Pith. sign in

REVIEW 4 major objections 5 minor 28 references

BAR Conjecture: the Feasibility of Inference Budget-Constrained LLM Services with Authenticity and Reasoning

T0 review · 4 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read The paper proves a BAR theorem: for sufficiently long inputs, no LLM can meet a fixed inference budget while preserving both factual authenticity and multi-step reasoning.

desk verdict The BAR 'Theorem' is an assumption dressed as a proof: the authenticity cost is asserted in Premise A2, Lemma 2 just restates it, and the final result is a tautology, though the informal three-way taxonomy is a fair description of practice. read the letter →

arxiv 2507.23170 v2 pith:NTOWV7QE submitted 2025-07-31 cs.LG

Jinan Zhou , Rajat Ghosh , Vaishnavi Bhargava , Debojyoti Dutta , Aryan Singhal This is my paper · ORCID
classification cs.LG
keywords BARtheoremLLMinferencebudgetauthenticityreasoningchain-of-thoughtlowerboundretrieval-augmentedgenerationimpossibilityresultlatency
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 sets out to prove that LLM application design faces an unavoidable trade-off: once inputs grow past a critical length, no model can simultaneously keep inference within a fixed time budget, stay factually authentic, and reason through multi-step problems. The authors formalize this as the BAR theorem, define losses for budget, authenticity, and reasoning, and derive a threshold length $n^\star$ beyond which the latency cost of mandatory reasoning tokens and mandatory retrieval calls exceeds any fixed budget $T$. If the theorem is right, every deployed system must explicitly choose two of the three properties, which the paper maps onto three existing design types: budget–authenticity services such as real-time assistants, authenticity–reasoning services such as deep research tools, and budget–reasoning services such as coding assistants. The practical upshot is that the trilemma is structural, not something a better model alone removes.

What carries the argument

The load-bearing mechanism is the additive latency lower bound produced by three premises and two lemmas. Premise A1 says multi-step reasoning needs at least $\Omega(n)$ scratchpad tokens; Lemma 1 turns that into a required $c_1 \tau n$ seconds. Premise A2 says factual authenticity on non-synthetic queries needs at least $k \ge 1$ retrieval or verification calls; Lemma 2 turns that into a required $k\rho$ seconds. Premise A3 states that every forward pass, retrieval, and tool call draws from the same wall-clock budget, giving the combined lower bound $L_{\mathrm{lat}} \ge c_1 \tau n + k\rho$ (with a memory-bandwidth twin $(c_1\mu n + k\beta)/B_{\max}$). The threshold $n^\star = \lceil (T - k\rho)/(c_1\tau)\rceil$ is where the linear reasoning term alone crosses the fixed budget.

What would settle it

Search for a task family with unbounded input length and a deployed system that, on instances of every length $n$, simultaneously holds a fixed latency $T$, achieves the stated authenticity and reasoning thresholds, and makes zero retrieval calls at runtime. One concrete candidate is a fine-tuned closed-book QA model that answers long multi-step factual queries under a latency cap without scratchpad tokens and without external tools; if such a system clears all three bars for arbitrarily large $n$, Lemma 2 and Premise A2 fail.

Watch

Extended reading notes

Core claim

The central claim is Theorem 2.3: fix tolerances $\varepsilon_h, \varepsilon_r > 0$ and an inference budget $T > 0$; there exists an input length $n^\star$ such that no LLM can satisfy $L_{\mathrm{auth}} \le \varepsilon_h$, $L_{\mathrm{reason}} \le \varepsilon_r$, and $L_{\mathrm{lat}} \le T$ for any task of size $n \ge n^\star$. The proof adds two lower bounds: reasoning requires at least $c_1 n$ chain-of-thought tokens, each costing at least $\tau$ seconds, and authenticity requires at least $k$ external retrieval or verification calls, each costing at least $\rho$ seconds, so any compliant system pays $c_1 \tau n + k\rho$ in wall-clock time. With $n^\star = \lceil (T - k\rho)/(c_1\tau)\rceil$, all tasks of length $n \ge n^\star$ exceed the budget. The authors read this as a trilemma: for large tasks, a deployed system can prioritize at most two of budget, authenticity, and reasoning.

Load-bearing premise

The argument assumes that factually authentic answers require at least one external retrieval or verification call at inference time; the paper states this as Premise A2 without proving it, and it would fail if a model's weights alone could supply the necessary factual grounding.

Editorial extensions

If this is right

  • For any task family whose inputs can grow without bound, a fixed-latency service can satisfy at most two of the three BAR criteria once $n \ge n^\star$.
  • Designers should classify applications as budget–authenticity, authenticity–reasoning, or budget–reasoning; real-time assistants cannot also do deep multi-step reasoning, and deep research assistants cannot stay real-time.
  • Shrinking per-token latency $\tau$, retrieval latency $\rho$, or memory traffic only moves $n^\star$; the linear chain-of-thought term keeps the trilemma in force for sufficiently large inputs.
  • A system that skips retrieval to save budget still pays the reasoning lower bound, so the budget-versus-reasoning trade-off binds even for closed-book deployments.

Reading between the lines

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

  • The theorem's hardest step is Premise A2, asserted without proof; if authenticity can be achieved entirely through training, then the 'trilemma' reduces to a two-way reasoning-versus-budget trade-off.
  • The paper equates authenticity with runtime retrieval, but a model that has memorized the needed facts could pass the same authenticity criterion with $k=0$; testing that case would separate the genuine theorem from the premise.
  • A concrete empirical program follows from the proof: fix a model and benchmark, vary input length under a constant token or wall-clock cap, and measure where reasoning accuracy or factuality drops; the shape of that crossover would calibrate $n^\star$ per task family.
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 / 5 minor

Summary. The paper claims a formal impossibility result, the "BAR Theorem," stating that for sufficiently large inputs no LLM can simultaneously satisfy an inference-time budget, factual authenticity, and reasoning quality. The argument is built on three premises: reasoning requires Ω(n) chain-of-thought tokens (Premise A1), authenticity requires external retrieval or verification calls (Premise A2), and all such operations consume the inference budget (Premise A3). Lemmas 1 and 2 convert these premises into latency lower bounds, and Theorem 2.3 combines them with a bandwidth term to define a threshold n⋆ beyond which the budget constraint is violated. The paper also proposes a taxonomy of real-world systems that prioritize two of the three properties.

Significance. If the BAR Theorem were valid, it would be a substantive systems-level insight with direct design implications for LLM applications. The paper usefully assembles empirical observations and references on latency, retrieval cost, and reasoning overhead, and its BA/AR/BR taxonomy is a reasonable organizing framework. However, the formal claim is not established: the proof of Lemma 2 restates Premise A2 as its conclusion, the theorem's threshold n⋆ is defined so that the claimed violation follows algebraically from the premises, and Premise A2 is an unproven empirical assertion that is contradicted by retrieval-free factual systems. The paper's contribution is therefore an empirical conjecture and a taxonomy, not a proven theorem.

major comments (4)
  1. [Section 2.2.2, Lemma 2] The proof of Lemma 2 simply restates Premise A2 as its conclusion. Premise A2 says factuality "typically requires" at least k ≥ 1 external retrieval or verification calls, but Lemma 2 asserts a universal "must" for every model satisfying Lauth ≤ εa. The definition of Lauth in Eq. (6) is a KL divergence between q and pθ; nothing in that definition forces q to be obtained by runtime retrieval. A model that approximates q during training—for example by distillation from a RAG teacher, knowledge editing, or preference optimization—could satisfy Lauth ≤ εa with zero retrieval calls. Closed-book QA systems are an existence proof of this possibility. Since Lemma 2 is the only source of the kρ term in Eq. (8), the authenticity half of the impossibility theorem has no valid derivation. This is load-bearing: without Lemma 2, Theorem 2.3 reduces to a reasoning-only latency lower bound and does not establish the claimed trilemma.
  2. [Section 2.3, Theorem and Eq. (8)] The theorem's conclusion is essentially tautological given the definitions. n⋆ is defined as ⌈(T − kρ)/(c1τ)⌉, so for every n ≥ n⋆ the inequality c1τ n + kρ > T holds by construction. This establishes only that if a model must spend at least c1τ per input token and kρ per input, then a budget T smaller than those costs is violated for long inputs. It does not prove that any LLM must spend those amounts; those costs are imported from Premises A1 and A2. Moreover, the theorem quantifies over "no LLM" and "tasks of size n", but Lemma 1 applies only to a specific family of length-parameterised reasoning tasks, and Premise A2 is an empirical regularity. Therefore the formal claim of impossibility is not established.
  3. [Section 2.1, Premise A2] Premise A2 is the linchpin of the authenticity lower bound, but it is asserted as an axiom without proof. It is phrased as a hedged empirical observation ("typically requires"), and the cited evidence (WebGPT, RAG energy studies) concerns systems that choose retrieval, not systems for which retrieval is necessary. The paper does not rule out models whose factual competence is achieved at training time; indeed, instruction-tuned models with RLHF/DPO are explicitly discussed in the Introduction as improving faithfulness without runtime retrieval. An impossibility theorem cannot rest on an unproven and empirically contestable premise; at most, the paper can claim a conditional statement "if authenticity required retrieval, then...", which is far weaker than the abstract's claim.
  4. [Section 2.2.1, Lemma 1 and Premise A1] Premise A1 is also an assumption, not a theorem about all LLMs. The cited lower-bound results (Amiri et al.; Merrill et al.) hold for specific transformer classes (e.g., constant-depth, finite-precision, poly(n)-embedding) on specific tasks. The theorem, however, states an impossibility for "no LLM" on "tasks of size n ≥ n⋆" without restricting the task family or the model class. As stated, the theorem is at least ambiguous for tasks that do not require chain-of-thought (e.g., copying or memorized lookup), and the proof does not specify which task family is intended. The proof needs either a precise quantification over a task family or a more modest claim.
minor comments (5)
  1. [Title and Abstract] The title calls the contribution a "Conjecture" while the abstract and Section 2.3 call it a theorem; the terminology should be made consistent.
  2. [Section 2.1 and Lemma 2] Units are inconsistent: ρ is defined as a latency in seconds, but Lemma 2's statement refers to "additional inference FLOPs of at least kρ", and the theorem later treats kρ as seconds. The paper should consistently distinguish compute cost, latency, and memory traffic.
  3. [Section 2.1, Premise A1] There are grammatical errors in the premise statement (e.g., "it have been shown") and the phrase "poly(n) embedding size" should read "poly(n)-dimensional embedding".
  4. [Section 2.3, Theorem] The notation for the authenticity tolerance alternates between εa in Eq. (6) and εh in the theorem statement; please use one symbol consistently.
  5. [Section 2.5, Limitations] The limitations section states that architectural innovations "cannot remove the Ω(n) term mandated by Premise A1"; this treats an assumption as a proved fact and should be rephrased as conditional on the premise.

Circularity Check

1 steps flagged · score 6.0 of 10

The BAR theorem's authenticity leg is Premise A2 restated as Lemma 2; n⋆ inherits the assumed kρ, so the trilemma is partially assumed rather than derived.

  1. other [Section 2.2.2 Lemma 2; Premise A2 in Section 2.1; Eq. (8) in Section 2.3]
    "Any model satisfying this authenticity constraint must incur an additional inference FLOPs of at least kρ. Proof. By Assumption A2, achieving Lauth ≤ εa on non-synthetic queries requires at least k ≥ 1 external retrieval (or verification) calls to approximate the reference density q(· |x) and to compute or bound the KL divergence."

    The lemma's conclusion—'must incur ... at least kρ'—is exactly Premise A2's 'unavoidable additive latency of at least kρ', and the proof's only argument is 'By Assumption A2'. A2 is hedged ('typically requires') yet Lemma 2 upgrades it to a universal 'must'. The Lauth definition in Eq. (6) does not force runtime retrieval: q is a reference density, and a model trained offline to match q could satisfy Lauth ≤ εa with zero external calls. The theorem then carries kρ into n⋆ in Eq. (8), so the impossibility threshold is built from the assumed cost. The authenticity leg of the trilemma reduces to its input by construction.

full rationale

Lemma 1 is a valid algebraic consequence of Premises A1 and A3, so the reasoning leg is not circular. The circularity is concentrated in the authenticity leg: Lemma 2 does not prove a lower bound; it restates Premise A2 as a conclusion. Since satisfying a KL constraint on the model's distribution does not require computing or bounding that KL at inference, the 'must incur kρ' claim is not derived from Eq. (6). The theorem's n⋆ is chosen after kρ is imported from A2, so the advertised impossibility for n≥n⋆ is conditional on the very factuality-needs-retrieval claim it purports to establish. No self-citation chain is involved; the score is driven by an assumption dressed as a lemma and used as the theorem's pivotal lower bound.

Assumptions & free parameters 4 free parameters · 3 assumptions · 0 invented entities

The central impossibility result is an algebraic summation of three premises; none of the premises is derived in the paper. Premise A2 is asserted without evidence, and the theorem is essentially a restatement of the assumptions. The paper therefore adds no independent scientific content beyond the cited lower bounds.

free parameters (4)
  • c1
    Constant in the CoT lower bound C ≥ c1 n, imported from cited results; value unspecified and task-dependent.
  • tau
    Per-token inference time for intermediate reasoning tokens; hardware-dependent, not measured in this paper.
  • rho
    Latency per retrieval or verification call; hardware-dependent, cited from external benchmarks.
  • k
    Number of external retrieval calls required for authenticity; asserted to be at least 1 by Premise A2, not derived.
assumptions (3)
  • domain assumption Premise A1: any transformer solving length-n reasoning tasks requires Ω(n) chain-of-thought tokens.
    Rests on cited circuit-complexity lower bounds for constant-depth transformers; applied here to all LLMs and all reasoning tasks without stating the model class conditions.
  • ad hoc to paper Premise A2: authenticity on non-synthetic queries requires at least k ≥ 1 external retrieval or verification calls.
    Asserted without proof; not an established theorem and contradicted by closed-book QA systems; load-bearing for Lemma 2.
  • domain assumption Premise A3: every extra compute step, retrieval, or tool call consumes from a fixed inference budget.
    Plausible systems axiom, but the formal equation treats latency per token as linear in number of generated tokens, ignoring batching, caching, and parallel decoding.

how reviews work

0 comments
Cite this review

Pith. "Pith review of BAR Conjecture: the Feasibility of Inference Budget-Constrained LLM Services with Authenticity and Reasoning." pith.science (2026). https://pith.science/paper/NTOWV7QE

@misc{pith2026250723170,
  author       = {Pith},
  title        = {Pith review of: BAR Conjecture: the Feasibility of Inference Budget-Constrained LLM Services with Authenticity and Reasoning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/NTOWV7QE}},
  note         = {Machine review of arXiv:2507.23170}
}
read the original abstract

When designing LLM services, practitioners care about three key properties: inference-time budget, factual authenticity, and reasoning capacity. However, our analysis shows that no model can simultaneously optimize for all three. We formally prove this trade-off and propose a principled framework named The BAR Theorem for LLM-application design.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

28 extracted references · 1 canonical work pages

  1. [5]

    Edward J

    URL https://arxiv.org/abs/2502.18864. Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models,

    arXiv

  2. [6]

    Rania Iskander, Jaeyeon Hwang, Ankit Goyal, and Sharan Narang

    URL https: //arxiv.org/abs/2106.09685. Rania Iskander, Jaeyeon Hwang, Ankit Goyal, and Sharan Narang. The hidden cost of knowledge: Energy footprint of retrieval-augmented generation.arXiv preprint arXiv:2311.09765,

    arXiv

  3. [9]

    Junyi Li, Xiaoxue Cheng, Wayne Xin Zhao, Jian-Yun Nie, and Ji-Rong Wen

    URL https://arxiv.org/abs/2309.06180. Junyi Li, Xiaoxue Cheng, Wayne Xin Zhao, Jian-Yun Nie, and Ji-Rong Wen. Halueval: A large- scale hallucination evaluation benchmark for large language models, 2023a. URL https:// arxiv.org/abs/2305.11747. Xiaoyang Li, Zeyu Chen, Xinyun Fu, et al. Livecodebench: Evaluating code generation with live competitive programm...

    arXiv

  4. [10]

    Stephanie Lin, Jacob Hilton, and Owain Evans

    URL https://arxiv.org/abs/2402.12875. Stephanie Lin, Jacob Hilton, and Owain Evans. Truthfulqa: Measuring how models mimic human falsehoods,

    arXiv

  5. [12]

    William Merrill, Ashish Sabharwal, and Noah A

    URL https://arxiv.org/abs/2503.03961. William Merrill, Ashish Sabharwal, and Noah A. Smith. Saturated transformers are constant- depth threshold circuits. Transactions of the ACL , 10:843–858,

    arXiv

  6. [14]

    7 Alexander Novikov, Ngˆan V ˜u, Marvin Eisenberger, Emilien Dupont, Po-Sen Huang, Adam Zsolt Wagner, Sergey Shirobokov, Borislav Kozlovskii, Francisco J

    URL https://arxiv.org/ abs/2112.09332. 7 Alexander Novikov, Ngˆan V ˜u, Marvin Eisenberger, Emilien Dupont, Po-Sen Huang, Adam Zsolt Wagner, Sergey Shirobokov, Borislav Kozlovskii, Francisco J. R. Ruiz, Abbas Mehrabian, M. Pawan Kumar, Abigail See, Swarat Chaudhuri, George Holland, Alex Davies, Sebastian Nowozin, Pushmeet Kohli, and Matej Balog. Alphaevol...

    arXiv

  7. [15]

    URL https://arxiv.org/abs/2506.13131. Maxwell Nye, Anders Johan Andreassen, Guy Gur-Ari, Henryk Michalewski, Jacob Austin, David Bieber, David Dohan, Aitor Lewkowycz, Maarten Bosma, David Luan, Charles Sutton, and Au- gustus Odena. Show your work: Scratchpads for intermediate computation with language models, 2021a. URL https://arxiv.org/abs/2112.00114. M...

    arXiv

  8. [16]

    Bhrij Patel, Souradip Chakraborty, Wesley A

    URL https://arxiv.org/abs/2203.02155. Bhrij Patel, Souradip Chakraborty, Wesley A. Suttle, Mengdi Wang, Amrit Singh Bedi, and Dinesh Manocha. Aime: Ai system optimization via multiple llm evaluators,

    arXiv

Show all 28 references
  1. [17]

    Rewon Child Pope and Scott Gray

    URL https: //arxiv.org/abs/2410.03131. Rewon Child Pope and Scott Gray. Efficiently scaling transformer inference. In International Conference on Learning Representations (ICLR) ,

  2. [18]

    Samyam Rajbhandari et al

    URL https://arxiv.org/abs/2305.18290. Samyam Rajbhandari et al. Zero: Memory optimizations toward training trillion parameter models. In ACM/IEEE SC20: International Conference for High Performance Computing, Networking, Storage and Analysis,

    arXiv

  3. [19]

    Mrinank Sharma, Meg Tong, Tomasz Korbak, David Duvenaud, Amanda Askell, Samuel R

    URL https://arxiv.org/abs/2311.12022. Mrinank Sharma, Meg Tong, Tomasz Korbak, David Duvenaud, Amanda Askell, Samuel R. Bow- man, Newton Cheng, Esin Durmus, Zac Hatfield-Dodds, Scott R. Johnston, Shauna Kravec, Tim- othy Maxwell, Sam McCandlish, Kamal Ndousse, Oliver Rausch, N...

    arXiv

  4. [20]

    Noam Shazeer

    URL https://arxiv.org/abs/2310.13548. Noam Shazeer. Fast transformer decoding: One write-head is all you need. arXiv preprint arXiv:1911.02150,

    1911 arXiv

  5. [22]

    Megatron-lm: Training multi-billion parameter language models using gpu model parallelism

    Mohammad Shoeybi et al. Megatron-lm: Training multi-billion parameter language models using gpu model parallelism. arXiv preprint arXiv:1909.08053,

    1909 arXiv

  6. [24]

    Alon Talmor, Jonathan Herzig, Nicholas Lourie, and Jonathan Berant

    URL https://arxiv.org/abs/2308.03212. Alon Talmor, Jonathan Herzig, Nicholas Lourie, and Jonathan Berant. Commonsenseqa: A question answering challenge targeting commonsense knowledge. In Proceedings of NAACL-HLT,

    arXiv

  7. [25]

    Hierarchical reasoning model

    Guan Wang, Jin Li, Yuhao Sun, Xing Chen, Changling Liu, Yue Wu, Meng Lu, Sen Song, and Yasin Abbasi Yadkori. Hierarchical reasoning model. arXiv preprint arXiv:2506.21734,

    arXiv

  8. [27]

    Yiben Wang, Yao Fan, Zhilin Yang, Canwen Xu, Dong Yu, and Zhiyuan Liu

    URL https://arxiv.org/abs/2203.11171. Yiben Wang, Yao Fan, Zhilin Yang, Canwen Xu, Dong Yu, and Zhiyuan Liu. Humanity’s last exam: Evaluating large language models on complex reasoning. arXiv preprint arXiv:2401.10968 , 2024a. Zhongzhi Wang, Yaobo Zhang, Xiaoqian Wang, Shuli X...

    arXiv

  9. [28]

    Zhilin Yang, Peng Qi, Saizheng Zhang, Yoshua Bengio, William W

    URL https://arxiv.org/abs/2201.11903. Zhilin Yang, Peng Qi, Saizheng Zhang, Yoshua Bengio, William W. Cohen, Ruslan Salakhutdinov, and Christopher D. Manning. HotpotQA: A dataset for diverse, explainable multi-hop question answering. In Proceedings of the 2018 Conference on Em...

    2018 arXiv

  10. [30]

    Griffiths, Caiming Xiong, and Jingjing Liu

    Shunyu Yao, Dian Yu, Jeffrey Zhao, Izhak Shafran, Thomas L. Griffiths, Caiming Xiong, and Jingjing Liu. Large language models as analogical reasoners. arXiv preprint arXiv:2310.01714, 2023a. URL https://arxiv.org/abs/2310.01714. Shunyu Yao, Dian Yu, Jeffrey Zhao, Izhak Shafran...

    arXiv

  11. [32]

    URL https://arxiv.org/abs/2205.10625. 9

    arXiv

  12. [2017]

    Yannis Kalantidis, Matthijs Douze, and Herv´e Jegou

    URL https://arxiv.org/ abs/1705.03551. Yannis Kalantidis, Matthijs Douze, and Herv´e Jegou. Faiss on 20 million vectors: Gpu-efficient ann search for production rag. In Proceedings of the ACM SIGIR Conference ,

    arXiv

  13. [2018]

    doi: 10.18653/v1/D18-1259

    Association for Computational Linguistics. doi: 10.18653/v1/D18-1259. Haonan Yao, Ziyang Zhang, Yiming Wu, Xiang Zhang, Jialong Tang, and Xu Sun. Logicbench: A unified benchmark for evaluating natural language inference over logical reasoning patterns. arXiv preprint arXiv:2404.15522,

    arXiv

  14. [2019]

    Towards under- standing systems trade-offs in retrieval-augmented generation model inference

    Michael Shen, Muhammad Umar, Kiwan Maeng, G Edward Suh, and Udit Gupta. Towards under- standing systems trade-offs in retrieval-augmented generation model inference. arXiv preprint arXiv:2412.11854,

    arXiv

  15. [2020]

    Palm: Scaling language modeling with pathways

    Aakanksha Chowdhery et al. Palm: Scaling language modeling with pathways. arXiv preprint arXiv:2204.02311,

    arXiv

  16. [2021]

    org/abs/2110.14168

    URL https://arxiv. org/abs/2110.14168. Tri Dao, Daniel Y . Fu, Stefano Ermon, Atri Rudra, and Christopher R ´e. Flashattention: Fast and memory-efficient exact attention with io-awareness. In Advances in Neural Information Process- ing Systems (NeurIPS),

    arXiv

  17. [2022]

    William Merrill and Ashish Sabharwal

    URL https://arxiv.org/abs/2109.07958. William Merrill and Ashish Sabharwal. The expressive power of log-depth transformers. arXiv preprint arXiv:2503.03961,

    arXiv

  18. [2023]

    Shows retrieval can consume 30–60% of infer- ence energy

    URL https://arxiv.org/abs/2311.09765. Shows retrieval can consume 30–60% of infer- ence energy. Mandar Joshi, Eunsol Choi, Daniel S. Weld, and Luke Zettlemoyer. Triviaqa: A large scale distantly supervised challenge dataset for reading comprehension,

    arXiv

  19. [2024]

    Paul Gauthier

    URLhttps://arxiv.org/abs/ 2402.01306. Paul Gauthier. Aider polyglot benchmark for whole-file code editing. https://github.com/ paul-gauthier/aider,

    arXiv

  20. [2025]

    Albalak et al

    URL https://arxiv.org/abs/2501.12948. Albalak et al. Big-math: A large-scale, high-quality math dataset for reinforcement learning in language models

    arXiv

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.