Pith. sign in

REVIEW 2 major objections 4 minor 29 cited by

The paper argues that 'typicality bias' in preference data — annotators favoring familiar text — drives mode collapse in aligned LLMs, and that prompting for a verbalized probability distribution instead of a single answer restores the lost

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 ↗

T0 review · deepseek-v4-flash

2026-08-04 12:58 UTC pith:Y57MRET3

load-bearing objection The HelpSteer typicality-bias result is the paper's real contribution; the VS mechanism is not yet established, but the method works and the paper deserves serious review. the 2 major comments →

arxiv 2510.01171 v4 pith:Y57MRET3 submitted 2025-10-01 cs.CL cs.AI

Verbalized Sampling: How to Mitigate Mode Collapse and Unlock LLM Diversity

Jiayi Zhang , Simon Yu , Derek Chong , Anthony Sicilia , Michael R. Tomz , Christopher D. Manning , Weiyan Shi This is my paper
classification cs.CL cs.AI
keywords typicality biasmode collapseRLHFoutput diversitypreference dataverbalized samplingprompt engineeringalignment
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 paper tries to establish that mode collapse in post-alignment LLMs is not merely an optimization artifact but a property of the preference data itself: annotators systematically favor typical, fluent, familiar text, and this 'typicality bias' enters the reward signal. Formally, with reward r = r_true + α log π_ref + ε and α > 0, the optimal aligned policy is π* ∝ π_ref^γ exp(r_true/β) with γ = 1 + α/β > 1 — a strictly sharpened distribution, so even a perfect reward model and optimizer yield a model that over-favors stereotypical outputs. The paper verifies this bias on preference datasets (α̂ ≈ 0.57, p < 10⁻¹⁴) and then proposes Verbalized Sampling (VS): instead of asking for one response, prompt the model to list several responses with associated probabilities. VS roughly doubles creative-writing diversity, matches pretraining distributions much better on enumerative questions, and improves dialogue simulation and synthetic data without harming factuality or safety. A reader should care because the paper converts a training-time ailment into a tunable inference-time prompt, backed by a mechanism one can inspect.

Core claim

The core claim is an identity: with reward r = r_true + α log π_ref + ε, α > 0, the optimal RLHF policy is π* ∝ π_ref^γ exp(r_true/β), γ = 1 + α/β > 1. Any positive typicality weight sharpens the reference distribution, compressing mass onto its modes — mode collapse even with perfect rewards. The paper verifies the bias (α̂ ≈ 0.57, p < 10⁻¹⁴; above-chance across four datasets) and shows the three prompt types collapse to different objects: instance prompts return π_ref's mode, list prompts a uniform list, distribution prompts return an approximation of π_ref. VS nearly doubles creative-writing diversity, matches pretraining distributions (KL = 0.12), and improves dialogue simulation and syn

What carries the argument

Two pieces carry the argument. The typicality-bias reward model r(x,y) = r_true(x,y) + α log π_ref(y|x) + ε(x) uses the base model's own log-likelihood as a proxy for how typical a response is; plugging it into the closed-form KL-regularized RLHF solution gives the sharpening identity π*(y|x) ∝ π_ref(y|x)^γ exp(r_true/β), γ = 1 + α/β > 1. The second piece is the prompt-mode trichotomy: an instance-level prompt returns the single mode of π_ref, a list-level prompt returns a uniform distribution, and a distribution-level prompt — Verbalized Sampling — can return π_ref itself, because verbalized probabilities can encode it. Practically, VS is a prompt template asking for k responses with numeri

Load-bearing premise

The remedy rests on the premise that a mode-collapsed model, asked to list candidate responses with probabilities, will produce distinct responses whose verbalized probabilities approximate the pretraining distribution π_ref; the proof of that step builds the premise in by construction, and its direct empirical support is a single enumerative task judged against an acknowledged proxy for the pretraining corpus.

What would settle it

Take an open model whose pretraining corpus is known, prompt it with Verbalized Sampling on several enumerative questions (US states, countries, animals), and compare the verbalized probabilities against true pretraining frequencies and against the aligned model's own next-token probabilities. If the verbalized probabilities track surface plausibility instead of the reference distribution — or if the published US-states match fails to generalize to other enumerative domains — the diversity-recovery claim is falsified.

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

If this is right

  • Even with a perfect reward model and perfect optimization, any positive typicality bias in preference data sharpens the aligned distribution — so mode collapse cannot be fully eliminated by better algorithms on the same data.
  • Aligned models retain most of the base model's diversity underneath the collapse; it can be recovered by a training-free prompt, without sacrificing factual accuracy or safety.
  • Verbalized Sampling gives a continuous diversity dial: lowering the requested probability threshold in the prompt increases output diversity across tasks and models.
  • Diversity gains flow downstream: synthetic data produced by VS improves fine-tuned math performance and offline-RL training over direct-prompted data.
  • Larger models benefit more from VS, so the method's advantage should widen as models scale.

Where Pith is reading between the lines

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

  • Editorial inference: if verbalized probabilities really track π_ref, Verbalized Sampling doubles as a probe of closed models' pretraining distributions — recoverable even where training data is unpublished.
  • Editorial inference: the theory implies a training-time intervention the paper does not run — curating or re-weighting preference data to neutralize typicality (pushing α toward zero) should shrink mode collapse at its source, not just at decode time.
  • Editorial inference: since the sharpening identity follows from the α log π_ref term alone, any reference-distribution-regularized preference objective should exhibit the same collapse; the paper demonstrates the mechanism for RLHF-style training but the data-level cause is objective-independent.
  • Editorial inference: the claim that list-level prompts collapse to uniform distributions yields a cheap diagnostic — a peaked (non-uniform) list output measures how much residual mode collapse remains.

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

2 major / 4 minor

Summary. The paper argues that post-training alignment reduces LLM diversity because preference data contain a "typicality bias": annotators favor text that is familiar/typical, modeled as an additive log π_ref term in the reward. From this decomposition, the authors derive that the RLHF/DPO optimal policy is proportional to π_ref^γ with γ > 1 (Eq. 3), i.e., a sharper, more mode-collapsed distribution. They verify the bias empirically on HelpSteer using correctness-matched pairs, then propose Verbalized Sampling (VS), a prompting method that asks the model to output a list of responses with verbalized probabilities. They report large diversity gains over direct prompting across creative writing, dialogue simulation, open-ended QA, and synthetic data generation, and they claim that VS recovers the pretraining distribution, with a US-state experiment (KL = 0.12 vs. a RedPajama proxy) as direct evidence.

Significance. If the mechanism holds, the paper offers a genuinely novel data-centric explanation of mode collapse and a simple, training-free remedy. The HelpSteer verification is a particular strength: it uses 6,874 within-prompt pairs matched on correctness, cluster-robust standard errors, and estimates α̂ = 0.57–0.65 with p < 10⁻¹⁴, and the effect survives including correctness as a covariate. Eq. 3 is algebraically correct conditional on the stated reward decomposition. The paper also ships detailed prompts, hyperparameters, and code, which materially aids reproducibility. The empirical breadth—from creative writing to math data generation—is impressive. The main weakness is that the paper's central mechanistic claim, that VS recovers π_ref, rests on a proof in §E.4 that assumes the conclusion and on a single enumerative proxy-corpus experiment.

major comments (2)
  1. [§E.4, Claim 3] The proof of Claim 3 is circular in a load-bearing way. The construction sets p_i := π_ref(y_i | x_orig) and k = m, then concludes P(Y=y) = π_ref(y | x_orig). That is precisely the property needing demonstration. The appended Remark concedes this, saying the argument assumes π* is accurate enough to encode π_ref's distributional information. Since this premise is the bridge between the mode-collapse theory and the proposed fix, it cannot be left as an assumption. The only direct empirical support (§G.9) is the US-states enumerative task, which has a bounded answer space and is measured against a proxy corpus. Please either supply a direct validation of the premise on open-ended tasks (e.g., compare VS's verbalized distribution against the base model's own sample distribution on creative-writing prompts) or substantially weaken the wording that VS 'recovers' π_ref.
  2. [§G.9 / §4.1] The claimed alignment with the pretraining distribution (KL = 0.12 in Figure 3) is measured against RedPajama, not against the actual pretraining corpora of the closed API models (GPT-4.1, Claude-4-Sonnet). RedPajama is an acknowledged proxy, and the paper itself says so in §G.9, but the main text and Figure 3 repeatedly call it the 'pretraining distribution' without the same hedge. Since state-name frequency in a web corpus need not equal any particular model's true pretraining distribution, this experiment cannot bear the weight it is given. I recommend re-running the comparison on an open model whose pretraining data are known, or at minimum adding multiple corpora and consistently hedging the terminology.
minor comments (4)
  1. [Table 30] In the GPT-4.1-mini block, the row labeled ',→Combined' shows '0.153(↑0.43)0.266'; the 0.43 appears to be a typo for 0.043, and the 'Combined' name is inconsistent with the VS-CoT/VS-Multi labels used elsewhere.
  2. [§H.4 vs. §4.2] The tuning experiment says 'each word should be < p%' but the method description in the main text and Figure 2 uses 'probability of each response'. Please unify the wording.
  3. [Figure 4(e-f) caption] The caption reads 'Emergent Trend: in Diversity' and 'Cognitive Burden: in Quality' with blank spaces; presumably 'improvement' and 'change' were intended.
  4. [§G.9] The sentence 'Although it is a proxy, we refer to this distribution as ground-truth one in the following description' is confusing; please rephrase to avoid calling a proxy ground truth.

Axiom & Free-Parameter Ledger

3 free parameters · 7 axioms · 0 invented entities

The central claim rests on (a) an assumed reward decomposition with a fitted α, (b) the operationalization of typicality as base-model log-likelihood, (c) the standard RLHF closed form, (d) a flat-reward subset S, (e) an all-Y collapse simplification for the prompt-capability proofs, and (f) an acknowledged proxy pretraining corpus. No new physical or mathematical entities are introduced; 'typicality bias' is a named behavioral construct operationalized through existing quantities, and 'Verbalized Sampling' is a prompting method. The heaviest unvalidated load is item (e) combined with (a): the pervasiveness of α > 0 beyond HelpSteer is established only with a weaker metric.

free parameters (3)
  • alpha (typicality bias weight) = 0.57 ± 0.07 (Llama-3.1-405B-Base); 0.65 ± 0.07 (GLM-4.5-Base); 0.456–0.489 adjusted on HelpSteer
    Fitted via Bradley-Terry logistic regression on correctness-matched HelpSteer pairs (§E.1); positive α drives the entire sharpening argument (Eq. 3) and is the empirical 'verification' of the hypothesis.
  • k (number of candidates per VS call) = 5 (default; 1–20 ablated)
    Chosen by hand across experiments (§5, §H.1); affects VS performance but not the causal claim about typicality bias.
  • p (probability threshold for diversity tuning) = 1.0 (default); 0.001–0.9 in ablations
    Experimental control knob for VS diversity tuning (§H.4); hand-chosen, not fitted.
axioms (7)
  • domain assumption log π_ref(y|x) is a valid proxy for human typicality/familiarity
    Eq. 1; the entire empirical identification of typicality bias uses base-model log-likelihood as the operationalization. If base-model likelihood tracks quality/fluency rather than familiarity, the cognitive interpretation changes, though the sharpening mechanism survives.
  • standard math Bradley-Terry model and the KL-regularized RLHF closed form (Rafailov et al. 2024)
    Eqs. 2 and 4; standard results cited from the literature; the closed form is parameter-free with stated assumptions that do not include the target claim.
  • domain assumption r_true is (approximately) flat on the response set S
    §3.2; required to reduce π* to π_ref^γ on S and thereby show mode collapse; the paper states this can be relaxed to approximate flatness but does not prove the relaxed version.
  • ad hoc to paper Mode collapse holds on all of Y (stronger collapse for prompt analysis)
    §E.4; the proofs of Claims 2–3 assume π* is collapsed on all of Y, not just a subset S; justified only by observable sampling behavior.
  • domain assumption The collapsed model can still produce a diverse within-call list with faithful probabilities
    §4.1 Claims 2–3; the mode of a list prompt is assumed to be a list of k distinct items, and the mode of a distribution prompt a list with probabilities ≈ π_ref; directly tested only on the US-states task (§G.9).
  • domain assumption RedPajama frequency is a proxy for the pretraining distribution of API models (GPT-4.1, Claude-4-Sonnet)
    §G.9; the KL = 0.12 comparison is relative to this proxy, which the authors acknowledge is not the models' actual pretraining corpus.
  • domain assumption Cognitive psychology findings (mere exposure, fluency, schema congruity) transfer to LLM preference annotators
    §3.1; background motivation for why annotators might prefer familiar text; not independently validated in this paper.

pith-pipeline@v1.3.0-alltime-deepseek · 5453 in / 5541 out tokens · 293237 ms · 2026-08-04T12:58:38.480673+00:00 · methodology

0 comments
read the original abstract

Post-training alignment often reduces LLM diversity, leading to a phenomenon known as mode collapse. Unlike prior work that attributes this effect to algorithmic limitations, we identify a fundamental, pervasive data-level driver: typicality bias in preference data, whereby annotators systematically favor familiar text as a result of well-established findings in cognitive psychology. We formalize this bias theoretically, verify it on preference datasets empirically, and show that it plays a central role in mode collapse. Motivated by this analysis, we introduce Verbalized Sampling, a simple, training-free prompting strategy to circumvent mode collapse. VS prompts the model to verbalize a probability distribution over a set of responses (e.g., "Generate 5 jokes about coffee and their corresponding probabilities"). Comprehensive experiments show that VS significantly improves performance across creative writing (poems, stories, jokes), dialogue simulation, open-ended QA, and synthetic data generation, without sacrificing factual accuracy and safety. For instance, in creative writing, VS increases diversity by 1.6-2.1x over direct prompting. We further observe an emergent trend that more capable models benefit more from VS. In sum, our work provides a new data-centric perspective on mode collapse and a practical inference-time remedy that helps unlock pre-trained generative diversity.

Figures

Figures reproduced from arXiv: 2510.01171 by Anthony Sicilia, Christopher D. Manning, Derek Chong, Jiayi Zhang, Michael R. Tomz, Simon Yu, Weiyan Shi.

Figure 1
Figure 1. Figure 1: We show that typicality bias in preference data is a fundamental and pervasive cause of mode collapse, reducing output diversity. As a solution, we propose Verbalized Sampling (VS), a principled prompting method that returns distributions of responses, to improve diversity. ∗Equal contribution. See the contribution statement for further details. 1 arXiv:2510.01171v3 [cs.CL] 10 Oct 2025 [PITH_FULL_IMAGE:fi… view at source ↗
Figure 2
Figure 2. Figure 2: Ready-to-use Verbalized Sampling (VS) Prompt. See §I.2 for more variants and detail. As typicality bias is pervasive across human preference data, we look for solutions beyond the training process. Grounded in our theoreti￾cal insights, we propose a simple but principled prompting method to bypass mode collapse. As shown in [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Qualitative and quantitative examples on different tasks. For [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: a-c: Average semantic diversity scores (%) in poem (a), story (b) and joke (c) across methods and models. Our methods consistently outperform the baselines. We performed a one￾tailed t-test between VS-Standard and the baselines (* p < 0.05, ** p < 0.01, *** p < 0.001). d: Diversity vs. Quality trade-off for the poem task, where VS-Multi and VS-CoT approach the Pareto front. e-f: Emergent Trend where larger… view at source ↗
Figure 5
Figure 5. Figure 5: Image diversity using captions generated by different methods. We use different methods to generate descriptive captions given the topic, and then visualize these captions with images. Direct Prompting (top row) consistently converges on captions that will produce photorealistic images within a narrow range of scenarios, typically landscapes like deserts. In contrast, our Verbalized Sampling method (bottom… view at source ↗
Figure 6
Figure 6. Figure 6: Ablation study on temperature for poem generation across GPT-4.1 and Gemini-2.5- Flash models. We set k = 5 across experiments. Each plot shows the diversity-quality trade-off for three methods (Direct, Sequence, VS-Standard) at different temperature values (t). VS-Standard can be combined with temperature to further improve the trade-off, consistently outperforming baselines across both models. experience… view at source ↗
Figure 7
Figure 7. Figure 7: Diversity scores across post-training stages of Tulu-70B. “Tulu-Final-70B” is the model after RLVR. The red dashed line indicates the base model’s diversity level (45.4%). Baseline prompting methods experience major diversity drops (mode collapse) after SFT and DPO, with direct prompting showing the most severe drop. In contrast, VS maintains a higher diversity scores throughout all training stages, demons… view at source ↗
Figure 8
Figure 8. Figure 8: VS performance in Persuasive Dialogue Simulation. (a) Donation Amount Distri￾butions simulated by small, large, and reasoning models with direct and VS, compared against fine-tuned model (green) and human (blue). We see that VS simulates donation distributions more similar to human, especially for the larger and reasoning-focused models. (b) Linguistic Alignment on Distinct-1/2/3, semantic diversity, and r… view at source ↗
Figure 9
Figure 9. Figure 9: Results on the Open-Ended QA task averaged across models. We perform one-tailed t-test between VS-Standard and baselines (*p < 0.05, **p < 0.01, ***p < 0.001). (a) shows the average KL divergence between the response distribution and the corresponding pretraining distribution. VS achieves lower KL divergence compared to baseline methods, indicating closer alignment with the pretraining distribution. (b) sh… view at source ↗
Figure 10
Figure 10. Figure 10: Typicality bias rate across different preference datasets and base models. Typicality bias rate measures how often the human-preferred response in a preference pair is assigned a higher likelihood by a base model. All models show a systematic, above-chance bias (agreement >50%), with larger models generally exhibiting a stronger effect. We also show the 95% confidence intervals. The consistent above-chanc… view at source ↗
Figure 11
Figure 11. Figure 11: Semantic diversity (%) and quality scores on the [PITH_FULL_IMAGE:figures/full_fig_p042_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Semantic diversity (%) and quality scores on the [PITH_FULL_IMAGE:figures/full_fig_p044_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: Semantic diversity (%) and quality scores on the [PITH_FULL_IMAGE:figures/full_fig_p046_13.png] view at source ↗
Figure 14
Figure 14. Figure 14: Example interfaces of the Prolific human study for poem (top) and story (bottom). [PITH_FULL_IMAGE:figures/full_fig_p049_14.png] view at source ↗
Figure 15
Figure 15. Figure 15: Dice roll distributions from direct, se [PITH_FULL_IMAGE:figures/full_fig_p052_15.png] view at source ↗
Figure 16
Figure 16. Figure 16: Average diversity and quality results with GPT-4.1 on the [PITH_FULL_IMAGE:figures/full_fig_p054_16.png] view at source ↗
Figure 17
Figure 17. Figure 17: Comparison of model output distributions with the ground-truth distribution. Figure 17a Claude-4-Sonnet and Figure 17b GPT-4.1 results show that Direct prompting (blue) concentrates probability on few states, while Sequence prompting yields a uniform distribution (dashed line), missing the ground truth’s sharp peaks. VS-Standard (red) best matches the ground￾truth distribution (yellow) by preserving peake… view at source ↗
Figure 18
Figure 18. Figure 18: Analysis of the number of candidates (k) for poem generation across GPT-4.1 and Gemini-2.5-Flash. Each plot illustrates the diversity-quality trade-off as k is varied from 1 to 20. Increasing k generally improves diversity but lowers quality. VS-Standard consistently provides the best trade-off compared to the two baseline, approaching the Pareto front. We analyze the impact of the number of candidates (k… view at source ↗
Figure 19
Figure 19. Figure 19: Top-p sampling analysis for poem generation across GPT-4.1 and Gemini-2.5-Flash. The plots show the quality-diversity trade-off for varying p values. VS-Standard demonstrates a superior performance, with an optimal balance often found at p = 0.95. The inset provides a zoomed￾in view of each method’s performance curve. 10 12 14 16 18 20 22 Diversity 71 72 73 74 75 76 77 Quality p=0.0 p=0.01 p=0.02 p=0.05 p… view at source ↗
Figure 20
Figure 20. Figure 20: Min-p sampling analysis for poem generation across Qwen3-235B and Llama-3.1- 70B-Instruct. The plots show the quality-diversity trade-off for varying min-p values. Increasing min-p enhances diversity while reducing quality. VS-Standard outperforms the baselines, establishing a much more favorable Pareto front on both open-source models. 62 [PITH_FULL_IMAGE:figures/full_fig_p062_20.png] view at source ↗
Figure 21
Figure 21. Figure 21: Ablation of probability formats for Verbalized Sampling on the Poem Continuation Task. We evaluate VS-Standard (blue) and VS-Multi (red) on two models across two metrics: (a, c) Diversity (↑) and (b, d) Quality (↑). Subplots a–b report results on GPT-4.1, while c-d show results on Gemini 2.5 Flash. Prompt formats include Implicit, Explicit, Relative, Percentage, Confidence, NLL, and Perplexity. H.4 ABLATI… view at source ↗
Figure 22
Figure 22. Figure 22: Ablation of probability formats for Verbalized Sampling on the Open-Ended QA Task. We evaluate VS-Standard (blue) and VS-Multi (red) on two models across three metrics: (a, d) KL Divergence (↓), (b, e) Coverage-N (↑), and (c, f) Precision (↑). Subplots a–c report results on GPT-4.1, while d–f show results on Gemini 2.5 Flash. 1 0.1 0.01 0.001 VS Probability Threshold 10 12 14 16 18 Diversity Score Direct … view at source ↗
Figure 23
Figure 23. Figure 23: Diversity tuning results for Poem Continuation Task. Comparison of diversity scores across probability tuning parameters for GPT-4.1 (left) and Gemini 2.5 Flash (right). Notably, while VS-Multi initially falls behind VS-Standard at higher probability thresholds, its diversity improves more with diversity tuning. As the threshold decreases, VS-Multi’s diversity score catches up to that for GPT-4.1 (left) o… view at source ↗
Figure 24
Figure 24. Figure 24: Diversity tuning results for Story Generation. Comparison of diversity scores across probability tuning parameters for GPT-4.1 (left) and Gemini 2.5 Flash (right). The continuous y-axis shows the full range of diversity values. VS-Standard and VS-Multi maintain consistent performance advantages over baselines while exhibiting complementary tuning behaviors. The results demonstrate that diversity tuning pr… view at source ↗
Figure 25
Figure 25. Figure 25: Diversity tuning results for Joke Writing. Comparison of diversity scores across probability tuning parameters for GPT-4.1 (left) and Gemini 2.5 Flash (right). The x-axis shows probability thresholds in descending order from 1.0 to 0.001. VS-Standard and VS-Multi consistently outperform Direct and Sequence baselines across all parameter settings. Both VS variants show controllable diversity curves, with V… view at source ↗
Figure 26
Figure 26. Figure 26: Diversity tuning results for Open-Ended QA on Coverage-N. Results are shown for GPT-4.1 (left) and Gemini-2.5-Flash (right) across probability tuning parameters. Coverage-N measures the proportion of ground truth covered in the response distribution (higher is better). Both VS-Standard and VS-Multi consistently outperform the sequence baseline, with coverage increasing as probability decreases until ≤ 0.1… view at source ↗
Figure 27
Figure 27. Figure 27: Diversity tuning results for Open-Ended QA on KL Divergence over uniform distribution. Results are shown for GPT-4.1 (left) and Gemini-2.5-Flash (right) across probability tuning parameters. VS-Standard and VS-Multi achieve consistently lower divergence than the sequence baseline. The overall trend shows decreasing KL Divergence as probability decreases, indicating closer alignment with uniform distributi… view at source ↗
Figure 28
Figure 28. Figure 28: Diversity tuning results for Open-Ended QA on Precision. Results are shown for GPT-4.1 (left) and Gemini-2.5-Flash (right) across probability tuning parameters. 68 [PITH_FULL_IMAGE:figures/full_fig_p068_28.png] view at source ↗

discussion (0)

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

Forward citations

Cited by 29 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. ECHO: Learning Epistemically Adaptive Language Agents with Turn-Level Credit

    cs.MA 2026-06 unverdicted novelty 7.0

    ECHO is a clipped policy-gradient method that uses posterior-sensitive rewards to give turn-level epistemic credit in multi-turn information-seeking tasks, outperforming trajectory-level GRPO on a new Clue Selector Ga...

  2. The Piggyback Hypothesis of Generalization: Explaining and Mitigating Emergent Misalignment

    cs.CL 2026-06 conditional novelty 7.0

    Shared chat-template tokens piggyback narrow finetuning behaviors onto out-of-domain queries; regularizing their KV states (TReFT) reduces emergent misalignment and other off-topic generalization.

  3. When Discourse Pressures Conflict: Information Structure in Vision-Language Model Outputs

    cs.CL 2026-05 conditional novelty 7.0

    VLMs use Hungarian word order to mark Topic and Focus but drastically underproduce the variable strategies humans show under conflicting discourse pressures, resembling mode collapse.

  4. More Is Not More: What Matters for Diversity in LLM Opinions?

    cs.CL 2026-05 conditional novelty 7.0

    Diversity in LLM opinions comes mostly from the first persona sentence and from combining different interaction architectures, not from richer personas, temperature, or diversity instructions.

  5. Measuring and Mitigating Persona Distortions from AI Writing Assistance

    cs.CL 2026-04 conditional novelty 7.0

    AI writing assistance systematically distorts how writers are perceived across 29 social dimensions, and mitigating undesirable distortions reduces user preference for AI-assisted text.

  6. The Impact of AI-Generated Text on the Internet

    cs.CY 2026-04 unverdicted novelty 7.0

    By mid-2025 roughly 35% of new websites are AI-generated or AI-assisted, correlating with lower semantic diversity and higher positive sentiment but showing no significant drop in factual accuracy or stylistic diversity.

  7. Dynamic Context Evolution for Scalable Synthetic Data Generation

    cs.CL 2026-04 conditional novelty 7.0

    Dynamic Context Evolution prevents cross-batch mode collapse in LLMs by combining model self-assessment for idea filtering, embedding-based deduplication, and evolving prompts, yielding zero collapse and consistently ...

  8. PromptEvolver: Prompt Inversion through Evolutionary Optimization in Natural-Language Space

    cs.LG 2026-04 unverdicted novelty 7.0

    PromptEvolver recovers high-fidelity natural language prompts for given images by evolving them via genetic algorithm guided by a vision-language model, outperforming prior methods on benchmarks.

  9. Validity, Sparse Holes, and Breadth in Language Generation: Banach Density, Topology, and Geometry

    cs.DM 2026-04 unverdicted novelty 7.0

    Banach density reveals topological dichotomies in language generation: 1/2 is always achievable in 1D for finite-rank spaces but impossible in some infinite-rank cases, unlike asymptotic density; d>=2 needs nondegeneracy.

  10. Validity, Sparse Holes, and Breadth in Language Generation: Banach Density, Topology, and Geometry

    cs.DM 2026-04 conditional novelty 7.0

    Under the stricter Banach-density measure, valid generation in the limit guarantees the optimal 1/2 coverage exactly when the language collection has finite Cantor-Bendixson rank; other collections force arbitrarily l...

  11. Instruction-Tuned Language Models Cannot Sample from Distributions They Can Describe

    cs.AI 2026-07 conditional novelty 6.0

    Instruction-tuned LLMs collapse per-call outputs onto a single answer while still being able to state the target distribution accurately in one call; prompt perturbation recovers some variation.

  12. Automated IEP Generation from Traditional Chinese Parent-Teacher Interviews via Corpus-Grounded Feature Diffusion

    cs.CL 2026-06 unverdicted novelty 6.0

    Corpus-Grounded Feature Diffusion generates synthetic Traditional Chinese IEP training data from 25 seeds to fine-tune a 7B model that reaches BERTScore F1 0.779 on a 10-sample hold-out, beating several larger zero-sh...

  13. The Piggyback Hypothesis of Generalization: Explaining and Mitigating Emergent Misalignment

    cs.CL 2026-06 unverdicted novelty 6.0

    The Piggyback Hypothesis attributes emergent misalignment to chat-template tokens piggybacking finetuned behavior; Token-Regularized Finetuning (TReFT) mitigates it by regularizing prefix token representations.

  14. "I've Seen How This Goes": Characterizing Diversity via Progressive Conditional Surprise

    cs.CL 2026-06 unverdicted novelty 6.0

    Decan (D_Ca_n = C × a_n) measures text diversity as progressive conditional surprise from base LM log-probabilities, scoring 0.846 OCA on McDiv benchmark and detecting monotonic diversity drop across base→SFT→DPO→RLVR stages.

  15. KnowledgeGain: Evaluating and Optimizing Science News Generation for Reader Learning

    cs.CL 2026-05 unverdicted novelty 6.0

    Introduces KnowledgeGain metric to evaluate science news by reader knowledge gain, validated in human studies and used with an LLM simulator to select better articles.

  16. When Discourse Pressures Conflict: Information Structure in Vision-Language Model Outputs

    cs.CL 2026-05 unverdicted novelty 6.0

    VLMs over-regularize Topic/Focus realizations in Hungarian visually-grounded QA compared to humans who use variable strategies under discourse, grammatical, and definiteness pressures.

  17. Simulating Human Memory with Language Models

    cs.CL 2026-05 unverdicted novelty 6.0

    Language models show superior memory to humans on psych experiments but can be adjusted via prompting and compaction to forget more human-like, yielding better user simulators.

  18. Unlocking LLM Creativity in Science through Analogical Reasoning

    cs.AI 2026-05 conditional novelty 6.0

    Analogical reasoning increases LLM solution diversity by 90-173% and novelty rate to over 50%, delivering up to 13-fold gains on biomedical tasks including perturbation prediction and cell communication.

  19. Shepherd: Enabling Programmable Meta-Agents via Reversible Agentic Execution Traces

    cs.AI 2026-05 unverdicted novelty 6.0

    Shepherd provides a reversible execution trace substrate for LLM agents that enables meta-agents to inspect and transform runs, yielding reported gains on coding and terminal benchmarks via supervision, counterfactual...

  20. Annotations Mitigate Post-Training Mode Collapse

    cs.CL 2026-05 unverdicted novelty 6.0

    Annotation-anchored training reduces semantic diversity collapse in post-trained language models by a factor of six compared to standard supervised fine-tuning while preserving instruction-following and improving with scale.

  21. FlexSQL: Flexible Exploration and Execution Make Better Text-to-SQL Agents

    cs.CL 2026-05 unverdicted novelty 6.0

    FlexSQL reaches 65.4% on Spider2-Snow by allowing agents to flexibly explore schemas, generate diverse plans, choose SQL or Python execution, and apply two-tiered repair.

  22. Measuring and Mitigating Persona Distortions from AI Writing Assistance

    cs.CL 2026-04 conditional novelty 6.0

    AI writing distorts perceived writer personas across 29 dimensions in large experiments, and reward-model mitigation reduces but does not eliminate user preference for the AI.

  23. TurboEvolve: Towards Fast and Robust LLM-Driven Program Evolution

    cs.NE 2026-04 unverdicted novelty 6.0

    TurboEvolve improves LLM program evolution by running parallel islands with LLM-generated diverse candidates that carry self-assigned weights, an adaptive scheduler, and clustered seed injection to reach stronger solu...

  24. Seeing the Hivemind: A Consensus-Aware Interaction Technique for Mitigating AI Homogenization

    cs.HC 2026-06 unverdicted novelty 5.0

    Introduces Semantic Repulsion Technique (SRT) that boosts semantic diversity in AI creative outputs by 85-167% and receives higher usefulness and coherence ratings than baselines in a 16-person user study.

  25. Shepherd: Enabling Programmable Meta-Agents via Reversible Agentic Execution Traces

    cs.AI 2026-05 unverdicted novelty 5.0 partial

    Shepherd is a runtime system that formalizes meta-agent operations via typed execution traces, enabling fast forking and demonstrated improvements in agent intervention, optimization, and training on benchmarks.

  26. DORA Explorer: Improving the Exploration Ability of LLMs Without Training

    cs.CL 2026-04 unverdicted novelty 5.0

    DORA Explorer boosts LLM agent exploration without training by ranking diverse actions using log-probabilities and a tunable parameter, yielding UCB-competitive results on multi-armed bandits and gains on text adventu...

  27. MAR:Multi-Agent Reflexion Improves Reasoning Abilities in LLMs

    cs.AI 2025-12 reject novelty 4.0

    Multi-agent persona debate as a replacement for single-agent self-reflection modestly improves GPT-3.5's final-answer accuracy on HotPotQA and HumanEval, though the reported comparison is not trial-matched and lacks s...

  28. The Hitchhiker's Guide to Agentic AI: From Foundations to Systems

    cs.AI 2026-06 unverdicted novelty 2.0

    A comprehensive reference book organizing existing techniques for agentic AI systems across LLM substrate, reasoning, agent design patterns, inter-agent coordination, and production deployment.

  29. The Hitchhiker's Guide to Agentic AI: From Foundations to Systems

    cs.AI 2026-06 unverdicted novelty 1.0

    A survey-style reference book mapping the full agentic-AI stack from transformer internals to production deployment, with no new research result.

Reference graph

Works this paper leans on

22 extracted references · 7 linked inside Pith · cited by 23 Pith papers

  1. [1]

    Here, we can directly compare the probability assigned byπ ∗ andπ

    Instance-level prompts (the standard case) return only one completion. Here, we can directly compare the probability assigned byπ ∗ andπ

  2. [2]

    The natural assumption, without added information, is that each completion is chosen at random with equal probability

    List-level prompts return several possible completions, but no probabilities. The natural assumption, without added information, is that each completion is chosen at random with equal probability

  3. [3]

    Please write a poem starting with the following line: Fire up the sky with rocket sparks!

    Distribution-level prompts return completions together with probability estimates. In this case, it is reasonable to assume that sampling follows the provided probabilities. This distinction explains why distribution-level prompts can accurately replicate πref, as we prove next. It also aligns with our experimental results comparing π∗ under distribution-...

    2025

  4. [6]

    Peter West and Christopher Potts

    URLhttps://arxiv.org/abs/2411.04368. Peter West and Christopher Potts. Base models beat aligned models at randomness and creativity,

    Pith/arXiv arXiv

  5. [7]

    Justin Wong, Yury Orlovskiy, Michael Luo, Sanjit A

    URLhttps://arxiv.org/abs/2505.00047. Justin Wong, Yury Orlovskiy, Michael Luo, Sanjit A. Seshia, and Joseph E. Gonzalez. Simplestrat: Diversifying language model generation with stratification, 2024. URL https://arxiv.org/ abs/2410.09038. Jiancong Xiao, Ziniu Li, Xingyu Xie, Emily Getzen, Cong Fang, Qi Long, and Weijie J Su. On the algorithmic bias of ali...

    Pith/arXiv arXiv 2024

  6. [8]

    typicality bias

    URLhttp://arxiv.org/abs/2306.13063. arXiv:2306.13063 [cs]. Weijia Xu, Nebojsa Jojic, Sudha Rao, Chris Brockett, and Bill Dolan. Echoes in ai: Quantifying lack of plot diversity in llm outputs.Proceedings of the National Academy of Sciences, 122(35), August 2025. ISSN 1091-6490. doi: 10.1073/pnas.2504966122. URL http://dx.doi.org/ 10.1073/pnas.2504966122. ...

    Pith/arXiv arXiv 2025

  7. [9]

    Sentences Flow Naturally

  8. [10]

    (2025b), which achieved 80% agreement with human raters

    Overall Reader Engagement 78 Preprint Joke Evaluation.For the joke writing task, we also employed Claude-3.7-Sonnet (Anthropic, 2025b) with a slightly modified version of the autograder prompt from Narad et al. (2025b), which achieved 80% agreement with human raters. The prompt and rubric are provided below: Joke Autograder Rubric You will receive:

  9. [12]

    responses

    dataset. The prompts used for both direct and verbalized sampling prompting are as follows. Direct Prompt: You are an Amazon Mechanical Turk worker completing a 2-dollar communication task. •You are motivated by this task payment -- you value every cent you earn. •Act naturally as the person in the <persona> tag--think and respond as they would, including...

    2021

  10. [13]

    responses

    dataset to evaluate if verbalized sampling can retain or improve on the factual accuracy. The prompts are as follows. Direct Prompt: Provide your best guess for the given question. Output ONLY the response, no explanations or extra text. Verbalized Sampling (Standard) Prompt: Provide your {num_samplings} best guesses for the given question. Return the res...

    2023

  11. [14]

    Surprising and Creative

  12. [15]

    Imagery and Descriptive Quality

  13. [16]

    Well-earned Lightness or Darkness

  14. [17]

    Emotionally Engaging

  15. [18]

    Consistent Voice/Tone of Writing

  16. [21]

    The original joke prompt (may or may not contain a topic)

  17. [22]

    octopus,

    The model-generated joke. Your task is to evaluate the joke based on three qualitative metrics. Evaluation rules: - If the prompt includes a topic (e.g., "octopus," "coffee"), check whether the joke is on-topic and score Relevance from 0-5. - If the prompt does not include a topic (e.g., "Tell me a joke"), automatically assign Relevance = 5. - A good joke...

    2024

  18. [2019]

    doi: 10.18653/v1/P19-1566

    Association for Computational Linguistics. doi: 10.18653/v1/P19-1566. URL https: //aclanthology.org/P19-1566/. Yizhong Wang, Yeganeh Kordi, Swaroop Mishra, Alisa Liu, Noah A Smith, Daniel Khashabi, and Hannaneh Hajishirzi. Self-instruct: Aligning language models with self-generated instructions. In Proceedings of the 61st Annual Meeting of the Association...

    doi:10.18653/v1/p19-1566 2024

  19. [2022]

    doi: 10.18653/v1/2022.semeval-1.155

    Association for Computational Linguistics. doi: 10.18653/v1/2022.semeval-1.155. URL https://aclanthology.org/2022.semeval-1.155/. Yanran Chen, Hannes Gröner, Sina Zarrieß, and Steffen Eger. Evaluating diversity in auto- matic poetry generation. In Yaser Al-Onaizan, Mohit Bansal, and Yun-Nung Chen (eds.),Pro- ceedings of the 2024 Conference on Empirical Me...

    Pith/arXiv arXiv 2022

  20. [2023]

    URLhttps://arxiv.org/abs/2303.16694. 15 Preprint Stephen Casper, Xander Davies, Claudia Shi, Thomas Krendl Gilbert, J’er’emy Scheurer, Javier Rando, Rachel Freedman, Tomasz Korbak, David Lindner, Pedro J Freire, Tony Wang, Samuel Marks, Charbel-Raphaël Ségerie, Micah Carroll, Andi Peng, Phillip J. K. Christoffersen, Mehul Damani, Stewart Slocum, Usman Anw...

    Pith/arXiv arXiv 2023

  21. [2024]

    arXiv:2407.21787 [cs] version: 1

    URLhttp://arxiv.org/abs/2407.21787. arXiv:2407.21787 [cs] version: 1. Tristan J. B. Cann, Ben Dennes, Travis Coan, Saffron O’Neill, and Hywel T. P. Williams. Using se- mantic similarity and text embedding to measure the social media echo of strategic communications,

    Pith/arXiv arXiv

  22. [2025]

    Yann Dubois, Chen Xuechen Li, Rohan Taori, Tianyi Zhang, Ishaan Gulrajani, Jimmy Ba, Carlos Guestrin, Percy S Liang, and Tatsunori B Hashimoto

    URLhttps://arxiv.org/abs/2501.12948. Yann Dubois, Chen Xuechen Li, Rohan Taori, Tianyi Zhang, Ishaan Gulrajani, Jimmy Ba, Carlos Guestrin, Percy S Liang, and Tatsunori B Hashimoto. Alpacafarm: A simulation framework for methods that learn from human feedback.Advances in Neural Information Processing Systems, 36:30039–30069, 2023. Rudolph Flesch. A new rea...

    Pith/arXiv arXiv 2023