Pith. sign in

REVIEW 4 major objections 5 minor 14 references

An LLM can be made to aim its outputs at requested points in a frozen 3D coordinate space, turning generation into endpoint search.

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 ↗

Retrieval-grounded reinforcement learning makes an autoregressive LLM hit requested coordinates in a frozen encoder-defined PCA space, giving an outer loop a low-dimensional target to sweep or optimize.

T0 review reviewed 2026-08-03 challenge →

load-bearing objection New and honestly scoped interface — but z* is really an exemplar-library index, not a directly actuated coordinate, which limits the outer-loop claim; worth a careful referee. the 4 major comments →

arxiv 2601.21169 v1 pith:ZO64LCKH submitted 2026-01-29 cs.CL cs.AI

Output-Space Search: Targeting LLM Generations in a Frozen Encoder-Defined Output Space

Tobias Materzok This is my paper
classification cs.CL cs.AI
keywords output-space searchcontrollable generationretrieval groundingsequence-level RLlow-dimensional output spaceautoregressive decodinggeneration diversityblack-box optimization
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

This paper tries to establish that an autoregressive language model can be made to aim its outputs at requested points in a fixed, low-dimensional space added on top of a frozen encoder. An outer loop chooses a target coordinate once per sample; a retrieval-grounded policy trained with sequence-level reinforcement learning generates text or code whose realized coordinates land near that target under ordinary autoregressive decoding. The payoff is that generation becomes endpoint search: diversity can be produced by sweeping a small grid of targets in parallel, and objectives can be optimized by black-box search over the same coordinates without feeding the objective into the generator. The paper reports a 3.1x gain in LLM-judged story-branch diversity over prompt chaining, and improved performance on a withheld executable objective for constrained code under matched budgets. A sympathetic reader would say the core message is that a reusable coordinate interface for a model's output space is feasible, if the library of exemplars that grounds each target is dense enough.

Core claim

On the paper's own terms, the discovery is that an autoregressive model can act as an actuator for a frozen, encoder-defined 3D coordinate space. The coordinates are z(x)=U^T(E(x)-mu), fit once per domain; the model is trained with sequence-level RL to follow requests z* and to self-report its own coordinates. The downstream objective is never used during training, so the actuator is objective-agnostic. Evidence in both domains: story grid sweeps yield much higher judged branch diversity than chained sampling, and code Bayesian search over targets, updating at realized coordinates z(x), improves a withheld executable score under a matched budget of valid programs.

What carries the argument

The central object is the frozen coordinate map z(x)=U^T(E(x)-mu), a 3D projection of a frozen encoder's embedding, fit once per domain by PCA followed by Varimax rotation and then never changed during RL. The mechanism that makes a numeric target actionable is retrieval grounding: each request z* is instantiated by retrieving the two nearest exemplars to z* and one contrast exemplar near -z* from a library indexed in Z, and the policy is trained with group-based sequence-level RL to follow the neighborhood. Empirically, the numeric coordinate line can be hidden with almost no loss, while removing or mismatching exemplars collapses control, so exemplar selection in the frozen space is what c

Load-bearing premise

Control is carried by library coverage, not by the numeric target: the paper's ablations show nearly unchanged targeting when the numeric target is hidden but near-collapse when exemplars are removed, so any target region without dense exemplars disables the interface.

What would settle it

Pick a target z* in a region of Z that has no near exemplars and has not been seen in training, ask the trained controller to land within ±0.05 on all three axes, and score best-of-5; if success stays near zero, as it does when exemplars are removed (Bo5 ≈ 0.001), then the apparent control is library coverage rather than coordinate targeting.

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

If this is right

  • Diversity becomes parallel and context-free: branches sampled at different targets are conditionally independent, so context never grows and branches can be generated concurrently; the story results show a 3.1x judged-diversity ratio over the strongest chained baseline in the paper's harness.
  • Outer loops can optimize arbitrary withheld objectives over the same actuator; the code results show a surrogate-based search over targets improving a score that never appeared in the RL reward.
  • Search moves from high-dimensional discrete token or program space to a 3D continuous space, so standard continuous optimizers and grids become applicable to LLM sampling.
  • Best-of-k at a fixed target sharpens accuracy, giving a small, tunable accuracy-versus-compute knob at inference time.
  • The same coordinate space is reusable across prompts, and the code setup shows the exemplar library can grow online during RL, absorbing new valid examples without retraining the encoder.

Where Pith is reading between the lines

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

  • An implication the author stops short of: the ablations suggest the RL-trained policy is largely a vehicle for exemplar-following, so a sufficiently capable base model might achieve similar targeting with a dense enough library and no RL at all; this is testable by swapping the trained policy for a strong instruction-following model at the same prompt schema.
  • A testable extension: make reachability explicit by building a coverage map of Z from the exemplar library, and let the outer loop query only regions above a density threshold; this directly addresses the paper's own non-uniform reachability limitation.
  • An extension implicit in the anchoring experiment: other anchor sets could produce interpretable axes for genre, sentiment, or formality; the paper validates only one anchored axis, so the generality of axis semantics is open.
  • The paper's safety warning implies a constructive caution: if extreme-tail targets can elicit graphic content, a guardrail could be implemented as a bounding box or density constraint on z* in addition to content filtering.
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 / 5 minor

Summary. OS-Search proposes to turn LLM generation into endpoint search in a frozen, encoder-defined 3D output space Z. An outer loop proposes a target z*, a retrieval-grounded policy trained with GRPO/GSPO generates samples whose embedded <text> coordinates are rewarded to lie near z*, and the model self-reports a calibration estimate. The paper evaluates two applications: story diversity by sweeping a grid in an anchored Ztext (Tab. 2, LLMScore 57.15 vs. best chained baseline 18.45, rho_div ≈ 3.10), and code search over constrained make_seed() programs by Bayesian optimization over Zcode under a withheld CA++ objective (Fig. 6, best CA++ 0.395 vs. library baseline 0.371). The paper is explicitly framed as a feasibility/interface study and discusses its limitations in detail.

Significance. If the results hold, the paper contributes a concrete recipe for a reusable low-dimensional control interface for LLM outputs: a frozen representation, retrieval grounding, and sequence-level RL. The code-domain experiments are the strongest part: calibration curves, a clean prompt ablation (App. Tab. 5), three-seed BO ablations, and a disclosed synthetic benchmark. The paper is honest about non-uniform reachability and library dependence. However, the ablation that gives the paper credibility also reveals that the numeric target is nearly redundant in code, so the contribution is best characterized as an exemplar-neighborhood search interface indexed by z* rather than direct coordinate actuation. The story diversity claim lacks error bars and judge replication, which limits its evidentiary weight.

major comments (4)
  1. [§4.1 and App. Tab. 5] The load-bearing claim in §1 is that an outer loop chooses z* and the model generates x with z(x) near z*, enabling sweep/optimization of the coordinate. App. Tab. 5 shows that in the code domain, hiding the REQUESTED TARGET line leaves Bo1 Success@0.05 almost unchanged (0.658 vs. 0.664 with the default prompt), while removing exemplars collapses control to 0.001 and validity to 35.3%. Thus the policy primarily follows the retrieved near/contrast exemplars; z* acts as an index into the exemplar library. The outer loop is consequently an exemplar-neighborhood search, whose reachability is bounded by library density. The paper discloses non-uniform reachability (§6, Fig. 11), but the abstract and introduction still present the numeric coordinate as the control interface. Please either reframe the contribution accordingly or demonstrate direct numeric actuation (e.g., vary z* while holding
  2. [§4.3 / Tab. 2] The headline diversity gain (LLMScore 57.15 vs. 18.45, rho_div ≈ 3.10) is a point estimate over 20 prompts with a single GPT-5.1 judge configuration. No prompt-level variance, bootstrap interval, or repeated-judge agreement is reported. Because the method comparison rests on this ratio, report the per-prompt distribution and a confidence interval (e.g., bootstrap over prompts and/or judge calls), and state whether the judge is blind to the sampling condition.
  3. [§4.4 / Fig. 6 / App. I] The BO comparison may give the optimizer free evaluations: the warm-start phase uses 1,375 pre-scored random-z* rollouts before the plotted BO segment (App. I). If these scored programs are not counted in Nok for both BO and path-based sampling, the 'matched budget' claim is not satisfied. Please present a version where the warm-start evaluations are explicitly included in the budget for all methods, or run BO from scratch, and clarify the status of the Baseline (Zcode) sweep (fixed library) versus the budget-matched baseline.
  4. [§4.1 / Fig. 3 and Tabs. 3–4] Success@ε curves are computed on 'usable completions' only (stories N=6,946, code N=6,860), but invalid outputs count as failures in some places. Please make the denominator convention consistent across Fig. 3, Tabs. 3–4, and App. Tab. 5, and report the raw validity rate per domain. This matters because the code validity drop to 35.3% without exemplars affects the interpretation of control.
minor comments (5)
  1. [§3.1 / App. A] The Z-space hyperparameter selection uses a fixed scalar preference with fixed weights; report sensitivity of downstream results to these choices.
  2. [§4.3 / Tab. 2] Degen for OS-Search (0.09) is slightly higher than the best baseline (0.07); report whether this is significant and whether the LLMScore ratio is robust to Degen-matching.
  3. [§6 / Resources] Story-domain training data and checkpoints are not released; at minimum release the 20 WritingPrompts prompts and exact prompt templates to allow reproduction of Tab. 2.
  4. [App. G] The 15-branch truncation of the 18-target grid is deterministic but arbitrary; report whether the headline ratio changes if the omitted three targets differ.
  5. [Fig. 4] The Slop-Score correlation is on rollouts from the final 10% of training; clarify whether these rollouts used fixed prompts or varied prompts, and whether the correlation is dominated by a few prompts.

Circularity Check

0 steps flagged

No significant circularity: the paper's central claims are empirical and rest on external measures, with the key retrieval-dependence explicitly disclosed.

full rationale

The paper reports an empirical interface study rather than a derivation from first principles. The output space Z is constructed by a frozen encoder plus PCA/Varimax, and the controller is trained with sequence-level RL to place outputs near requested coordinates; headline evaluations use external measures (lexical/embedding overlap, GPT-5.1 judge, Slop-Score analyzer, executable CA++ scores). Target-tracking accuracy is a trained behavior, not a prediction from an input, and the paper does not disguise it as an independent result; the calibration curves and ablations (App. Tab. 5) explicitly show that actuation is carried largely by exemplar retrieval, which scopes but does not circularize the claim. The z1/Slop-Score correlation is explicitly labeled a post-hoc axis-validation diagnostic, and the anchor construction is disclosed, so it is not presented as an independent derivation. The diversity gain from choosing a spread-out target grid is the intended mechanism of the method, not a circular prediction. No load-bearing self-citation or uniqueness argument appears; the only self-citation (Materzok 2025) is contextual related work. Hence no step reduces by construction to its inputs.

Axiom & Free-Parameter Ledger

7 free parameters · 6 axioms · 0 invented entities

The central claim rests on many hand-set engineering choices (anchor set, reward weights 3/3/1.5, Z-construction grid search, target bands, CA++ weights), all disclosed in appendices, plus domain assumptions about the encoder, PCA/Varimax geometry, the RL optimizer, the LLM judge, and exemplar-library coverage that are only checked empirically. No new physical entities are introduced; 'output space', 'state-like target', and 'endpoint search' are framing terms for a frozen PCA projection of embeddings. The honest reading: a large but disclosed parameter surface, with external evaluations for each headline claim.

free parameters (7)
  • Anchor set A for Z_text z1 axis = small hand-picked set of Qwen3-generated 'default-style' stories (size not stated)
    Defines the semantic meaning of the first story axis; the Slop-Score correlation (r≈0.57) validates this axis post hoc, but the direction itself is chosen by the authors (App. A).
  • Reward shaping weights (3, 3, 1.5) = 3·R_dist, 1.5·R_hon
    Hand-chosen coefficients in the RL reward (Sec. 3.2, App. D), not ablated; they set the target-hitting vs honesty trade-off.
  • Z-construction hyperparameters = d_inter ∈ {8..48}, whitening on/off, dz = 3 (best by preference grid search)
    Selected by an ad hoc scalar preference score with weights 'reported in the code release' (App. A); the space geometry itself is tuned on these.
  • Distance exponent alpha curriculum = annealed 1.5 → 0.8 per 1500-update phase
    Controls how sharply reward concentrates near the target; hand-scheduled (App. C, Fig. 7).
  • Target sampling bands and story grid targets = text ≤1.75·s_i, code ≤1.5·s_i; story grid z1∈{½s1,s1}, z2,z3∈{-s_i,0,+s_i}
    Hand-picked magnitudes and grid levels; the story diversity claim rests on these well-separated targets (App. C, App. G).
  • CA++ benchmark weights = 0.30 act + 0.20 div + 0.25 pent + 0.20 ccont + 0.05 bal; T_N=64/96; LIFE/HIGHLIFE/SEEDS
    The withheld objective is entirely author-designed (App. H); the BO result optimizes this specific composite score.
  • Frozen per-axis scales s_i = max(|q0.10|, |q0.90|, 1e−4)
    Data-derived from reference corpus; sets request ranges and reward normalization (App. C).
axioms (6)
  • domain assumption Frozen encoder embeddings (all-mpnet-base-v2 for text, jina-embeddings-v2-base-code for code) capture task-relevant style/semantic structure
    The whole approach inherits its geometry from these encoders (§3.1, App. A); if embeddings do not track the properties an outer loop cares about, Z is useless.
  • domain assumption PCA + (anchored) Varimax directions correspond to coherent output variation
    Used to define z(x); validated only for z1 via Slop-Score, while z2/z3 remain uninterpreted (App. A, Fig. 9).
  • domain assumption GRPO/GSPO group-relative RL reliably learns the target-hitting policy
    The controller's accuracy (Fig. 3, Tabs. 3–4) is the empirical check of this training assumption; there is no formal guarantee.
  • domain assumption GPT-5.1 judge scores diversity and degeneration validly for these outputs
    The headline story claim (3.1x) uses this judge; the paper notes judge bias is possible (§6, App. G) and reports no human agreement.
  • ad hoc to paper CA++ composite score is a meaningful objective for make_seed programs
    Author-designed synthetic benchmark (App. H); the paper concedes it is not representative of real software engineering (§6).
  • domain assumption Exemplar library covers the reachable part of Z
    Control is carried by retrieval (App. Tab. 5); sparse regions cannot be actuated, and App. I Fig. 11 notes target combinations that are effectively unattained.

reviewed 2026-08-03 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Output-Space Search: Targeting LLM Generations in a Frozen Encoder-Defined Output Space." pith.science (2026). https://pith.science/paper/ZO64LCKH

@misc{pith2026260121169,
  author       = {Pith},
  title        = {Pith review of: Output-Space Search: Targeting LLM Generations in a Frozen Encoder-Defined Output Space},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ZO64LCKH}},
  note         = {Machine review of arXiv:2601.21169}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

We introduce Output-Space Search (OS-Search), which turns LLM generation into endpoint search. An outer loop selects a target z* in a frozen encoder-defined 3D output space Z, and a retrieval-grounded policy trained with sequence-level RL generates outputs whose coordinates land near z* under standard autoregressive decoding. This enables parallel sweeps and black-box optimization in Z without path-dependent token/program search. On stories, sweeping Z (text) yields 3.1x higher LLM-scored diversity than prompt-chaining. On code, Bayesian optimization over Z (code) improves an objective withheld from the controller under matched inference budgets while preserving validity.

Figures

Figures reproduced from arXiv: 2601.21169 by Tobias Materzok.

Figure 1
Figure 1. Figure 1: Path- vs state-like control (schematic 2D slice of Z). (a) Standard decoding: a fixed prompt yields token-space paths whose endpoints concentrate in one region of the output space Z. (b) z ∗ -conditioned decoding: the same base prompt plus a requested target z ∗ (grounded by retrieved exemplars) shifts endpoints toward z ∗ in Z (up to targeting error), while the decoder remains autoregressive. (c) External… view at source ↗
Figure 2
Figure 2. Figure 2: Prompt schema for OS-Search. The prompt in￾cludes retrieved exemplars with stored realised coordinates z(xj ) and the numeric request z ∗ S. The model outputs a struc￾tured completion y containing task content x and a self-report zˆS. We compute realised coordinates by embedding only <text> to obtain z(x). 3.1 Representation: constructing the output space for text and code We construct a low-dimensional ou… view at source ↗
Figure 3
Figure 3. Figure 3: Success–tolerance curves for 3D control in Z. Dashed curves show one-shot sampling, and solid curves show best-of-5 at the same request. Malformed/invalid outputs count as failures. Best checkpoint (blue) against first 5% of the RL run (gray). by exemplar selection rather than direct numeric￾coordinate parsing. Calibration plots, with targets spanning q0.01 to above q0.99, show code track￾ing within ±0.05 … view at source ↗
Figure 4
Figure 4. Figure 4: Relationship between requested z ∗ 1 and Slop￾Score. Slop-Score as a function of the requested target on the first axis, z ∗ 1 , for completions from late-training rollouts (final 10% of GRPO updates) generated under 1D control (S = {1}) (N ≈ 7,000). Points are coloured by local density. The solid line shows a LOWESS smooth, the dotted line the global linear regression fit (r ≈ 0.57), and the shaded band a… view at source ↗
Figure 5
Figure 5. Figure 5: Programs in CA++ behavioral space (code do￾main). Each point is a valid make_seed() program scored by CA++. Axes are PCs of the CA++ trajectory-statistic feature vectors of the plotted programs (so proximity indicates simi￾lar CA dynamics), not Zcode. Hexbins show random-z ∗ con￾troller samples coloured by mean score per bin. Red outlines mark bins reached by the N=188 program-library baseline, and stars m… view at source ↗
Figure 6
Figure 6. Figure 6: Best-so-far CA++ score vs scored-program bud￾get. Nok counts scored valid programs (invalid/malformed outputs are discarded). The grey segment indicates the warm￾start phase for BO (Nok = 1375). Thick lines show the median over three seeds, and thin lines show the min/max range (split y-axis for readability). our surrogate-update ablation (App. Tab. 7). Under a matched Nok budget, BO improves the best-so￾f… view at source ↗
Figure 7
Figure 7. Figure 7: Evolution of the curriculum reward surface. Re￾ward profiles for a target value z ∗ 1 = 0.2 are shown within a single curriculum phase. In each phase, the distance ex￾ponent α is linearly annealed from 1.5 to 0.8 over the 1 500 RL updates (inset), progressively narrowing the reward basin in the magnitude regime to demand higher precision later in that phase, while maintaining a distinct penalty step in the… view at source ↗
Figure 8
Figure 8. Figure 8: Calibration of state-like control in Z. Binned mean realised coordinate zi versus requested target z ∗ i for each constrained axis. The diagonal indicates perfect calibration. Blue: best checkpoint. Orange: early/baseline (first 5% of RL). The grey band marks the reference tolerance |zi − z ∗ i | ≤ 0.05 (raw z units). Vertical markers show reference-corpus quantiles (q0.01, . . . , q0.99) of zi Curves are … view at source ↗
Figure 9
Figure 9. Figure 9: Slop-Score vs realised coordinates in Z. Hexbin plots of Slop-Score against realised coordinates z1, z2, z3 for all story completions. Each panel shows a global linear regression fit (black dashed line) and the corresponding Pearson correlation r with two-sided p-value. Only the first axis exhibits a strong positive association with Slop-Score (r ≈ 0.56, left), while correlations on the remaining axes are … view at source ↗
Figure 10
Figure 10. Figure 10: Qualitative excerpts along the anchored story axis z1 (GRPO training prompt template). Three independent completions to the same prompt template used during GRPO training, shown at increasing realised z1 (with z2, z3 roughly matched). We report frozen encoder-defined coordinates z(x) in Ztext and show only short excerpts (opening + ending) for space. The shift toward more “default-style”/templated narrati… view at source ↗
Figure 11
Figure 11. Figure 11: Score structure in the (z1, z2) plane (code domain). Mean CA++ score as a function of (left column) the requested targets (z ∗ 1 , z∗ 2 ) and (right column) the corresponding realised coordinates (z1, z2) = z(x) of generated programs in Zcode. Top: the combined candidate pool from the Baseline (Zcode) sweep and random z ∗ samples (last valid 13,752 GRPO rollouts). Bottom: warm-start subset (1375 points) s… view at source ↗
Figure 12
Figure 12. Figure 12: Qualitative evolution of the BO-best seed under CA++ rules (N=16 horizon). Spatiotemporal snapshots of the best-performing BO-discovered make_seed() program, visualized over the CA++ horizon T16 = 64 on the 16 × 16 torus under LIFE, HIGHLIFE, and SEEDS. Each panel shows five snapshots from t = 0 to t = 64 (earlier layers are lighter). Update Growing Scale Best@2000 Best@End z(x) no 2.5 0.376 [0.376,0.384]… view at source ↗
Figure 13
Figure 13. Figure 13: BO-best make_seed() program BO-best make_seed() program used to generate [PITH_FULL_IMAGE:figures/full_fig_p023_13.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

14 extracted references · 8 linked inside Pith

  1. [2]

    feed all previously generated outputs back into the input

    story_2 . . . PROMPT:<original prompt> where story_1, . . . ,story_b-1 are the previously generated branches for that prompt, stripped of any <think> content. We include all earlier branches as negatives. This procedure reproduces the “feed all previously generated outputs back into the input” sampling baselines in Park et al. (2025). Decoding grids.We sw...

    2025

  2. [3]

    , TN } and cell location (i, j), we ex- tract the 3×3 Moore neighborhood pattern around (i, j)with wraparound, yielding a 9- bit code in {0,

    Patch entropy: for every time t∈ {0, . . . , TN } and cell location (i, j), we ex- tract the 3×3 Moore neighborhood pattern around (i, j)with wraparound, yielding a 9- bit code in {0, . . . ,511}. Let pk be the em- pirical frequency of codekover all neighbor- hoods across the full spacetime. The normal- ized entropy is pent = 1 9 − X k:p k>0 pk log2 pk , ...

  3. [4]

    ArXiv:2506.13131 [cs]

    AlphaEvolve: A coding agent for scientific and algorithmic discovery.arXiv preprint. ArXiv:2506.13131 [cs]. OpenAI. 2026. GPT-5.1 Model | OpenAI API. OpenAI, Aaron Jaech, Adam Kalai, Adam Lerer, Adam Richardson, Ahmed El-Kishky, Aiden Low, Alec Helyar, Aleksander Madry, Alex Beutel, Alex Carney, Alex Iftimie, Alex Karpenko, Alex Tachard Passos, Alexander ...

    Pith/arXiv arXiv 2026

  4. [5]

    ArXiv:2509.02170 [cs]

    Avoidance Decoding for Diverse Multi-Branch Story Generation.arXiv preprint. ArXiv:2509.02170 [cs]. 12 William H. Paulsen. 2003. Creating Large Life Forms with Interactive Life.Complex Systems, 14(3). Nils Reimers and Iryna Gurevych. 2019. Sentence-BERT: Sentence Embeddings using Siamese BERT-Networks.arXiv preprint. ArXiv:1908.10084 [cs]. Bernardino Rome...

    Pith/arXiv arXiv 2003

  5. [8]

    default-style

    Group Sequence Policy Optimization.arXiv preprint. ArXiv:2507.18071 [cs]. Yaoming Zhu, Sidi Lu, Lei Zheng, Jiaxian Guo, Weinan Zhang, Jun Wang, and Yong Yu. 2018. Texygen: A Benchmarking Platform for Text Generation Models. arXiv preprint. ArXiv:1802.01886 [cs]. Appendices A Output-space construction and evaluation We detail the construction of the frozen...

    Pith/arXiv arXiv 2018

  6. [9]

    not X, but Y

    as an automatic proxy for stereotyped, tem- plated “AI slop” writing. The analyzer compares a model completion against a fixed list of over- represented slop words and trigrams, and also de- tects overused contrast constructions of the form “not X, but Y” using a two-stage pattern-matching pipeline (regex followed by a lightweight POS- tag-based check) (P...

    2025

  7. [11]

    Define c= 1 TN TNX t=1 c(xt),div = min 1, c N 2/4

    Diversity (components): let c(xt) be the number of 4-neighbor connected components of live cells in xt (with wraparound adjacency on the torus). Define c= 1 TN TNX t=1 c(xt),div = min 1, c N 2/4

  8. [13]

    Compression contrast: we serialize the full spacetime (xt)TN t=0 into a row-major ASCII string over {0,1} of length L= (T N + 1)N2. We compress this byte string using DEFLATE at maximum compression level (9), let r= |compressed| |raw| be the byte-length ratio, and define ccont = 2 min(r,1−r) with |compressed| never being larger than |raw|

  9. [14]

    spaceship

    Balance: let ℓt = 1 N 2 P i,j xt[i, j]be the live- cell fraction and ¯ℓ= 1 TN PTN t=1 ℓt its mean (ex- cluding x0). We define the balanced-density score bal = 4 ¯ℓ(1− ¯ℓ). CA++ composite score.For each run (N,rule) we compute the weighted sum score(N,rule) = 0.30 act + 0.20 div + 0.25 pent + 0.20 ccont + 0.05 bal. The overall CA++ score is the mean over a...

    2000

  10. [2016]

    Mastering the game of Go with deep neural networks and tree search.Nature, 529(7587):484–489. David Silver, Thomas Hubert, Julian Schrittwieser, Ioannis Antonoglou, Matthew Lai, Arthur Guez, Marc Lanctot, Laurent Sifre, Dharshan Kumaran, Thore Graepel, Timothy Lillicrap, Karen Simonyan, and Demis Hassabis. 2017. Mastering Chess and Shogi by Self-Play with...

    Pith/arXiv arXiv 2017

  11. [2023]

    ArXiv:2212.08307 [cs]

    Controllable Text Generation via Probability Density Estimation in the Latent Space.arXiv preprint. ArXiv:2212.08307 [cs]. Rafael Gómez-Bombarelli, Jennifer N. Wei, David Duvenaud, José Miguel Hernández-Lobato, Benjamín Sánchez-Lengeling, Dennis Sheberla, Jorge Aguilera-Iparraguirre, Timothy D. Hirzel, Ryan P. Adams, and Alán Aspuru-Guzik. 2018. Automatic...

    Pith/arXiv arXiv 2018

  12. [2024]

    ArXiv:2405.03553

    AlphaMath Almost Zero: Process Supervision without Process.arXiv preprint. ArXiv:2405.03553. Paul Christiano, Jan Leike, Tom B. Brown, Miljan Martic, Shane Legg, and Dario Amodei. 2023. Deep reinforcement learning from human preferences.arXiv preprint. ArXiv:1706.03741 [stat]. Hoagy Cunningham, Aidan Ewart, Logan Riggs, Robert Huben, and Lee Sharkey. 2023...

    Pith/arXiv arXiv 2023

  13. [2025]

    ArXiv:2503.20783 [cs]

    Understanding R1-Zero-Like Training: A Critical Perspective.arXiv preprint. ArXiv:2503.20783 [cs]. Ilya Loshchilov and Frank Hutter. 2019. Decoupled Weight Decay Regularization.arXiv preprint. ArXiv:1711.05101 [cs]. Tobias Materzok. 2025. COS(M+O)S: Curiosity and RL-Enhanced MCTS for Exploring Story Space via Language Models.arXiv preprint. ArXiv:2501.171...

    Pith/arXiv arXiv 2019

  14. [2026]

    ArXiv:2601.16175 [cs]

    Learning to Discover at Test Time.arXiv preprint. ArXiv:2601.16175 [cs]. Chujie Zheng, Shixuan Liu, Mingze Li, Xiong-Hui Chen, Bowen Yu, Chang Gao, Kai Dang, Yuqiong Liu, Rui Men, An Yang, Jingren Zhou, and Junyang Lin

    Pith/arXiv arXiv

This paper was first reviewed by deepseek-v4-flash on August 3, 2026.