the evidence supports, because every Table 1 entry is that method's best held-out checkpoint from a single training run, and the paper's own Appendix B says the bootstrap intervals condition on the chosen checkpoints.
What's actually new. FOCUS frames extra continuation sampling as active learning for a learned critic: allocate rollouts to states where expected reduction in critic squared error (Eq. 7, Beta posterior, kappa=1) is largest, then use one-level TreeRPO-style backups at compaction boundaries to get segment TD advantages. The detail I like most is Appendix E, where the authors show selection should use raw observed continuation outcomes rather than the recursive backups used for training. Recursive backups attenuate the critic-outcome residual by a factor of four per edge (Eq. 8), which would destroy the selection signal; the worked example makes the point concrete. The derivations check out, and the paper is transparent about what it does not claim: g is explicitly a heuristic, the Eq. 10 unbiasedness caveat is flagged, and the compaction-quality link is left to future work. The BrowseComp-Plus ablation (refinement rollouts drive the gain) and the critic-MSE diagnostics both support the mechanism.
Where it's soft. The empirical core has one genuine problem: best-of-training checkpoint selection over single runs inflates each method's expected reported score by an amount that need not be equal across methods, so the magnitudes of the headline margins are unverified, not just their significance. Several intervals include zero: LocBench File@1 and File@5, Deontic Hard, SARA-N, and 27B Overall. Some intervals are supplied rather than recomputed from per-case outcomes, and the 9B Deontic intervals come from an old export that does not match the corrected results. The MATH preliminary shows checkpoint choice alone can flip a ranking. To be fair, none of this is hidden — the authors say it plainly — and Figure 3's curves have FOCUS ahead at most checkpoints on BrowseComp-Plus and DeonticBench, so the direction of the result is plausible. The honest claim is "ranks first at selected checkpoints with consistently better training curves on the headline cells," not "outperforms on all three benchmarks." What settles it: fixed checkpoints, multiple seeds, and ideally per-case outcome data.
Who this is for: anyone working on RLVR for long-horizon agents, especially critic-based credit assignment and tree-structured rollouts. It deserves a serious referee. I would send it to review with an expectation of major revision — fix the checkpoint/seed comparison or soften the claims to match the evidence.
The strongest objection
The central claim is comparative: FOCUS "outperforms GRPO and VinePPO" on legal reasoning, deep search, and code localization with matched per-step generation budgets (b6, b24), yet every Table 1 entry is each method's best held-out checkpoint over one training run (b112). Appendix B states the paired 90% intervals are "conditional on the selected checkpoints" and "do not account for held-out checkpoint selection, training-seed variation, or repeated stochastic decoding" (b358-b359). The conditioning is practical, not formal: LocBench columns pair FOCUS checkpoint 35 against different GRPO checkpoints per metric (1 for File@1, 89 for File@5); the BrowseComp-Plus interval is "supplied" rather than recomputed from per-question outcomes; and the DeonticBench-9B intervals are "preserved" from an export that "does not represent these corrected results." Zero-containing intervals cover LocBench File@1 (+2.3, -1.1..+5.7), File@5 (+1.1, 0.0..+3.4), Deontic Hard (-1.4, -8.45..+5.63), SARA-N (-4.0, -16.0..+8.0), and 27B Overall (+2.0, -1.18..+5.33); only BrowseComp-Plus, Deontic Airline, and 9B Deontic Overall have strictly positive lower endpoints. Best-of-training checkpoint selection inflates each method's expected reported score by an amount that need not be equal across methods, so the magnitudes (not just the significance) of the headline margins are unverified; the MATH preliminary shows checkpoint choice can change a ranking (b84). Figure 3's curves showing FOCUS ahead at most 9B checkpoints partially mitigate this for the two headline cells, but single runs still do not bound the effect size. I found no hidden formal defect in the mechanism: Eq. 7 follows from differentiating the stated MSE model, Eq. 8's induction is correct, and Appendix E explicitly labels g a heuristic and flags the Eq. 10 unbiasedness caveat (b475, b489).
I reviewed the method's equations and the appendices before judging the empirical claim. The selection and refinement math is internally consistent: Eq. 6 matches a Beta(κv, κ(1−v)) prior posterior, Eq. 7 is the stated MSE model's marginal reduction (Appendix E also calls it a heuristic, b489), and Eq. 8's attenuation example follows by induction. The authors explicitly acknowledge the Eq. 10 unbiasedness caveat (b475), the heuristic nature of g (b489), and the unresolved compaction-quality link (b173), so none of those are new findings. Corroborating diagnostics (Figure 4b critic Spearman improvement, Figure 5b refinement-rollout ablation) and the Figure 3 curves, which show FOCUS ahead at most 9B checkpoints, support the mechanism. What remains load-bearing is the statistical legitimacy of the headline deltas: Table 1 reports best held-out checkpoints over single runs, and Appendix B states the intervals are conditional on the selected checkpoints and ignore checkpoint selection, seed variation, and decoding noise (b358–b359). Several cells include zero (LocBench both metrics, Deontic Hard, SARA-N, 27B Overall), and the Deontic-9B and BrowseComp intervals are "supplied" or "preserved" rather than recomputed from per-case outcomes. The MATH preliminary shows the checkpoint choice can change the ranking (FOCUS leads at updates 8–24 but trails by one answer at 32, b84). The reader's weakest_assumption identified the same premise, and my independent pass confirms it; I found no additional defect that would move the verdict. CONDITIONAL remains the right call, pending the fixed-checkpoint multi-seed comparison and per-case outcome release.
The referee's report
The methodological core is coherent and well motivated. Casting selective continuation branching as active learning for the critic, while deliberately separating the selection statistics (observed continuation outcomes) from the training targets (recursive backups), is a clean and non-obvious design, and the Appendix E analysis showing that recursive backups attenuate the selection signal by a factor of four per edge under kappa=1 (Eq. 8) is a genuinely useful insight. The modular context-management interface is a real contribution to the long-horizon agent RL literature, and the ablations (compaction operator, refinement rollouts, structural vs. free-form prompts) plus the critic diagnostics (Figure 4a) support the internal mechanism. The authors are commendably explicit about the heuristic status of the acquisition score and about the scope of the unbiasedness statement in Eq. (10), and they release code.
Pith's reading of the paper (a machine's summary, not the authors')
Long-horizon language-model agents—tools, search, and code edits over dozens of steps—receive a single sparse reward at the end, so a model cannot tell which early actions mattered. FOCUS is a reinforcement-learning method that treats the extra rollout budget (additional continuations sampled from intermediate states) as an active-learning resource for a learned critic network, spending each continuation where the critic's value estimate is expected to improve the most. The refined values then score each segment between context compactions with a temporal-difference ($\mathrm{TD}(\lambda)$) advantage, giving every segment its own credit signal. At a matched per-step generation budget with Qwen3.5-9B, the paper reports that FOCUS beats the critic-free baselines GRPO and VinePPO on legal reasoning, deep search, and code localization, gaining 7.2 points on BrowseComp-Plus (55.4 vs. 48.2) and 4.1 points on DeonticBench (56.5 vs. 52.4). If the claim holds, a fixed training budget produces better long-horizon agents when extra samples are aimed where the critic is wrong rather than spread uniformly.
What carries the argument. The load-bearing object is the acquisition score $g(\bar{c}) = (2\kappa^2\hat{e}^2 + (N-\kappa)\hat{\sigma}^2)/(\kappa+N)^3$ (Eq. 7), an estimate of the marginal reduction in the critic's squared estimation error from one more continuation at a state, built from a Beta posterior (Eq. 6) that treats the critic prediction as $\kappa$ pseudo-observations against $N$ observed continuation outcomes. Around it sits a rollout tree whose nodes are resumable states (in the agent experiments, the points where context compaction has replaced the transcript), a bottom-up value refinement that blends the critic with a one-level recursive backup (Eqs. 1–2), and a segment temporal-difference residual that becomes the actor's advantage (Eq. 4). A deliberate design choice carries the method: selection scores use raw continuation outcomes, not the recursive backups, because backups attenuate the critic-error signal geometrically—by a factor of four per level under the default $\kappa=1$ (Appendix E)—which would misrank states for further sampling. This same machinery converts compaction boundaries, otherwise just a way to fit a long conversation into a context window, into cheap reusable checkpoints for value learning and credit assignment.
Core claim. The paper's central claim is that rollout allocation itself is the lever: under a finite budget of additional continuations, intermediate value estimates—and the policy gradients built from them—improve most when each new continuation is assigned to the state with the largest expected reduction in the critic's squared estimation error. FOCUS runs each episode as a search tree whose nodes are resumable states, taken in the agent benchmarks to be the boundaries where context compaction has summarized the transcript; each completed continuation contributes one outcome observation at its starting state. The acquisition score $g(\bar{c}) = (2\kappa^2\hat{e}^2 + (N-\kappa)\hat{\sigma}^2)/(\kappa+N)^3$ (Eq. 7) ranks states by marginal MSE reduction, and collected evidence is blended with the critic prediction $\tilde{V}(\bar{c}) = (\kappa V_\phi(\bar{c}) + n(\bar{c})\bar{G}(\bar{c}))/(\kappa+n(\bar{c}))$ (Eq. 2) to refine values bottom-up. Those refined values produce segment-level temporal-difference residuals $\delta_i = R_i + \gamma^{d_i}\tilde{V}(\bar{c}_i) - \tilde{V}(\bar{c}_{i-1})$ (Eq. 4) for the actor and recursive backup targets for the critic (Eq. 3). At Qwen3.5-9B with matched per-step generation budgets, the paper reports FOCUS outperforms GRPO and VinePPO on all three benchmarks, avoids the late-training collapse in tool use seen under trajectory-level group credit, and loses most of its BrowseComp-Plus gain when the refinement rollouts are removed.
The premise it rests on. The headline comparisons assume that the best held-out checkpoint from a single training run stands in for each method's expected performance, and the selection rule assumes that a continuation resumed from a compacted context is an unbiased sample of that state's value even though the compaction discards information.
If the paper is right (Pith's inferences, not the authors')
Under a matched per-step generation budget, FOCUS claims higher held-out accuracy than GRPO and VinePPO on deep search, legal reasoning, and code localization, with the largest gain (7.2 points) on BrowseComp-Plus at Qwen3.5-9B scale.
Refined segment values keep long-horizon exploration stable: the paper reports that GRPO's mean tool calls collapse late in training as it submits right after compaction, while FOCUS and VinePPO stay near their initial levels.
The critic itself improves over training: rank agreement with an oversampled, critic-blind MCTS target rises on LocBench, and the mean acquisition score $g(\bar{c})$ falls by about half over 32 updates on MATH.
The gains are not tied to one compaction implementation—in-context and in-memory compaction score similarly on DeonticBench—but the refinement rollouts themselves are necessary: without them FOCUS trains visibly more slowly on BrowseComp-Plus.
The advantage persists at Qwen3.5-27B with LoRA on DeonticBench, though the margin narrows, which the paper attributes in part to the already-high untrained baseline.
A natural extension the paper leaves implicit is an adaptive stopping rule: instead of a fixed refinement budget, allocate continuations until $g(\bar{c})$ falls below a threshold, which would turn the fixed-budget comparisons into an accuracy-versus-tokens cost curve.
The unbiasedness premise of Eq. (10) is directly testable: comparing value estimates from compacted versus uncompacted resumptions of the same states would probe whether lossy compaction biases the critic's training targets, and would speak to the compaction-quality and training-stability link the authors defer to future work.
The allocation rule is logically separate from the credit-assignment scheme that consumes its evidence; swapping the advantage estimator (for example, group-relative or process-reward) while keeping the $g(\bar{c})$ scheduler would isolate how much of the win comes from where samples are spent versus how they are scored.
The selection posterior could be upgraded to graded, non-binary outcomes or a full predictive distribution for the critic; the authors fix $\kappa=1$ pseudo-observations and note that $\kappa^*=\sigma^2/e^2$ would be optimal in their simplified MSE model, making per-state tuning of $\kappa$ a cheap and informative experiment.
Check this review
Pith stored no record of this review's passes and took no fingerprint of it. Reviews made by code from 13 Nov 2026 on carry both.
How often reviews like this one are right
Reviews like this one: written by deepseek-v4-flash, where what was read was not recorded.
How often objections from reviews like this one are right: not yet measured
Reviews by people
This review was written by a machine. No person has reviewed this paper on Pith yet. A scientist listed on Pith can review it for a fee, and their review is published here whatever it concludes.
The paper
Abstract
Sparse outcome rewards make it difficult for long-horizon language-model agents to assign credit to earlier actions. Learned critics provide value estimates for intermediate states, but these estimates can be inaccurate, while estimating intermediate values with repetitive continuation samples from points within a trajectory incurs substantial costs. To mitigate these issues, we introduce FOCUS, an RL method that uses selective branching to help the critic network learn its value estimator. The rollout controller facilitates active learning for the critic by sampling continuations from states that are expected to most reduce the critic's estimation error. These refined value estimates then score the segment between consecutive states with a TD(λ) advantage, while the sampled evidence supplies training targets for the critic. FOCUS supports long-horizon agentic reinforcement learning through a modular context-management interface that decouples context management from rollout allocation. Our experiments show that using compaction-induced segment boundaries for value estimation and credit assignment, FOCUS outperforms GRPO and VinePPO with matched per-step generation budgets on legal reasoning, deep search, and code localization benchmarks, improving held-out accuracy at Qwen3.5-9B scale by 7.2 points over the strongest baseline on BrowseComp-Plus (55.4 vs. 48.2) and by 4.1 points on DeonticBench (56.5 vs. 52.4).
The paper's own sentences, arranged, kept beside the review and separate from it. Each label is drawn from a closed vocabulary; the number beside it is the probability the classifier gave that label. No sentence in this record was written by a model.
abstract compiled · 2026-10-27
The authors' main claim, as the classifier picked it out 1.00
“Our experiments show that using compaction-induced segment boundaries for value estimation and credit assignment, FOCUS outperforms GRPO and VinePPO with matched per-step generation budgets on legal reasoning, deep search, and code localization benchmarks, improving held-out accuracy at Qwen3.5-9B scale by 7.2 points over the strongest baseline on BrowseComp-Plus (55.4 vs. 48.2) and by 4.1 points on DeonticBench (56.5 vs. 52.4).”
How the main claim is stated: asserted 1.00
What the abstract itself states, as probabilities: claims first 0.04 · states limitation 0.20 · human participants 0.04 · controlled comparison 0.95 · randomized assignment 0.02 · states quantitative result 0.99 · compares with prior results 0.98 · mentions code or data availability 0.03
The abstract, sentence by sentence
This table labels the abstract only (arXiv listing abstract v1). The number is the probability the classifier gave the label.
| Sentence | Role | The authors' own work | States the main conclusion | Strength |
|---|---|---|---|---|
| Sparse outcome rewards make it difficult for long-horizon language-model agents to assign credit to earlier actions. | background 0.98 | 0.08 | 0.02 | asserted 1.00 |
| Learned critics provide value estimates for intermediate states, but these estimates can be inaccurate, while estimating intermediate values with repetitive continuation samples from points within a trajectory incurs substantial costs. | background 0.74 | 0.14 | 0.02 | asserted 0.92 |
| To mitigate these issues, we introduce FOCUS, an RL method that uses selective branching to help the critic network learn its value estimator. | contribution 0.98 | 0.84 | 0.13 | asserted 1.00 |
| The rollout controller facilitates active learning for the critic by sampling continuations from states that are expected to most reduce the critic's estimation error. | method 0.98 | 0.79 | 0.04 | asserted 1.00 |
| These refined value estimates then score the segment between consecutive states with a TD($\lambda$) advantage, while the sampled evidence supplies training targets for the critic. | method 1.00 | 0.84 | 0.05 | asserted 1.00 |
| FOCUS supports long-horizon agentic reinforcement learning through a modular context-management interface that decouples context management from rollout allocation. | method 0.69 | 0.79 | 0.06 | asserted 1.00 |
| Our experiments show that using compaction-induced segment boundaries for value estimation and credit assignment, FOCUS outperforms GRPO and VinePPO with matched per-step generation budgets on legal reasoning, deep search, and code localization benchmarks, improving held-out accuracy at Qwen3.5-9B scale by 7.2 points over the strongest baseline on BrowseComp-Plus (55.4 vs. 48.2) and by 4.1 points on DeonticBench (56.5 vs. 52.4). | result 1.00 | 0.97 | 0.87 | asserted 1.00 |
What kind of paper the classifier reports this is: method or tool 1.00 · evidence from observation or experiment 0.98
What was examined, and what was not
Sources read
- arXiv listing abstract v1
Not examined
- full text
- figures and tables
- references and citing sentences
- relations to other papers
Levels run
- P abstract
Not run
- S sections
- T body sentences
- C citations
- R cross-paper relations
How the labels were made
Compiled by classifier-1.13.0 via TypeSafe API, question bank abstract-p1-20260923, 2026-10-27: 1 request, 5,707 input tokens.
Every quoted sentence on this record is the authors'. Every label is from Pith's closed vocabulary, with the probability shown beside it. This record as JSON
Figures (5)
Figures are the authors', reproduced from the paper.
Challenges and discussion
Disagree with the review, or know something it missed? Leave a comment, or challenge this review formally: a challenge becomes a permanent entry in this paper's record, and its decision is published here.
Come back by email
A short list of the week's newly reviewed papers. Nothing is sent until you confirm from your inbox, and every email has a link that stops it. No account.
No comments yet.
Sign in to comment