Machine Learning · arXiv:2611.09701 v1 · submitted 6 Nov 2026 · 20 references
Where Should Learning Live? Regret Analysis for Text, Parameter, and Joint Adaptation
The paper says
“Language-model systems can select prompts and parameter adapters repeatedly during deployment, using feedback to guide later choices. Earlier choices may change the context or environment inherited by later configurations.”
more of the abstract
We study selection among preconstructed candidates and compare cumulative loss with the best fixed prompt–parameter pair used from the beginning. This excess loss separates into three terms: selection regret within the available family, the gap from restricting that family, and the benefit or harm of the actual deployment history. Selection is evaluated using the history each candidate would have produced if used throughout, which separates candidate quality from the effects of switching. When loss functions are fixed independently of selector randomness, past choices affect loss for only a bounded number of rounds, and every candidate's own-history loss is revealed, limiting switches yields sublinear expected regret within the selected family. We also quantify the costs and additional assumptions of independent probing and deployment-only feedback. Controlled experiments show that fewer switches can either lower or raise deployed loss, depending on whether inherited histories are harmful or helpful. With Qwen3.5-9B on Spider, text-only selection has lower within-family regret than joint selection, yet produces more failures because its candidate family excludes stronger configurations.
from the abstract
Pith says
Machine review · deepseek-v4-flash · 9 Nov 2026 · what was read was not recorded · not checked by a person
It can be wrong. Challenge this review or reply to it.
Holds up. Solid, honestly-scoped regret analysis; the genuinely new result is the joint-versus-coordinatewise separation, while the headline sublinear bound lives behind a counterfactual oracle the authors themselves show deployment logs cannot supply. This line is wrong
3 minor
How often objections from reviews like this one are right: not yet measured
Not checked: figures, code, what the 20 references say.
3 minor objections
The paper says, at Conclusion
Pith says: minor objection 1 of 3
The conclusion’s opening sentence states that low regret, eventual optimal selection, and infrequent switching can each occur alongside higher cumulative loss. This is accurate, but because the abstract’s sublinear-regret headline is often read in isolation, I suggest adding a phrase such as “under the diagonal-oracle assumptions of Theorem 2” to the first sentence of the conclusion so the conditional character of the main guarantee is visible without reading Appendix A.
The paper says, at Appendix A.3.2
Pith says: minor objection 2 of 3
The sentence “Assumption 2 requires valid simulation or cloning of each constant history” is the most important scope caveat of the paper; consider promoting it to a short paragraph at the end of Section 4, since a reader of only the main text could miss Appendix A.3.2’s explicit statement of the oracle requirement.
show 1 more
The paper says, at Section 5, Eq. (19)
Pith says: minor objection 3 of 3
The factor b multiplying the sum over g_j in the definition of Γ_b(Q) is easy to overlook; one sentence in the main text stating that Γ_b is measured in cumulative-loss units over the bJ complete blocks would improve readability.
What would settle it
Extend the controlled fixed-advantage study of Appendix C.5 (same loss law, m_Z=m_W=4, lambda=0.2, four candidates, uniform prior) to horizons T=256, 1,024, 4,096, computing exact expected per-round within-family policy regret for (i) full-diagonal lazy EW with rate (12) and (ii) deployment-only block EW with b=1 and rate (170). If per-round regret decays as T^{-1/2} under (i) but only as T^{-1/3} or slower under (ii), Assumption 2 is confirmed load-bearing and the sublinear headline holds only in the oracle regime; if (ii) matches (i)'s scaling at this m=4 instance, deployment-only feedback nearly recovers the oracle rate and the practical weight of Assumption 2 is much smaller than the bounds suggest.
from the skeptic's reading
How the review went
4 readings by deepseek-v4-flash · 9 Nov 2026
- Reader holds up
- Skeptic agrees
- Editor wrote the verdict above
- Referee accept: 3 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
First: this paper deserves a serious referee, and the caveat the stress-test flags is real but already disclosed by the authors. Second: the result genuinely worth stealing is the joint-versus-coordinatewise separation, not the headline regret bound.
What is new. The loss decomposition (Prop 1) is add-and-subtract algebra, but it is the right accounting tool: diagonal selection, family restriction, and signed state effect are exactly the quantities practitioners conflate when comparing prompt, parameter, and joint adaptation. Prop 4 and Cor 1 are the real results — a one-sweep coordinatewise sampler can suffer Omega(sqrt(T)) regret against joint sampling while both eventually fix on the optimal configuration. That separation is nontrivial and survives reading. Theorem 2's lazy-EW bound under bounded influence assembles standard machinery (Geulen–Voecking–Winkler coupling, Hoeffding, KL change of measure) cleanly; I re-derived the chain through Lemmas 1–2 and the tuning in (12)–(13) and it holds.
The paper is unusually honest about scope. Appendix A.3.2 says Assumption 2 requires valid simulation or cloning of each constant history; Appendix A.4.9 shows one-step importance weighting does not identify diagonal values and that window importance sampling is invalid under lazy logging; the deployment-only rate is T2/3 with the reaction term possibly linear. Nothing is hidden.
Soft spots, in proportion. The load-bearing condition is Assumption 2, and in the persistent-workspace setting that motivates the paper, that oracle is exactly what you lack. The abstract's 'limiting switches yields sublinear expected regret' is true only under parallel counterfactual simulation. The authors say this; still, the headline framing does more work than the assumptions support, and I would want the abstract to carry the caveat's emphasis. Their own Table 2 shows the practical distance: excess at T=1,024 is 61.84 with full diagonal versus 232.26 deployment-only at b=1.
Empirically, the controlled synthetic studies are competent and the decomposition checks hold to floating-point tolerance, but the SQL evaluation is one model, two benchmarks, no code or data artifact, and several key contrasts have standard errors comparable to the observed differences (joint versus parameter on Spider; lazy versus independent on SParC). The authors mostly report this honestly.
Who it is for: anyone designing evaluation protocols for deployed LLM agents, and online-learning readers interested in policy regret under memory. The increments are genuine and cleanly separated from what is assumed.
Recommendation: send it to a serious referee. The main revision ask is reframing the headline to match the oracle dependence; the math is in good shape.
The strongest objection
I re-derived the main proof chain: Lemma 1's marginals and refresh bound, Lemma 2's contamination accounting, Theorem 2's combination with Theorem 1, the tuning in (12)-(13), the Proposition 4/Corollary 1 constants, and Corollary 12's T^{2/3} rate all check out. The load-bearing condition is Assumption 2, the diagonal counterfactual oracle: sublinear within-family policy regret requires d_t(a)=F_t(a,...,a) revealed for every candidate after every round. This is acknowledged (Appendix A.3.2 requires 'valid simulation or cloning of each constant history'), and Appendix A.4.9 shows it is not slack: one-step importance weighting does not identify diagonal values, and window importance sampling (Lemma 8) is invalid under lazy logging ('lazy logging violates Assumption 3') with second moment exponential in m. What remains unresolved is the reach, not the validity, of the claim. In the persistent-workspace setting of Section 1 these counterfactuals are unavailable; Theorem 4 and Corollary 12 then leave deployment-only regret at O(T^{2/3}) with the causal-reaction term Gamma_b 'uncontrolled' and possibly linear. The paper's own Table 2 shows the practical gap: mean excess at T=1,024, lambda=0.2 is 61.84 (full diagonal) versus 232.26 (deployment-only b=1). The abstract's headline, 'limiting switches yields sublinear expected regret,' therefore holds only under a parallel cloning-and-simulation protocol, not under deployment feedback. This is an honestly disclosed scope limitation, not an inconsistency, but it is the single condition on which the central claim's practical force rests.
I verified the central arguments rather than accepting them on trust. Lemma 1's lazy-coupling kernel preserves marginals with expected refreshes at most eta(T-1); Lemma 2's pathwise contamination bound is consistent with Assumption 1's coordinate windows; Theorem 2's decomposition into diagonal selection plus state charge, with the tuning yielding 2 sqrt(T C_Q (m_max+1/8)), is arithmetically correct. Proposition 4 and Corollary 1 constants reproduce (weights (4n)^{-gamma}, per-coordinate retention (4k/(4k+1))^2, Pr(E_t) >= e^{-1/2}/(4 sqrt(n)), Borel-Cantelli fixation), and Theorem 4 plus Corollary 12's O(T^{2/3}) rate with the reduction to Dekel et al.'s switching-cost lower bound hold. The empirical sections carry independent support: the decomposition residuals ran at or below 2.8e-11 and additive EW equivalence held to floating-point tolerance (Appendices C.2-C.3), so implementation error is unlikely to drive the headline numbers; the SParC lazy-minus-independent contrast (2.375 +/- 1.626) is reported with honest uncertainty. The single load-bearing condition is Assumption 2's diagonal oracle, exactly the reader's weakest assumption. My sharpening is that the paper's own identification analysis (Appendix A.4.9) shows the oracle cannot be replaced by estimation from logged deployment data for the lazy algorithm, so the sublinear regime exists only when K parallel counterfactual trajectories are maintained from round one. The paper discloses this and Theorem 4/Corollary 12 delimit the oracle-free regime. The title question, where learning should live, is deliberately left open by the theory ('different upper bounds alone do not establish which mode performs better'), and the empirical answer rests on one model and benchmark pair; these are scope limitations the reader already priced into a moderate-confidence ACCEPT. Nothing found changes the verdict.
The referee's report
The contribution is solid and honestly scoped. The decomposition (2) is a simple but useful accounting identity for the setting, and the paper is careful to state what each feedback model can and cannot guarantee. The proofs in Appendix B use standard arguments—Hoeffding, KL change of measure, lazy coupling, importance weighting—and I found the constants internally consistent; the refresh-count lemma, the contamination accounting in Lemma 2, and the tuning in (12)–(13) all check out. The central sublinear regime is conditional on Assumption 2’s full diagonal counterfactual oracle, and the manuscript explicitly acknowledges in Appendix A.3.2 and A.4.9 that this requires valid simulation or cloning of constant histories and is not identifiable from ordinary logged deployment data; the deployment-only analysis retains the causal-reaction term and is reported as possibly linear. This is a scope limitation rather than an internal inconsistency, and it is handled with unusual candor. The empirical component is auxiliary and finite-family, with appropriate uncertainty caveats; it supports, rather than overclaims, the theory.
Pith's reading of the paper (a machine's summary, not the authors')
When a language-model agent repeatedly picks a prompt and an adapter during deployment, each pick rewrites the context that the next pick inherits, so any comparison against "the best fixed configuration" mixes several things at once. The paper establishes an exact accounting split: deployed excess loss over the best fixed prompt–parameter pair equals within-family selection regret plus family restriction plus a signed state effect, with no hidden remainder, and only the restriction term is never negative. Its main theorem then shows that selection regret can be made sublinear in a stateful world: if past choices influence current loss for only $m_{\max}$ bounded rounds and every candidate's own constant-history loss is revealed after each round, lazy, switch-limiting exponential weights achieve $O(\sqrt{(1+m_{\max})T})$ expected policy regret within the candidate family. The cost of weaker feedback is quantified rather than hidden — sparse probes pay an extra factor in the selection term, and deployment-only feedback degrades to a $T^{2/3}$ rate with an uncontrolled reaction term. These distinctions matter in practice: on Spider, text-only selection had lower within-family regret than joint selection and yet worse deployed loss, because its candidate family excluded stronger configurations.
What carries the argument. Three mechanisms carry the argument. First, the diagonal-loss accounting identity of Proposition 1: adding and subtracting the selected family's cumulative diagonal loss and the best within-family value splits deployed excess loss exactly into selection regret, family restriction, and the signed state effect, so each term can be measured and bounded separately. Second, the diagonal loss $d_t(a) = F_t(a,\ldots,a)$, which evaluates each candidate on the constant history it would have produced if used from the start; Assumption 2 postulates an oracle that reveals all of these after every round, and Assumption 1 bounds how long a past choice can keep influencing current loss through hidden state, so a switch can contaminate at most $m_{\max}$ future rounds. Third, lazy exponential weights coupled by a shrinking-dartboard scheme: the deployed action keeps the exponential-weights marginals $q_t$, but it is retained with probability $e^{-\eta d_t(A_t)}$ and only refreshed otherwise, giving at most $\eta(T-1)$ expected refreshes; combining the standard EW selection bound with this refresh count produces the state charge $m_{\max}\eta(T-1)$ that Theorem 2 adds to the selection term.
Core claim. The central discovery is a conditional theorem plus the accounting identity that exposes its condition. Proposition 1 states that, on every trajectory and in expectation, $L^{\mathrm{dep}}_T - V^{ZW}_{T,C}$ splits exactly into diagonal selection regret $L^{\mathrm{diag}}_T - V^e_{T,C}$, family restriction $V^e_{T,C} - V^{ZW}_{T,C}$, and the signed state effect $L^{\mathrm{dep}}_T - L^{\mathrm{diag}}_T$, where each term compares against the history a fixed candidate would have produced if used throughout. Under Assumption 1 (component-specific bounded influence, horizon $m_{\max}$) and Assumption 2 (a diagonal counterfactual oracle revealing every candidate's own-history loss each round), Theorem 2 proves that lazy exponential weights — a low-switching Hedge variant that keeps its current candidate with probability $e^{-\eta d_t(A_t)}$ — satisfy $R_{\mathrm{pol}}^T(Q) \le C_Q/\eta + \eta[T/8 + m_{\max}(T-1)]$, which tunes to $O(\sqrt{(1+m_{\max})T\,C_Q})$ expected policy regret against any fixed comparator distribution $Q$. The state penalty enters as the influence horizon times the expected number of refreshes, not through the candidate count. The paper is equally explicit about what the oracle buys: with one independent probe per round the selection term becomes $\eta K T/2$, and with deployment-only block feedback the rate falls to $O((mK C_Q)^{1/3}T^{2/3})$ with an uncontrolled causal-reaction term that can remain linear.
The premise it rests on. After every round the selector is assumed to see, for every candidate, the loss that candidate would have incurred had it alone been deployed from the start — a counterfactual history that requires valid simulation or cloning of each constant trajectory, and that ordinary one-step importance weighting from logged deployment data cannot recover, as the paper itself notes.
If the paper is right (Pith's inferences, not the authors')
The state cost of adaptation is set by the influence horizon, not the candidate count: freezing one coordinate (text-only or parameter-only selection) drops that coordinate's switching contribution, so modes whose changes echo longer in the workspace pay more.
Low within-family regret, eventual selection of the optimum, and infrequent switching can each coexist with high cumulative loss; none of the three alone certifies deployed performance.
Deployment-only feedback cannot reach the oracle rate: under bounded influence the block algorithm achieves only $O((mK C_Q)^{1/3}T^{2/3})$, with an uncontrolled causal-reaction term, so the counterfactual evaluator — not the learner — is what buys sublinearity.
Joint sampling and a single coordinatewise sweep are equivalent under additive losses with a product prior, but a barrier interaction separates them: coordinatewise regret can be $\Omega(\sqrt{T})$ even when both selectors eventually deploy the optimum on every sufficiently late round.
On Spider, text-only selection shows lower within-family regret than joint selection yet higher deployed loss because its candidate family excludes stronger configurations; equalizing candidate counts narrows or reverses the observed mode ordering.
The decomposition doubles as an engineering diagnostic: measure selection regret, family gap, and signed state effect separately on any deployed agent, and the dominant term tells you where improvement effort belongs — the Spider results imply that optimizing within-family regret alone can chase a term that is not the binding constraint.
The diagonal oracle is the practical bottleneck the theory quietly depends on; the sparse-probe theorem prices a partial remedy (one probe per round for a factor-$K$ evaluation saving at an extra $K/2$ selection cost), but a natural extension the paper leaves open is a probe-budget scheduler that spends more probes on plausible candidates.
Proposition 7's matched-marginal coupling suggests a field protocol: run two copies of an agent that differ only in switching laziness, and the sign of their deployed-loss gap reveals whether the workspace's inherited history is harmful (favor laziness) or helpful (favor fresh draws) before committing to a selection rule.
Bounded influence is effectively a design prescription: agents that truncate or reset their workspace memory within a fixed horizon — fresh sandboxes, dialogue resets, bounded context windows — qualify for the sublinear guarantee, so the theory gives a concrete reason to prefer architectures that limit how long one choice can contaminate the future.
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
Language-model systems can select prompts and parameter adapters repeatedly during deployment, using feedback to guide later choices. Earlier choices may change the context or environment inherited by later configurations. We study selection among preconstructed candidates and compare cumulative loss with the best fixed prompt–parameter pair used from the beginning. This excess loss separates into three terms: selection regret within the available family, the gap from restricting that family, and the benefit or harm of the actual deployment history. Selection is evaluated using the history each candidate would have produced if used throughout, which separates candidate quality from the effects of switching. When loss functions are fixed independently of selector randomness, past choices affect loss for only a bounded number of rounds, and every candidate's own-history loss is revealed, limiting switches yields sublinear expected regret within the selected family. We also quantify the costs and additional assumptions of independent probing and deployment-only feedback. Controlled experiments show that fewer switches can either lower or raise deployed loss, depending on whether inherited histories are harmful or helpful. With Qwen3.5-9B on Spider, text-only selection has lower within-family regret than joint selection, yet produces more failures because its candidate family excludes stronger configurations.
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 · 2026-11-09
The authors' main claim, as the classifier picked it out 0.57
“Controlled experiments show that fewer switches can either lower or raise deployed loss, depending on whether inherited histories are harmful or helpful.”
How the main claim is stated: asserted 0.98
What the abstract itself states, as probabilities: claims first 0.04 · states limitation 0.87 · human participants 0.04 · controlled comparison 0.89 · randomized assignment 0.04 · states quantitative result 0.23 · compares with prior results 0.88 · 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 |
|---|---|---|---|---|
| Language-model systems can select prompts and parameter adapters repeatedly during deployment, using feedback to guide later choices. | background 1.00 | 0.09 | 0.03 | asserted 0.60 |
| Earlier choices may change the context or environment inherited by later configurations. | background 1.00 | 0.11 | 0.03 | hedged possibility 0.94 |
| We study selection among preconstructed candidates and compare cumulative loss with the best fixed prompt--parameter pair used from the beginning. | contribution 0.91 | 0.84 | 0.05 | asserted 0.94 |
| This excess loss separates into three terms: selection regret within the available family, the gap from restricting that family, and the benefit or harm of the actual deployment history. | contribution 0.43 | 0.84 | 0.16 | asserted 1.00 |
| Selection is evaluated using the history each candidate would have produced if used throughout, which separates candidate quality from the effects of switching. | method 0.97 | 0.80 | 0.06 | asserted 1.00 |
| When loss functions are fixed independently of selector randomness, past choices affect loss for only a bounded number of rounds, and every candidate's own-history loss is revealed, limiting switches yields sublinear expected regret within the selected family. | result 0.85 | 0.74 | 0.23 | asserted 0.99 |
| We also quantify the costs and additional assumptions of independent probing and deployment-only feedback. | contribution 0.88 | 0.85 | 0.03 | asserted 0.97 |
| Controlled experiments show that fewer switches can either lower or raise deployed loss, depending on whether inherited histories are harmful or helpful. | result 0.98 | 0.94 | 0.23 | asserted 0.98 |
| With Qwen3.5-9B on Spider, text-only selection has lower within-family regret than joint selection, yet produces more failures because its candidate family excludes stronger configurations. | result 0.99 | 0.95 | 0.26 | asserted 1.00 |
What kind of paper the classifier reports this is: theoretical result 0.92 · evidence from observation or experiment 0.77
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, 2026-11-09: 1 request, 6,822 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 (12)
Showing 10 of the paper's 12 figures. Figures are the authors', reproduced from the paper.
References (20)
- [1] The counterexample invalidates substituting current-action losses on the realized prefix, not Theorem 2’s diagonal oracle. Proposition 10(External regret versus few switches).Letℓ 1 ≡0andℓ t(a) =1{a=A t−1}for t≥2, and let ST = TX t=2 1{At ̸=A t−1}.(86) Every learner satisfies ... on Pith
- [2] arXiv preprint arXiv:2609.00196. J. Kirkpatrick, R. Pascanu, N. Rabinowitz, J. Veness, G. Desjardins, A. A. Rusu, K. Milan, J. Quan, T. Ramalho, A. Grabska-Barwinska, D. Hassabis, C. Clopath, D. Kumaran, and R. Hadsell. Over- coming catastrophic forgetting in neural networks.P... on PithPith's reviewarXiv
- [3] Joint access need not measure interaction: L(z, w) = 1−0.4z−0.4won{0,1} 2 with base(0,0)has a positive joint-effect gap but zero factorial interaction. These gaps also differ from finite-candidate restriction in the Spider comparison and per-use target discrepancyε e; see Appe... on Pith
- [4] arXiv preprint arXiv:2605.12484. D. Wu, H. Wang, W. Yu, Y . Zhang, K.-W. Chang, and D. Yu. LongMemEval: Benchmarking chat assistants on long-term interactive memory. InProceedings of the International Conference on Learning Representations (ICLR), on PithPith's reviewarXiv
- [5] Exogenous tasks are shared across counterfactual histories, while deployment-responsive tasks, users, and environments are rerun under each history. The accounting identity (2) conditions on this response tree. We refine its restriction and averaging conventions below. For tex... on Pith
- [6] A frozen language model can induce the text prior, while the parameter prior may encode rank, precision, sparsity, or provenance. Corollary 3(Prior-atom description-length regret).Under the product priorµ=π⊗ρ, for positive- mass atomsz ⋆, w⋆, putC π =−logπ({z ⋆})andC ρ =−logρ(... on Pith
- [7] The paired lazy-minus- independent difference was2.375±1.626failures, or0.197±0.135percentage points; see Fig- ure 7(b). Here1.626is the MCSE of the paired mean, conditional on the evaluated banks, bench- mark, and model responses. The estimated difference is about 1.5 MCSEs, ... on PitharXiv
- [8] Bounded multiplicative lossesd t(z, w) =gt(z)ht(w)generally do not factorize it, but nonadditivity alone need not separate regret. Exact Constants for the Fixed-Rate Separation.For the barrier lossfof Proposition 4, the exact constants are as follows. For every fixedη t =η >0a... on Pith
- [9] Our guarantees apply after the family and prior are fixed. For nonempty measurableC e =C ∩ Ae,e∈ {Z, W}, define V eT,C = infa∈Ce TX t=1 Lt(a),(56) DisceT (Ce) =V eT,C −V eT ≥0,(57) ΛeT,C (Qe) = TX t=1 Z Ce Lt(a) dQe(a)−V eT,C ≥0,(58) whereQ e is supported onC e. Discovery char... on Pith
- [10] A.4 FEEDBACKREQUIREMENTS ANDEXTENSIONS The feedback interface determines control of selection and state in (2). Full diagonal feedback evaluates each candidate’s own trajectory, sparse probing reduces evaluations, and deployment-only learning requires control of causal reactio... on Pith
- [11] provide further precedents. A.4.1 ACCURACY OFFULLDIAGONALEVALUATION These results relax Assumption 2 conditional on an exogenous response tree. For random trees, the relevant independence, mean, or error assumptions must hold conditionally. Unconditional unbi- asedness does no... on Pith
- [13] Refined Bound for Arbitrary Probe Distributions.Theorem 3 in Section 5 specializes a more detailed inequality. Under Assumption 3, form=m max, every fixedQandη >0satisfy RpolT (Q)≤ CQη + η2 TX t=1 EVt +mη T−1X t=1 E⟨qt, dt⟩.(97) An infinite expected second moment makes this bo... on Pith
- [15] Without assuming total finite influence, RpolT (Q)≤ bCQη + ηbKJ2 + min{h0, b}(J−1) +A b(Q) +r.(113) Sublinear regret requiresA b(Q) =o(T)and appropriateb, η, neither of which the definition sup- plies. A suffix-sensitivity bound can control reaction without exact forgetting. P... on Pith
- [16] Se- quential evaluation requires sequential intervention (Jiang & Li, 2016). Lemma 8(Window importance sampling in a sufficient model).Consider a fixed deterministic response table with total joint influence horizonm(for coordinate influence horizons,m=m max is sufficient),t≥m... on Pith
- [17] The difference on the remainingrrounds is at mostr, since both losses lie in [0,1]. Proof of Lemma 7.The first block’s common empty prefix gives zero discrepancy. In later blocks, u≥m+ 1repetitions make both coordinate windows constant, equating the losses in (19). At most min... on Pith
- [20] Sparse feedback uses one twenty-fourth as many logical diagonal entries, with paired sparse-minus- full mean differences of8.807±2.752expected failures on Spider and16.630±1.872on cleared SParC. The increase is larger and has a smaller training-seed SE on cleared SParC. These ... on Pith
- [2012] cannot universally be improved to √ Tunder deployment-only feedback (Dekel et al., 2014). Appendix B.4 gives the reduction to bandits with switching costs. This leaves optimalmand prior dependence in (111) open. Sparse probes provide stronger information. To separate initial w... on Pith
- [2018] and SParC (Yu et al., 2019). Construction, evaluation, and replay protocols are fixed before compar- ison. Selection constructs no new prompts or adapters. Spider and SParC were jointly partitioned by database, keeping identical schemas together: their official development dat... on Pith
- [2023] arXiv preprint arXiv:2310.08560. R. Pryzant, D. Iter, J. Li, Y . Lee, C. Zhu, and M. Zeng. Automatic prompt optimization with “gradient descent” and beam search. InProceedings of the Conference on Empirical Methods in Natural Language Processing (EMNLP), on PithPith's reviewarXiv
- [2026] arXiv preprint arXiv:2606.05661. P. Auer, N. Cesa-Bianchi, Y . Freund, and R. E. Schapire. The nonstochastic multiarmed bandit problem.SIAM Journal on Computing, 32(1):48–77, on PithPith's reviewarXiv
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