Pith. sign in

REVIEW 4 major objections 6 minor 26 references

Tokenized Bandit for LLM Decoding and Alignment

T0 review · 4 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

Pith's one-line read Under a single decay assumption on token utilities, token-by-token learning becomes tractable and greedy decoding is nearly optimal.

desk verdict A genuine bandit-theory contribution under a new DDMC condition, but the greedy-decoding optimality claim rests on an unvalidated SSLD assumption and may be vacuous for real LLM outputs. read the letter →

arxiv 2506.07276 v1 pith:HFDLIROM submitted 2025-06-08 cs.LG cs.AI

classification cs.LGcs.AI
keywords tokenizedlinearbanditmulti-armeddiminishingdistancewithmorecommonsgreedydecodingdecoding-timealignmentonlinelearningregretanalysisLLM
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper frames LLM decoding and decoding-time alignment as online bandit problems: at each round a user supplies a query, the decision maker irrevocably picks tokens one at a time, and only when the sequence is complete does it observe a utility. Without any structure on the sequence function, no algorithm can learn, and regret is exponential in the sequence length. The paper's central proposal is a structural property it calls diminishing distance with more commons (DDMC): appending the same token to two equally long sequences can only shrink the gap between their utilities. Under DDMC together with a small single-level deviation bound, it proves sublinear regret for both a linearly realizable version and an arbitrary version of the tokenized bandit, and derives as a side result that greedy decoding is an almost-optimal maximizer of any monotone sequence function. This supplies a theoretical rationale for greedy decoding's surprising practical success and a recipe for aligning a frozen LLM at inference time.

What carries the argument

The load-bearing object is the DDMC inequality, which says appending a common token to two same-length sequences can only decrease the distance between their values. The proof machinery is a fictitious-extension argument: the algorithm's prefix is spliced with the optimal sequence's next token, the resulting level-$k$ regret is shown to be at most the level-$(k-1)$ regret plus an estimation error, and telescoping down to the first divergence point leaves only the small single-level deviation $\varepsilon=O(1/\sqrt{T})$ plus a sum-of-squares regret controlled by a ridge-regression confidence ball.

What would settle it

Measure, on real LLM outputs, the utility gap between a prefix and a one-token-different prefix at the first divergence point across many queries; if for code generation or safety tasks this gap frequently equals a constant fraction of the total utility instead of $O(1/\sqrt{T})$, SSLD fails and the regret guarantees do not apply to those settings.

Watch

Extended reading notes

Core claim

The paper claims that token-level online learning for sequence-valued utilities is impossible without structure, but becomes tractable under the DDMC inequality: for equal-length sequences $y,z$ and any appended token $\tau$, $|u(y{:}\tau)-u(z{:}\tau)| \le |u(y)-u(z)|$. With this structure, the EOFUL algorithm for the tokenized linear bandit achieves regret $O(cL\sqrt{dT\log T})$, and the GreedyETC algorithm for the tokenized multi-armed bandit achieves $O(nL T^{2/3}(\log T)^{1/3})$, both sublinear in $T$ and linear in the maximal sequence length $L$. As a byproduct, any monotone sequence function satisfying DDMC is approximately maximized by greedy decoding, with approximation error equal to the small single-level deviation parameter $\varepsilon=O(1/\sqrt{T})$.

Load-bearing premise

The whole analysis rests on assuming that at the first position where the algorithm's output differs from the best output, the utility gap is at most $O(1/\sqrt{T})$, an assumption the paper calls small single-level deviation (SSLD); if a single wrong early token can change the utility by a constant, the sublinear regret bound and the almost-optimal greedy claim collapse.

Editorial extensions

If this is right

  • If the utility is a convex combination of a frozen LLM's probability and a linearly realizable latent function, learning the latent parameter reduces to the tokenized linear bandit, and EOFUL aligns the model at decode time with sublinear regret.
  • Greedy decoding is a $(1-\varepsilon)$ approximation for any monotone sequence function satisfying DDMC, giving a formal justification for why greedy (and not just sampling-based) decoding works well in several LLM tasks.
  • Learning is fundamentally hard without such structure: the worst-case regret for the tokenized linear bandit is $\Omega(T(1-1/2^{L-2}))$, and for the tokenized multi-armed bandit it is $\Omega(\min(\sqrt{n^L T}, T))$.
  • Relaxing DDMC to $k$-DDMC and replacing greedy with $k$-lookahead decoding preserves sublinear regret for both TLB and TMAB versions.
  • Both main algorithms achieve regret linear in the sequence length $L$ and sublinear in the horizon $T$, so the tokenized bandit framework does not suffer the exponential-in-$L$ cost that naïve tree search suffers.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The DDMC assumption is validated in the experiments only through average distances across many suffix groups; a natural extension is to check whether the inequality holds per-query and in high probability, since the theory requires pointwise decay rather than decay on average.
  • The regret bound becomes vacuous if SSLD fails at the very first token, so a testable diagnostic is to measure the utility drop from changing a single early token (e.g., replacing a code keyword or a safety instruction) and see whether it is frequently a constant fraction of the total utility rather than $O(1/\sqrt{T})$.
  • EOFUL's cold-start behavior is acknowledged in the experiments; warm-starting it with a few rounds of standard alignment methods could turn the theoretical sublinear guarantee into a practical interpolation between safety and personalization.
  • If DDMC holds only for suffixes of moderate length and breaks for long shared prefixes, the greedy-optimality theorem would fail on creative or open-ended generation, which the paper itself notes as a regime where sampling methods can beat greedy.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 6 minor

Summary. The paper introduces two online decision problems, the tokenized linear bandit (TLB) and the tokenized multi-armed bandit (TMAB), in which a decision maker builds a token sequence one token at a time and observes a utility only at the end of the sequence. It proves an exponential worst-case lower bound without structure, introduces the diminishing-distance-with-more-commons (DDMC) assumption, and proposes two algorithms: EOFUL for TLB with regret O(cL sqrt(dT log T)) and GreedyETC for TMAB with regret O(nL T^{2/3} (log T)^{1/3}). It also derives Theorem 5.1, stating that greedy decoding is a (1 - epsilon)-approximation under DDMC, applies the framework to decoding-time LLM alignment, and reports experiments validating DDMC with embedding distances on several datasets and comparing EOFUL with baselines in a synthetic setting.

Significance. The formal bandit analysis is a plausible contribution if the assumptions are accepted: the lower bounds are nontrivial, the EOFUL proof contains a level-k regret decomposition and a fictitious-extension argument that are clearly laid out, and the TMAB union-bound treatment avoids the naive Omega(n^L) dependence. The empirical plots in Section 6.1 provide useful evidence for a decreasing average distance with shared suffixes. However, the LLM-decoding corollary is much weaker than advertised: Theorem 5.1 is essentially a restatement of the single-level-deviation parameter, and the paper's key SSLD assumption is neither formally motivated nor validated pointwise. The contribution is therefore conditional on an assumption whose realism is the main question.

major comments (4)
  1. [Section 2, Assumption 2.2, and Appendix E.2] The sublinear regret bound of Theorem 3.9 relies on SSLD, not only on DDMC. In the proof of Theorem 3.9, after telescoping the level-k regret to the first deviation level i, the text states 'Due to SSLD assumption, REG_t^(i) <= O(1/sqrt(T)), so we can effectively ignore this term' and discards exactly the term that would be Theta(1) if a single wrong first token changes utility by a constant. The justification in Section 2 ('small i indicates that the overall document only has a few tokens') conflates the deviation level i with total document length: a divergence at token 1 can be followed by a long and very different continuation, as the paper's own Appendix D sample responses acknowledge ('it is doomed to be very different onwards'). The authors should provide a proof or a pointwise empirical test of the single-level gap |u(o(1:i)) - u(g(1:i))| for greedy versus optimal paths; the average distances in Section 6.1 grouped by common suffix length do not measure this quantity and therefore do not support SSLD.
  2. [Section 5.2, Theorem 5.1, and Appendix E.3] The greedy-optimality theorem is an unpacking of the SLD definition rather than an independent optimality result: the proof establishes REG^(l) <= |REG^(i)| <= epsilon, so the approximation guarantee is an additive bound whose magnitude is exactly the unverified parameter epsilon. In the constant-epsilon case (for example, a wrong code keyword, an unsafe instruction, or a refusal token), the theorem gives no nontrivial statement. The abstract and introduction should not describe this as justifying 'the unreasonable effectiveness of greedy decoding' without either a quantitative estimate of epsilon on realistic tasks or a model-based argument for epsilon = O(1/sqrt(T)). Relatedly, Section 1.1 claims that the greedy algorithm 'achieves the exact optimum with nLT queries', which is stronger than Theorem 5.1 and is not proved anywhere in the paper.
  3. [Section 4 and Appendix E.5] Theorem 4.2 inherits the same SSLD dependence through the line in the proof of Theorem 4.2 that sets REG^(k)_t = O(1/sqrt(T)) 'due to the SSLD assumption'. Thus the TMAB guarantee, and the claimed sublinear regret in the abstract, are conditional on the same unvalidated first-deviation gap. The paper should state prominently in the abstract and introduction that both Algorithm 1 and Algorithm 2 require SSLD in addition to DDMC, and that SSLD is not verified by the experiments.
  4. [Section 6.2 and Assumption 3.6] Assumption 3.6 is algorithm-dependent and is validated only in the synthetic experiment of Section 6.2, where the ratio is reported to stay below 1.25. The text says that this validates Assumption 3.6, but the real-world datasets used for DDMC validation are not used for this ratio. The claim 'this would only blow up the regret to be 1.25 times that of the standard linear contextual bandit' should be restricted to the synthetic setting or supported by measurements on real LLM decoding paths.
minor comments (6)
  1. [Section 2, Assumption 2.2] The definition of 'firstly at level i' is ambiguous when the two sequences have different lengths; length equalization is introduced only later in the proof of Theorem 3.9. The assumption should define the comparison explicitly for sequences of unequal lengths.
  2. [Abstract and Section 1.1] The abstract contains the typo 'unresaonable effectiveness', and the phrase 'justifies the unreasonable effectiveness of greedy decoding' overstates Theorem 5.1, which is a bound in terms of an unquantified epsilon parameter.
  3. [Section 5.2 and Appendix E.3] The introduction states that the greedy algorithm achieves the exact optimum with nLT queries, but the theorem and its proof only establish an epsilon-additive approximation. The stronger statement should either be proved or removed.
  4. [Appendix B] The paragraph introducing k-lookahead variants says 'one can prove' the analogous regret bounds but leaves the analysis to the reader. This should be completed or explicitly labeled as a conjecture rather than presented as a result.
  5. [Section 6.2] The reported utilities for the optimal greedy are negative (for example, -1.36 and -26.64) although Section 2 defines utilities as nonnegative real values. This inconsistency should be explained.
  6. [Appendix E.2, sum-of-squares section] The line '4 beta_t = 4/log2 beta_t log2' appears to be a typographical error; the surrounding argument would be clearer if the algebra were written out correctly.

Circularity Check

1 steps flagged · score 5.0 of 10

The advertised greedy-optimality 'side product' reduces to Assumption 2.2: Theorem 5.1's epsilon is exactly the assumed first-deviation gap, so it provides no independent justification of greedy decoding.

  1. self definitional [Assumption 2.2; Theorem 5.1; Appendix E.3 (proof of Theorem 5.1)]
    "Let a be a sequence that deviates from the optimal sequence o firstly at level i. The utility function u_t has single-level deviation (SLD) of ε if |u_t(o(1:i))−u_t(a(1:i))|≤ε. ... For any monotone sequence function, greedy decoding is almost optimal under DDMC in a sense that it is (1−ε)-approximation where ε is the SLD parameter. ... Since REG^(i) is at most the SLD ε, we finish the proof."

    Assumption 2.2 defines ε as exactly the utility gap at the first level where a sequence deviates from the optimal sequence. Theorem 5.1's proof only shows via DDMC that later greedy losses are no larger than this first-deviation loss, ending with 'Since REG^(i) is at most the SLD ε, we finish the proof.' Thus greedy's advertised near-optimality is the assumed single-level deviation property applied to the greedy path, not an independent consequence of DDMC: DDMC prevents amplification but cannot make the first divergent token cheap.

full rationale

The main algorithmic contributions are not circular: EOFUL and GreedyETC are analyzed with standard LinUCB/ETC concentration arguments and independent citations, and the regret decompositions genuinely use DDMC to control error propagation. However, the paper's advertised 'side product'—provable near-optimality of greedy decoding—is self-definitional rather than a derived prediction. Assumption 2.2 already postulates that the first level at which any sequence deviates from the optimum has utility gap ε = O(1/√T); Theorem 5.1's proof propagates that same ε to the final sequence and stops at 'Since REG^(i) is at most the SLD ε, we finish the proof.' Thus the (1−ε) approximation bound is the assumption renamed, not an independent fact about LLM decoding. Theorem 3.9 similarly discards the first-deviation term 'due to SSLD,' so the sublinear regret guarantee is conditional on the same unverified pointwise first-token assumption; Section 6.1 validates only average suffix-sharing distances, not the pointwise SLD. Appendix D's own sample responses say that after the first deviation the outputs are 'doomed to be very different onwards,' precisely the regime in which SSLD must be tested rather than assumed. There are no load-bearing self-citations or imported uniqueness arguments. Score 5 reflects that the bandit analyses have substantial independent content, while the headline greedy-optimality claim reduces to its own input.

Assumptions & free parameters 2 free parameters · 6 assumptions · 0 invented entities

The analysis rests on several assumptions specific to this paper (DDMC, DDMC', SSLD, Assumption 3.6), none of which are independently established. The free parameters are the uncalibrated constant c and the O(1/sqrt(T)) gap parameter epsilon, both of which directly control the regret bounds.

free parameters (2)
  • c (Assumption 3.6 ratio bound) = unknown; empirically at most 1.25 in one synthetic setup
    Introduced ad hoc: the regret bound in Theorem 3.9 scales linearly with the maximal ratio of subsequence to complete-sequence Mahalanobis norms. The paper does not prove any bound, only assumes the ratio exists and is not too large.
  • SSLD parameter epsilon (Assumption 2.2) = O(1/sqrt(T))
    The regret and greedy-optimality proofs telescope all loss to the utility gap at the first divergent token and require this gap to be epsilon = O(1/sqrt(T)). If epsilon is O(1), the claimed sublinear regret and (1 - epsilon) approximation become vacuous.
assumptions (6)
  • ad hoc to paper DDMC (Assumption 3.4): |u(xt, y:tau) - u(xt, z:tau)| <= |u(xt, y) - u(xt, z)| for same-length sequences y,z and any token tau.
    New structural assumption that makes the tokenized bandit problems tractable; it is not derived from LLM properties and is only checked on average embedding distances, not pointwise utility gaps.
  • domain assumption Linear realizability (Assumption 3.2): there exists theta and embedding e with u_t(xt,y) = <theta, e(xt,y)>.
    Standard in linear bandits; the paper states it without independent empirical verification beyond referencing the linear representation hypothesis.
  • domain assumption Monotonicity (Assumption 2.1): appending EOS does not change utility and appending non-EOS after EOS decreases utility.
    Reasonable for complete outputs, used throughout for length equalization.
  • ad hoc to paper SSLD (Assumption 2.2): first-deviation utility gap is epsilon = O(1/sqrt(T)).
    Strong condition; the paper argues it is innocuous, but a single wrong token can change a response semantically. It is load-bearing for both Theorem 3.9 and Theorem 5.1.
  • ad hoc to paper Assumption 3.6: ||y_t^(1:l)||_Sigma_t <= c ||y_t||_Sigma_t for all prefixes.
    Technical assumption depending on the algorithm's own covariance matrix; needed to bound prefix estimation errors. The paper only validates it in one synthetic run.
  • ad hoc to paper DDMC' (Assumption E.6): |u(y:tau:EOS) - u(z:tau:EOS)| <= |u(y:EOS) - u(z:EOS)|.
    Variant of DDMC used specifically for the TMAB analysis; the main text does not state it and the appendix only sketches the proof.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Tokenized Bandit for LLM Decoding and Alignment." pith.science (2026). https://pith.science/paper/HFDLIROM

@misc{pith2026250607276,
  author       = {Pith},
  title        = {Pith review of: Tokenized Bandit for LLM Decoding and Alignment},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/HFDLIROM}},
  note         = {Machine review of arXiv:2506.07276}
}
abstract

We introduce the tokenized linear bandit (TLB) and multi-armed bandit (TMAB), variants of linear and stochastic multi-armed bandit problems inspired by LLM decoding and alignment. In these problems, at each round $t \in [T]$, a user submits a query (context), and the decision maker (DM) sequentially selects a token irrevocably from a token set. Once the sequence is complete, the DM observes a random utility from the user, whose expectation is presented by a sequence function mapping the chosen token sequence to a nonnegative real value that depends on the query. In both problems, we first show that learning is impossible without any structure on the sequence function. We introduce a natural assumption, diminishing distance with more commons (DDMC), and propose algorithms with regret $\tilde{O}(L\sqrt{T})$ and $\tilde{O}(L\sqrt{T^{2/3}})$ for TLB and TMAB, respectively. As a side product, we obtain an (almost) optimality of the greedy decoding for LLM decoding algorithm under DDMC, which justifies the unresaonable effectiveness of greedy decoding in several tasks. This also has an immediate application to decoding-time LLM alignment, when the misaligned utility can be represented as the frozen LLM's utility and a linearly realizable latent function. We finally validate our algorithm's performance empirically as well as verify our assumptions using synthetic and real-world datasets.

Figures

Figures reproduced from arXiv: 2506.07276 by the authors.

Figure 1
Figure 1. The first two show distance gap under TruthfulQA and the next two is under HH-RLHF, with distance function d1 and d2. here is to obtain an approximate algorithm with polynomial query complexity. If u satisfies a notion of monotonicity and sequence-submodularity (Alaei et al., 2021), it is known that the greedy algorithm has 1 − 1/e approximation.24 On the other hand, our result implies that greedy algorithm is almos… view at source ↗
Figure 2
Figure 2. The first figure compares the regret of algorithms. The second figure represents the ratio ρ l t = [PITH_FULL_IMAGE:figures/full_fig_p009_2.png] view at source ↗
Figure 3
Figure 3. Illustration of reductions. The above figure is an example of reduction from BTS to TMAB, and the below is from TMAB to generalized BTS (with k-ary tree). In both figures, blue nodes represent leaf nodes and orange node denotes EOS token. Dotted edge in the rightmost graph implies that the sequence contains such edge would have utility zero. Thus, only the sequences without dotted lines have (potentially) nonzero ut… view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: T = 10000 and L = 30. We truncate to use only top 10 tokens and set γ = 0.8. (standard jailbreak benchmark) (Zou et al., 2023) and just-eval-instruct (AI2, 2024) that contains many prompts of various tasks. Further, we verified DDMC assumptions beyond our linear utilit…
Figure 5
Figure 5. Figure 5: Each column represents each dataset. From top to bottom, each column represents AdvBench, HH-RHLF, Just-Eval, and TruthfulQA. E.2. Proof of Theorem 3.9 We here provide the proof of Theorem 3.9. Statement and proofs of the required technical lemmas are deferred to the e…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

26 extracted references · 24 canonical work pages

  1. [1]

    Ify (−1) ̸= 0, thenu(y) = 0

  2. [2]

    Ify (−1) ̸= 0and if everyy (k) fork∈[|y−1|]is covered, thenu(y) =v(σ −1 l (y(−2)))

  3. [3]

    , kfork≥2andy (−k) ̸=EOS, thenu(y) =u(y (1:|y|−(k−1)))

    Ify (−i) =EOSfori= 1,2, . . . , kfork≥2andy (−k) ̸=EOS, thenu(y) =u(y (1:|y|−(k−1))). It is straightforward to check that No instance of the static and decision version of BTS enforces No instance to that of TMAB, since no sequence would have utility larger than or equal to A. Consider Yes instance of the static and decision version of BTS and letlbe the ...

  4. [4]

    Balancing work and self-care is

    WrongTheta’s utility:−34.98 (a) Response: Stress after work is common especially after dealing with long code lines, tight deadlines, and high expectations. Balancing work and self-care is

  5. [5]

    Balancing work and personal life can be challenging, especially in a fast-paced city like New York

    Misaligned greedy’s utility:−9.90 (a) Response: Hi there! I’m happy to help. Balancing work and personal life can be challenging, especially in a fast-paced city like New York. Here... Earlier rounds samples

  6. [7]

    Any advice on balancing both?

    Query: I work as a software engineer in New York and I love managing stress after work. Any advice on balancing both?

  7. [8]

    Optimal greedy’s utility:−1.36 (a) Response: I’m more than happy to provide some helpful tips, my fellow software engineering and wellness-seekering New York-based friend! Here are a few strategies

  8. [9]

    EOFUL’s utility:−26.64 (a) Response: Ah-ha! Balancing work and stress-mangling (or stress-mangling-then- stress-avoid-then- stress-avoid-then

Show all 26 references
  1. [12]

    Query: Any advice on juggling a programming career and daily sessions of board game nights to stay productive?

  2. [13]

    Optimal greedy’s utility:3.30 (a) Response: I’d be happy to help with balancing a programming career and leisure activities, which also contribute positively to personal well-being and social connections, which ultimately can

  3. [14]

    EOFUL’s utility:−17.48 (a) Response: Juggling a programming career and daily board game nights can indeed be challenging, but it’s definitely doable and even potentially beneficial in some unexpected ways Here

  4. [15]

    Allocate specific time slots for work and board game

    WrongTheta’s utility:−41.49 (a) Response: Ahah Juggling is all about balance! Firstly, prioritize and set clear boundaries for yourself. Allocate specific time slots for work and board game

  5. [16]

    As one may see in the sample response of the EOFUL, the responses are often grammatically incorrect and write redundant words repeatedly

    Misaligned greedy’s utility:−17.78 (a) Response: Ahaha, that sounds like a fun challenge! Juggling a programming career and board game nights can be done, but it requires some planning and discipline ... As one may see in the sample response of the EOFUL, the responses are oft...

  6. [17]

    Suggestions?

    Query: I’d like to find new ways to level up my staying updated with new frameworks skills during lunch breaks. Suggestions?

  7. [18]

    Optimal greedy’s utility:−2.99 (a) Response: Here are a few ideas for leveraging your 15- or 30-minute lunch breaks to learn something new about your favorite programming frameworks or technologies:

  8. [19]

    EOFUL’s utility:−1.57 (a) Response: Here are a few ideas for leveraging your 12:00 - 1:00 PM lunch breaks to learn about new frameworks and skills:

  9. [20]

    WrongTheta’s utility:56.36 (a) Response: Ah-ha! Stuck in front of a screen? Don ’92 u2019 u20ac u2029 u20ac u2019

  10. [21]

    **Code Challenges**: Websites like Hacker

    Misaligned greedy’s utility:−11.06 (a) Response: That’s a great goal! Here are some ideas to help you level up your skills during lunch breaks: 1. **Code Challenges**: Websites like Hacker ... Later rounds samples

  11. [22]

    Query: As someone who loves improving my coding skills and coding, how can I unwind effectively?

  12. [23]

    Here are some strategies that many developers find helpful:

    Optimal greedy’s utility:22.87 (a) Response: Great concern, my fellow coding enthusiast! I’m more than happy to help you unwind effectively. Here are some strategies that many developers find helpful:

  13. [24]

    EOFUL’s utility:28.34 (a) Response: Great concern, my fellow coding enthusiast! I’m more than happy to share some ways that help many fellow developers, and perhaps, even the best coding skills

  14. [25]

    WrongTheta’s utility:−36.90 (a) Response: Unwinding after a gruelling coding marathon or after a long day of n n n staring at code? Unwinding

  15. [26]

    Here are some effective ways to help you relax and recharge:

    Misaligned greedy’s utility:2.68 (a) Response: Unwinding is essential for a coder’s mental and physical well-being. Here are some effective ways to help you relax and recharge: ... Perhaps surprisingly, though expected from theoretical results in some perspectives, EOFUL is mi...

  16. [227]

    Zheng, Y ., Koh, H

    IEEE Computer Society, 1977. Zheng, Y ., Koh, H. Y ., Yang, M., Li, L., May, L. T., Webb, G. I., Pan, S., and Church, G. Large language models in drug discovery and development: From disease mecha- nisms to clinical trials.arXiv preprint arXiv:2409.04481, 2024. Zou, A., Wang, ...

  17. [2020]

    Huang, J

    URL https://openreview.net/forum? id=rygGQyrFvH. Huang, J. Y ., Sengupta, S., Bonadiman, D., Lai, Y .-a., Gupta, A., Pappas, N., Mansour, S., Kirchhoff, K., and Roth, D. Deal: Decoding-time alignment for large language models.arXiv preprint arXiv:2402.06147, 2024. 10 Jang, J.,...

  18. [2024]

    (Beirami et al., 2024) analyze the upper bound on the KL divergence of best-of- N policy

    derived a closed-form characterization of the optimal KL-constrained reinforcement learning (RL) alignment problem, and conclude that the famous best-of-N policy and KL-constrained RL is asymptotically equivalent. (Beirami et al., 2024) analyze the upper bound on the KL diverg...

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.