Pith. sign in

REVIEW 4 major objections 4 minor 36 references

A committee of cheap models with a small LLM critic matches a much larger standalone reasoner on financial sentiment classification.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

A divergence-routed VADER+FinBERT+LLM committee reaches ~0.87 F1 with a 1.5B critic, matching 7B with far less cost, while same-size persona voting regresses to 0.66.

T0 review reviewed 2026-08-01 challenge →

load-bearing objection Plausible central claim, but a load-bearing FPB subset mislabeling keeps the headline numbers conditional; still deserves a serious referee. the 4 major comments →

arxiv 2607.19794 v1 pith:6I5Y2JTV submitted 2026-07-22 cs.CL cs.CEcs.DBcs.LG

TriAgent: Divergence-Aware Multi-Agent Committees for Cost-Efficient Financial Sentiment Analysis

Isabel Xu (The Overlake School) Cynthia Xu (The Overlake School) Rachel Ren (Edwards Vacuum Inc.) Cong Guo (The University of Memphis) Jiacheng Ding (The University of Memphis) This is my paper
classification cs.CL cs.CEcs.DBcs.LG
keywords financial sentiment analysismulti-agent committeesemantic divergence indexcost-efficient routingcritic plateaucross-lingual cachehallucination detectionLLM routing
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The reading

TriAgent proposes that a three-tier committee — a word-level sentiment lexicon, a sentence-level financial transformer, and a small reasoning LLM — can match or beat a much larger standalone language model at financial sentiment classification while cutting inference costs dramatically. The central empirical claim is the 'critic plateau': when the small LLM is re-tasked from answering directly to adjudicating disagreements between the two cheaper agents, F1 hovers around 0.87 for models from 1.5B to 7B parameters, while a same-size three-persona vote collapses to 0.66. The same divergence signal that gates routing also serves as a post-hoc hallucination detector and a cross-lingual cache key. If the claim holds, production systems can get large-model accuracy from small models plus a critic, reducing bills by orders of magnitude and making edge deployment viable.

Core claim

The paper's central empirical claim is that interaction substitutes for parameters within a model family: re-tasking a Qwen2.5 LLM (1.5B–7B) as a critic that reconciles VADER's word-level and FinBERT's sentence-level outputs yields F1 ≈ 0.87 on Financial PhraseBank, with bootstrap 95% confidence intervals overlapping across model sizes, whereas the same LLM used as one of three persona voters falls to F1 = 0.66. This plateau is attributed to granularity-stratified diversity: the three tiers fail on different instances (pairwise error-set Jaccard overlap only 0.13–0.15), so a critic that can read all three has enough signal to correct most errors even at 1.5B parameters. Three corollaries fol

What carries the argument

The Semantic Divergence Index (SDI): three pairwise absolute differences between the continuous polarity scores of the word-level lexicon (VADER), the sentence-level transformer (FinBERT), and the cross-sentence LLM reasoner. Thresholding the two key pairs (lexicon–specialist and specialist–reasoner) partitions queries into four quadrants (consensus, domain shift, ambiguous, mixed) that determine whether to trust the cheapest tier, escalate to the LLM, or call an interactive protocol (vote, critic, or debate). The critic protocol feeds the LLM the original sentence plus the two smaller agents' predictions and asks for a final label; this is the mechanism that produces the plateau.

Load-bearing premise

The whole argument rests on the benchmark being the 'sentences all agree' subset of Financial PhraseBank, but the reported sentence counts (4,838; 604/2,872/1,362) match the full or 50%-agreement set, not the standard all-agree subset; if the labels are noisier than claimed, every F1 number and the cost/F1 frontier shift.

What would settle it

Take the same TriAgent system and run it separately on the standard 'sentences all agree' Financial PhraseBank subset (about 2.3k sentences) and on the full 4,838-sentence set. If the critic plateau at F1≈0.87 persists on the clean all-agree subset, the mechanism is real; if it drops, the plateau was inflated by label noise. Also, replace FinBERT with a weaker sentence-level model and see if the plateau height falls; the paper's own running example suggests it should.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • Production sentiment pipelines can substitute a small critic LLM (1.5B) for a large one (7B) with no measured F1 loss, as long as the critic is supported by a word-level and sentence-level specialist pair.
  • Routing by the same divergence signal yields a continuous cost–accuracy frontier: at the 'Balanced' operating point on FPB, cost drops 48x versus always running the LLM, with only about 9pp F1 sacrifice.
  • The Shared Consensus Dictionary lets a new deployment inherit cached committee answers across languages: Chinese queries match English labels 95% of the time at F1 = 0.99, effectively a zero-cost cross-lingual canonicalization.
  • The divergence signal doubles as a post-hoc hallucination detector (AUC = 0.90), so systems already running a specialist and an LLM in parallel can add trust scoring without extra models.
  • The same SDI-routed signals produce better risk-adjusted returns in a 20-ticker backtest (Sharpe 3.50 vs. 1.36 for always-FinBERT and 0.11 for always-LLM), suggesting the quality signal transfers to downstream trading decisions.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The critic plateau may be a more general phenomenon: any task where a small model can adjudicate between two decorrelated cheap estimators might show the same saturation. A testable extension is to replicate the three-tier split on other domains (e.g., medical or legal text) and see whether the 1.5B plateau persists.
  • The paper's cross-lingual result likely underestimates production difficulty: the 1500-sentence translated set is a narrow domain, and the 0.99 F1 cache-hit accuracy may degrade on out-of-domain Chinese queries whose nearest English neighbor is not a true paraphrase.
  • The weakest link in the chain is the specialist, not the LLM: on the running example, the critic inherits the lexicon+specialist's false positive ('net profit halved' as positive) and stays wrong, suggesting plateau height is anchored to the smaller agents' error patterns. A system designer should therefore invest in specialist quality first.
  • If the FPB benchmark is actually the full 50%-agreement set rather than the 'sentences all agree' subset (the reported counts match the full set), all F1 numbers and the cost/F1 frontier are computed against noisier labels; the plateau could be partly an artifact of label noise.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 4 minor

Summary. The paper proposes TriAgent, a three-tier financial sentiment committee (VADER, FinBERT, a Qwen2.5 LLM) whose queries are gated by a Semantic Divergence Index (SDI) computed as pairwise absolute differences among the three agents' continuous polarity scores. The central claim is a 'critic plateau': when the LLM acts as a critic over VADER's and FinBERT's outputs, macro-F1 plateaus near 0.87 for Qwen 1.5B--7B, with overlapping bootstrap confidence intervals, whereas a same-size 3-persona vote achieves only 0.66. The same SDI signal is used for cost-aware routing, a Shared Consensus Dictionary for cross-lingual caching, a hallucination detector, and a backtested trading strategy. The paper reports a large cost saving at 10M-user scale and releases code and data.

Significance. If the central finding holds, it is an important and cost-relevant result: within a model family, interaction among small, granularity-stratified agents can substitute for raw parameter count, and an inexpensive lexicon+specialist+small-LLM committee can rival a much larger standalone reasoner. The paper has genuine strengths: the SDI signal is defined independently of the labels; the critic plateau is supported by bootstrap CIs; the same-size persona-vote ablation is a clean negative control; pairwise kappa and Jaccard error-overlap analyses make the diversity mechanism concrete; and all code, lexicons, and predictions are released. The main load-bearing concern is that the evaluation appears to use the full/noisy Financial PhraseBank while describing it as the 'sentences all agree' subset, and several thresholds and operating points are selected on the same test set. These issues are fixable but currently prevent the results from being taken at face value.

major comments (4)
  1. [§5 Setup] The setup states that the experiments use the 'sentences all agree' subset of Financial PhraseBank, but the reported counts (4,838 sentences; 604 negative / 2,872 neutral / 1,362 positive) match the full 50%-agreement FPB distribution, not the standard all-agree subset of about 2.3k sentences with a different class balance. Every headline F1 number — the critic plateau, the persona-vote F1=0.66, the Pareto frontier, and the cross-lingual SCD results — is therefore computed on noisier, majority-based labels than claimed. Since label noise can compress or shift F1 gaps between models, the central claim needs to be re-verified on the correct benchmark subset, or the benchmark description and all downstream comparisons need to be corrected.
  2. [§3.3, §5.3, Table 3] The routing thresholds (θ_LE=0.3, θ_ER=0.7), the SDI quadrant definitions, the SCD similarity threshold τ, and the three operating points in Table 3 (Budget/Balanced/Premium) are all chosen after inspecting the same test set. Figure 7 explicitly says the Pareto frontier is parameterized continuously by the SDI thresholds, and Table 3 selects points from that frontier. This is in-sample threshold selection. The paper should use a nested validation procedure or a separate tuning split; otherwise the reported F1 at the Balanced/Premium points and the associated cost/F1 frontier are optimistically biased. The same issue affects the backtest: the SDI single-stage and two-stage strategies appear to be selected from the same SDI sweep that defines their parameters, so the Sharpe=3.50 result may not reflect out-of-sample performance.
  3. [§5.2 and §1 running example] The critic plateau height is not independent of the smaller agents' errors. The running example in Section 1 shows that critic@1.5B and critic@7B are anchored on VADER+FinBERT's wrong positive label and stay wrong; only round-2 debate recovers. Thus the plateau reflects the critic's ability to arbitrate cases where V+F disagree, not its ability to overrule V+F when both are wrong. The paper does not report the joint distribution of critic errors conditional on V+F agreement/disagreement. A clearer decomposition (e.g., critic accuracy when V+F agree vs. disagree, and error overlap with the V+F pair) would strengthen the claim that the plateau is general and not an artifact of the particular scaffold.
  4. [§5.6 and Figure 10] The cross-lingual SCD result (95% of Chinese queries hit the English cache at F1=0.99) is obtained by translating 1,500 FPB sentences into Mandarin. Because these are parallel translations of the same English sentences, the Chinese queries are near-duplicates of the English cache entries under a multilingual sentence encoder, making a high hit rate and cached-label F1 somewhat trivial. This does not demonstrate generality to naturally occurring Chinese financial text. A test on an independent Chinese financial sentiment corpus is needed before claiming the SCD is a cross-lingual canonicalizer at 'zero marginal cost.'
minor comments (4)
  1. [Abstract] Typo: 'griven' should be 'driven'.
  2. [Table 2 caption] The caption describes quadrant behavior; consider defining how quadrant membership is assigned from SDI thresholds, since the thresholds are not given in that table and are central to interpreting the counts.
  3. [References] The FPB dataset reference should cite the exact repository version and the specific file used (e.g., 'sentences_allagree.txt' vs. 'sentences_50agree.txt'). This is related to the major benchmark-subset issue.
  4. [§7] The discussion honestly notes that the critic regresses on TFNS and that cross-family scaling is not universal. These limitations are useful but are presented as side notes; they should be integrated into the main results so readers do not overgeneralize the plateau claim.

Circularity Check

0 steps flagged

No circular derivation: SDI is defined independently of labels, the critic plateau is an empirical result rather than a fitted parameter, and there are no load-bearing self-citations.

full rationale

The central derivation chain is non-circular. SDI (Eq. 1) is defined as pairwise absolute differences of agent polarity scores, with no dependence on gold labels, fitted parameters, or the benchmark labels used for evaluation. The critic plateau is an observed empirical outcome of a protocol that feeds VADER and FinBERT predictions to the LLM; while the running example shows the critic can inherit V/F errors, nothing in the equations forces the LLM to copy either agent, so F1≈0.87 is a measurement rather than a construction. The paper contains no self-citations by the authors and invokes no uniqueness theorem; citations to prior work (query-by-committee, ensemble disagreement, LLM-as-judge) are explicit context, not load-bearing proofs. Corollary claims such as the SCD cross-lingual hit rate, the hallucination-detector AUC, and the backtest Sharpe are in-sample evaluations of the same SDI signal with thresholds selected on FPB; that is an over-claiming/correctness concern (as is the apparent mismatch between the claimed 'sentences all agree' FPB subset and the reported 4,838-sentence counts), but selecting an operating point from the same benchmark does not make the measurement equivalent to its input. The paper even discloses limitations: FPB is 'a news proxy; we do not claim real-market prediction' and adversarial detection is only partial. No circularity step can be exhibited, so the score is 0.

Axiom & Free-Parameter Ledger

4 free parameters · 4 axioms · 0 invented entities

Three thresholds/operating points and a backtest design are the main hand-set inputs. No new physical entities. The paper's claims are engineering results; the largest hidden input is the assumption that FPB F1 transfers to production and to tradable signals.

free parameters (4)
  • Routing thresholds θ_LE and θ_ER = θ_LE=0.3, θ_ER=0.7
    Used in Eq. 2/3 for quadrant partition and routing; no tuning procedure is given, and the Pareto/operating-point results depend on sweeping these values on FPB.
  • SDI_max edge-predictor target threshold = 0.7
    The XGBoost routing classifier is trained to predict SDI_max > 0.7; this threshold is hand-set and matches θ_ER.
  • SCD similarity threshold τ = 0.85 (FPB sweet spot), 0.70 (cross-lingual), 0.95/0.50 in sweep
    Cache hit rate and cached-label F1 are reported at operating points selected from the test set after sweeping τ, not from a held-out tuning split.
  • Backtest design parameters = 20 tickers, 20 sentences/week, 5-day hold, 10% capital, 10bps slippage
    Chosen by the authors; no sensitivity analysis or error bars are provided for Sharpe/return.
axioms (4)
  • domain assumption FPB labels are gold-standard ground truth and the 'sentences all agree' subset is the standard one
    All F1 claims are computed against these labels; the reported counts suggest the subset may be mislabeled (Section 5 Setup).
  • domain assumption Scalar polarity scores from VADER, FinBERT, and LLM are commensurable and their absolute differences are a valid disagreement signal
    SDI in Eq. 1 treats s_V, s_F, s_L on a common [-1,1] scale even though the models' outputs are calibrated differently.
  • domain assumption GPT-4o-mini/GPT-4 public API prices are the correct baseline for cost savings
    The $9.3M/yr saving uses list prices and assumes all queries would otherwise go to the cloud LLM; self-hosted alternatives are cheaper.
  • domain assumption News-sentence sentiment from FPB is a valid proxy for tradable signals
    The backtest uses 20 sampled FPB sentences per ticker per week; paper itself disclaims real-market prediction (Section 5.7).

reviewed 2026-08-01 · how reviews work

0 comments
Cite this review

Pith. "Pith review of TriAgent: Divergence-Aware Multi-Agent Committees for Cost-Efficient Financial Sentiment Analysis." pith.science (2026). https://pith.science/paper/6I5Y2JTV

@misc{pith2026260719794,
  author       = {Pith},
  title        = {Pith review of: TriAgent: Divergence-Aware Multi-Agent Committees for Cost-Efficient Financial Sentiment Analysis},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/6I5Y2JTV}},
  note         = {Machine review of arXiv:2607.19794}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Production LLM-based financial sentiment analysis faces a structural cost trap: most queries are trivially classifiable, yet expensive cloud reasoners process them all, and the bill scales linearly with user count. We present TriAgent, a multi-agent committee stratified by contextual granularity -- a word-level lexicon (VADER), a sentence-level domain transformer (FinBERT), and a cross-sentence reasoner (Qwen2.5, 0.5B-14B-4bit, with Mistral-7B and Phi-3.5-mini cross-family checks). A three-way Semantic Divergence Index (SDI) measures pairwise disagreement across granularities and routes each query accordingly. Our central finding is the critic plateau: when the LLM is re-tasked as a critic over the smaller agents' outputs, F1 plateaus at ~0.87 across 1.5B-7B Qwen (bootstrap 95% CIs overlap), while a same-size 3-persona vote drops to F1=0.66, which is driven by granularity-stratified diversity. Three corollaries follow from the same SDI signal: (i) a Shared Consensus Dictionary on multilingual sentence-BERT answers 95% of Chinese queries from an English cache at F1=0.99 -- cross-border canonicalization at zero marginal cost; (ii) SDI doubles as a post-hoc LLM-hallucination detector at AUC=0.90; (iii) the SDI single-stage strategy attains the best risk-adjusted return (Sharpe=3.50) on a 20-ticker back-test, dominating both always-FinBERT (1.36) and always-LLM (0.11). At 10M-user scale, TriAgent saves $9.3M/year vs. a GPT-4o-mini baseline. Code, lexicons, and the SCD are released.

Figures

Figures reproduced from arXiv: 2607.19794 by Isabel Xu (The Overlake School) Cynthia Xu (The Overlake School) Rachel Ren (Edwards Vacuum Inc.) Cong Guo (The University of Memphis) Jiacheng Ding (The University of Memphis).

Figure 1
Figure 1. Figure 1: Annual inference cost vs. user count (10 queries/user/day). [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: TriAgent system architecture. A query first [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Three-way SDI on FPB. Left: the three pairwise SDIs. Centre: four-quadrant scatter. Right: SDILE by gold class — negative has the highest mean disagreement. VADER FinBERT Qwen-7B VADER FinBERT Qwen-7B 1.00 0.27 0.19 0.27 1.00 0.61 0.19 0.61 1.00 (a) Pairwise Cohen's κ (lower = more diverse) 0.0 0.5 1.0 1.5 Per-sentence disagreement entropy (bits) 0 500 1000 1500 2000 2500 Sentence count all-agree (0 bits) … view at source ↗
Figure 4
Figure 4. Figure 4: Pairwise Cohen’s κ (left) and per-sample disagreement entropy (right). Low κ; bimodal entropy shows the committee is either fully aligned or split 2-vs-1. 0.5B 1.5B 3.0B 7.0B Qwen parameter count (log scale) 0.4 0.5 0.6 0.7 0.8 0.9 F1-Macro on FPB gap that scaling cannot close (7B still −7pp vs. FinBERT) Scaling Qwen alone cannot beat the specialist 3B dip (3B < 1.5B) FinBERT (specialist, F1=0.88) VADER (l… view at source ↗
Figure 5
Figure 5. Figure 5: Scaling Qwen alone cannot beat the specialist: 7 B is [PITH_FULL_IMAGE:figures/full_fig_p006_5.png] view at source ↗
Figure 9
Figure 9. Figure 9: SCD accuracy / hit-rate trade-off as τ varies. Deployable sweet spot at τ ≈0.85. 1.5B 3B 7B Qwen-N reasoning tier 0.4 0.5 0.6 0.7 0.8 0.9 1.0 rate / F1 Naive cross-lingual committee EN ↔ ZH committee agreement EN committee F1 vs gold ZH committee F1 vs gold 0.50 0.55 0.60 0.65 0.70 0.75 0.80 0.85 Similarity threshold τ 0.5 0.6 0.7 0.8 0.9 1.0 rate / F1 SCD as cross-lingual canonicalizer ZH→EN cache hit rat… view at source ↗
Figure 10
Figure 10. Figure 10: Cross-lingual deployment. Left: naive committees agree 81–85% with their English counterparts. Right: the SCD as cross￾lingual canonicaliser — Chinese queries find English cached an￾swers at hit-rate 0.95 / F1 0.99. [Chen et al., 2023] Lingjiao Chen, Matei Zaharia, and James Zou. FrugalGPT: How to use large language mod￾els while reducing cost and improving performance. arXiv:2305.05176, 2023. [Dettmers e… view at source ↗
Figure 8
Figure 8. Figure 8: Per-class F1 sorted by F1 on the (hard) negative class. [PITH_FULL_IMAGE:figures/full_fig_p007_8.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

36 extracted references · 10 linked inside Pith

  1. [1]

    FinBERT: Financial sen- timent analysis with pre-trained language models

    [Araci, 2019] Dogu Araci. FinBERT: Financial sen- timent analysis with pre-trained language models. arXiv:1908.10063,

    Pith/arXiv arXiv 2019

  2. [6]

    MetaGPT: Meta programming for a multi-agent collaborative framework

    [Hong and others, 2024] Sirui Hong et al. MetaGPT: Meta programming for a multi-agent collaborative framework. InThe Twelfth International Conference on Learning Rep- resentations (ICLR),

    2024

  3. [7]

    [Hutto and Gilbert, 2014] C. J. Hutto and Eric Gilbert. V ADER: A parsimonious rule-based model for sentiment analysis of social media text. InICWSM,

    2014

  4. [9]

    Jiang et al

    [Jiang and others, 2023] Albert Q. Jiang et al. Mistral 7B. arXiv:2310.06825,

    Pith/arXiv arXiv 2023

  5. [10]

    Is BERT really robust? a strong baseline for natural language attack on text classifi- cation and entailment

    [Jin and others, 2020] Di Jin et al. Is BERT really robust? a strong baseline for natural language attack on text classifi- cation and entailment. InAAAI,

    2020

  6. [14]

    CAMEL: Commu- nicative agents for “mind” exploration of large language model society

    [Li and others, 2023] Guohao Li et al. CAMEL: Commu- nicative agents for “mind” exploration of large language model society. InAdvances in Neural Information Pro- cessing Systems (NeurIPS),

    2023

  7. [15]

    Encouraging di- vergent thinking in large language models through multi- agent debate

    [Liang and others, 2024] Tian Liang et al. Encouraging di- vergent thinking in large language models through multi- agent debate. InProceedings of the 2024 Conference on Empirical Methods in Natural Language Processing (EMNLP),

    2024

  8. [16]

    When is a liability not a liability? Textual analysis, dictionaries, and 10-Ks.The Journal of Finance, 66(1):35–65,

    [Loughran and McDonald, 2011] Tim Loughran and Bill McDonald. When is a liability not a liability? Textual analysis, dictionaries, and 10-Ks.The Journal of Finance, 66(1):35–65,

    2011

  9. [18]

    Gales, Mark J.˙ SelfCheckGPT: Zero-resource black-box hallucination detection for generative large lan- guage models

    [Manakulet al., 2023 ] Potsawee Manakul, Adian Liusie, and F. Gales, Mark J.˙ SelfCheckGPT: Zero-resource black-box hallucination detection for generative large lan- guage models. InProceedings of the 2023 Conference on Empirical Methods in Natural Language Processing (EMNLP),

    2023

  10. [19]

    RouteLLM: Learn- ing to route llms with preference data

    [Ong and others, 2024] Isaac Ong et al. RouteLLM: Learn- ing to route llms with preference data. arXiv:2406.18665,

    Pith/arXiv arXiv 2024

  11. [20]

    Ig- nore previous prompt: Attack techniques for language models

    [Perez and Ribeiro, 2022] F´abio Perez and Ian Ribeiro. Ig- nore previous prompt: Attack techniques for language models. arXiv:2211.09527,

    Pith/arXiv arXiv 2022

  12. [21]

    Sentence-BERT: Sentence embeddings using siamese BERT-networks

    [Reimers and Gurevych, 2019] Nils Reimers and Iryna Gurevych. Sentence-BERT: Sentence embeddings using siamese BERT-networks. InEMNLP,

    2019

  13. [22]

    Making monolingual sentence embeddings multilingual using knowledge distillation

    [Reimers and Gurevych, 2020] Nils Reimers and Iryna Gurevych. Making monolingual sentence embeddings multilingual using knowledge distillation. InEMNLP,

    2020

  14. [23]

    Active learning literature sur- vey

    [Settles, 2009] Burr Settles. Active learning literature sur- vey. Technical Report 1648, University of Wisconsin– Madison,

    2009

  15. [25]

    Qwen2.5 technical report

    [Team, 2024] Qwen Team. Qwen2.5 technical report. arXiv:2412.15115,

    Pith/arXiv arXiv 2024

  16. [26]

    https://huggingface.co/datasets/zeroshot/ twitter-financial-news-sentiment,

    [tfn, 2022] Twitter financial news sentiment dataset. https://huggingface.co/datasets/zeroshot/ twitter-financial-news-sentiment,

    2022

  17. [27]

    Mixture of cheap and expensive models for cost-effective inference

    [Wang and others, 2022] Hanrui Wang et al. Mixture of cheap and expensive models for cost-effective inference. InNeurIPS Workshop on Efficient Natural Language and Speech Processing,

    2022

  18. [28]

    Self- consistency improves chain of thought reasoning in language models

    [Wang and others, 2023] Xuezhi Wang et al. Self- consistency improves chain of thought reasoning in language models. InThe Eleventh International Confer- ence on Learning Representations (ICLR),

    2023

  19. [29]

    HuggingFace’s transformers: State-of-the-art natural language process- ing

    [Wolf and others, 2020] Thomas Wolf et al. HuggingFace’s transformers: State-of-the-art natural language process- ing. InProceedings of the 2020 Conference on Empir- ical Methods in Natural Language Processing: System Demonstrations,

    2020

  20. [30]

    BloombergGPT: A large language model for finance

    [Wu and others, 2023] Shijie Wu et al. BloombergGPT: A large language model for finance. arXiv:2303.17564,

    Pith/arXiv arXiv 2023

  21. [31]

    AutoGen: En- abling next-gen LLM applications via multi-agent conver- sation framework

    [Wu and others, 2024] Qingyun Wu et al. AutoGen: En- abling next-gen LLM applications via multi-agent conver- sation framework. InFirst Conference on Language Mod- eling (COLM),

    2024

  22. [32]

    PIXIU: A large language model, instruction data and evaluation bench- mark for finance

    [Xie and others, 2023] Qianqian Xie et al. PIXIU: A large language model, instruction data and evaluation bench- mark for finance. InAdvances in Neural Information Processing Systems (NeurIPS), Datasets and Benchmarks Track,

    2023

  23. [33]

    FinBERT-tone- Chinese: Domain-specific BERT for chinese finan- cial sentiment

    [Yang and others, 2021] Yi Yang et al. FinBERT-tone- Chinese: Domain-specific BERT for chinese finan- cial sentiment. https://huggingface.co/yiyanghkust/ finbert-tone-chinese,

    2021

  24. [34]

    FinBERT: A pretrained language model for financial communications

    [Yanget al., 2020 ] Yi Yang, Mark Christopher Siy Uy, and Allen Huang. FinBERT: A pretrained language model for financial communications. arXiv:2006.08097,

    Pith/arXiv arXiv 2020

  25. [35]

    FinGPT: Open-source financial large language models

    [Yanget al., 2023 ] Hongyang Yang, Xiao-Yang Liu, and Christina Dan Wang. FinGPT: Open-source financial large language models. arXiv:2306.06031,

    arXiv 2023

  26. [36]

    Judging LLM-as-a-Judge with MT-Bench and Chatbot Arena

    [Zheng and others, 2023] Lianmin Zheng et al. Judging LLM-as-a-Judge with MT-Bench and Chatbot Arena. InAdvances in Neural Information Processing Systems (NeurIPS), Datasets and Benchmarks Track, 2023

    2023

  27. [2009]

    Large language model routing with benchmark datasets

    [Shnitzer and others, 2023] Tal Shnitzer et al. Large language model routing with benchmark datasets. arXiv:2309.15789,

    Pith/arXiv arXiv 2023

  28. [2011]

    Good debt or bad debt: Detecting semantic orientations in economic texts

    [Maloet al., 2014 ] Pekka Malo, Ankur Sinha, Pekka Korho- nen, Jyrki Wallenius, and Pyry Takala. Good debt or bad debt: Detecting semantic orientations in economic texts. InJASIST,

    2014

  29. [2014]

    Survey of hallucination in natural language generation.ACM Computing Surveys, 55(12):1–38,

    [Ji and others, 2023] Ziwei Ji et al. Survey of hallucination in natural language generation.ACM Computing Surveys, 55(12):1–38,

    2023

  30. [2017]

    A survey of large lan- guage models in finance (FinLLMs)

    [Lee and others, 2024] Jean Lee et al. A survey of large lan- guage models in finance (FinLLMs). arXiv:2402.02315,

    Pith/arXiv arXiv 2024

  31. [2019]

    Deployable sweet spot atτ≈0.85

    0.5 0.6 0.7 0.8 0.9 Similarity threshold τ 0.60 0.65 0.70 0.75 0.80 0.85 0.90 0.95F1-Macro on test set Shared Consensus Dictionary — cost / accuracy trade-off SCD hybrid F1 (cache + miss→ committee) Always-committee F1 = 0.876 Cache hit rate 0.0 0.2 0.4 0.6 0.8 1.0 Cache hit rate Figure 9: SCD accuracy / hit-rate trade-off asτvaries. Deployable sweet spot...

    Pith/arXiv arXiv 2023

  32. [2020]

    Billion-scale similarity search with GPUs

    [Johnsonet al., 2021 ] Jeff Johnson, Matthijs Douze, and Herv´e J ´egou. Billion-scale similarity search with GPUs. InIEEE Trans. Big Data,

    2021

  33. [2021]

    Simple and scal- able predictive uncertainty estimation using deep ensem- bles

    [Lakshminarayananet al., 2017 ] Balaji Lakshminarayanan, Alexander Pritzel, and Charles Blundell. Simple and scal- able predictive uncertainty estimation using deep ensem- bles. InNeurIPS,

    2017

  34. [2022]

    Tenenbaum, and Igor Mordatch

    [Duet al., 2024 ] Yilun Du, Shuang Li, Antonio Torralba, Joshua B. Tenenbaum, and Igor Mordatch. Improving fac- tuality and reasoning in language models through multia- gent debate. InProceedings of the 41st International Con- ference on Machine Learning (ICML),

    2024

  35. [2023]

    LLM.int8(): 8-bit matrix multiplication for transformers at scale

    [Dettmerset al., 2022 ] Tim Dettmers, Mike Lewis, Younes Belkada, and Luke Zettlemoyer. LLM.int8(): 8-bit matrix multiplication for transformers at scale. InAdvances in Neural Information Processing Systems (NeurIPS),

    2022

  36. [2024]

    https://huggingface.co/datasets/ChanceFocus/ flare-fpb,

    [fla, 2023] FLARE-FPB: Financial phrasebank parquet mirror. https://huggingface.co/datasets/ChanceFocus/ flare-fpb,

    2023

This paper was first reviewed by deepseek-v4-flash on August 1, 2026.