Which Compute, Which Data?

Polar Switch: optimize the circuit before approximating polar

Muon's cost depends on reusing matrix products, not only polynomial degree. For normalized X, Polar Switch searches polynomial circuits with limited buffers; the corresponding singular-value map is an odd polynomial. [1]

Loop depth versus backward depth: truncation need not save compute

Let T be forward loops and B≤T backward loops. Across 592 runs with 58–919M non-embedding parameters, repeated computation is mapped to effective parameters; N_once / N_rec count non-recurrent / recurrent parameters. [2]

More data makes the validation distribution matter

Eight real tabular tasks scale to ten million samples, evaluated under random splits and temporal/group shifts. Error separates into a decreasing term and a floor for dataset d and method m: [3]

Tuning on in-distribution validation can worsen out-of-distribution performance as data grows. Deployment-representative validation mitigates this in MLP experiments; adding OOD examples to PFN context does not always fix it. [3]

T-MUON: tensorized spectral updates

Extends Muon using t-SVD and transform-domain operators. Preferred tensorizations vary across 51M–1.3B pretraining models; one transform's advantage is not universal. [4]

A statistical model of Muon

Connects Adam, Shampoo and Muon through structured gradient covariance, proposing MN-Muon. It beats Muon in small GPT comparisons, but is comparable to or slightly below NorMuon in the best medium-scale configurations. [5]

DP-Muon: clipping matters alongside privacy noise

Analyzes clipping alongside orthogonalization; the best optimizer in language-model experiments still depends on batch size. The abstract does not support an unconditional rate or uniform superiority to DP-Adam. [6]

Halcyon: does text-to-image need language pretraining?

Trains from scratch on 100M image–text pairs. Controlled experiments find no T2I gain from text-only pretraining. Cross-model compute ratios compare different training recipes, not strictly matched single-model ablations. [7]

Ropedia W0: human experience and robot data

Uses about 10k hours of human video and 3k hours of robot experience for a cross-embodiment world model, reporting benefits from both data axes; no power-law exponent is verified here. [8]

Physical RSI: observational capability forecasting

Across 18 manipulation tasks with fixed interaction budgets, capability proxies and early feedback forecast self-improvement. Observational relationships do not identify causal effects of individual scaling axes. [9]

Answer delivery also costs tokens

Charges finalization prompts to the shared budget. On 100 fresh GSM8K problems per checkpoint, accuracy rises 8/7 percentage points, while physical token-slot savings remain below 1%. [10]

BOSS: generate, judge and stop under one budget

Charges both prompt and completion tokens while allocating generation, judging and stopping online. Best in 22 of 24 configurations, within the compared policies and budget grid. [11]

RAG: corpus growth also grows distractors

Holds relevant documents fixed while adding distractors to study retrieval rankings and storage/retrieval granularity. This is corpus scaling, not parameter scaling. [12]

Indic: mixing synthetic pretraining data

Builds a 2.83T-token resource across 22 languages and studies synthetic/natural text and English mixtures in continued pretraining. Mixture gains remain language- and evaluation-dependent. [13]

LP-SMC: particle count is not lineage diversity

Uses Chopthin and retention to reduce lineage collapse in diffusion-language decoding. Oracle coverage improves, with final accuracy improving in 11/12 cases. Reset ESS alone does not establish restored diversity. [14]

RLVR: limits of adaptive decoding support

Expands top-k for a 4B code model under 150M tokens. It beats fixed k=10 but is close to the best fixed k=20; the single test run ties. This does not establish universal superiority to fixed policies. [15]

Decision models after truncated reasoning

Evaluates 35 model pairings with a small decision model reading truncated traces. Cost benefits are offline held-out estimates, not direct measurements of online deployment savings. [16]

Spherical fixed-point reasoning and adaptive halting

Constrains recurrent states to a sphere and chooses depth through fixed-point residuals. The abstract describes conditional stability; without full assumptions, it is not reported as unconditional convergence. [17]

A horizon ceiling for curriculum learning

At matched compute, curriculum learning improves implicit long-horizon reasoning, but look-ahead saturates; scaling training does not remove the observed task-specific ceiling. [18]

ATEST: budgets respond to system load

Includes GPU load and energy in inference-budget control, trading accuracy against latency and throughput. Equal token budgets need not mean equal system cost. [19]

Noise-rejuvenated SMC for flow matching

Refreshes latent noise to preserve search diversity for test-time scaling of deterministic flow matching. This is not simply replacing the original ODE with a reverse diffusion SDE. [20]

P(Yes): what is the baseline behind 15×?

Search uses a frozen decision model's yes score; fully solved puzzles rise from 2 to 30 out of 128. The 15× ratio starts from a small baseline and should not replace absolute success rates. [21]

ATLAS: concentrate search in selected latents

Uses attention to target local candidates in visual autoregressive generation under a search budget. Relative gains on targeted categories are not overall image-quality gains. [22]

GeoART: geometric selection for camera-conditioned video

Scales test-time compute with geometry-aware candidate refinement and confidence gating, without extra training. The abstract is insufficient to verify strictly matched-budget superiority. [23]

BatchParScale: sparse communication across streams

Mixes states between parallel streams at selected positions, improving language modeling. Loss gains and factual-retrieval gains do not coincide and should not be conflated. [24]

E-process stopping: which error is controlled?

Uses online signal tests to shorten reasoning. The Ville-type bound controls signal false alarms under the relevant null, not the final-answer error rate. [25]

DeGMatch: data matching under a budget

Centers gradients, removes leading directions and gates matches to select target-relevant reasoning data. Evidence is experimental across budgets and model sizes, not a universally optimal mixture theorem. [26]

Scaling models also changes reliability profiles

Across 20 models and six benchmarks, paraphrase sampling separates accessibility from reliability. Higher mean scores do not imply reliability improves on every problem. [27]

ECHO: reuse the final block for depth scaling

Uses a shared final block and multi-step objectives to increase effective depth, comparing gains within budgets. Parameter sharing alone does not establish end-to-end cost savings. [28]

The gains and durability of pretraining filtering

Data-filtering effects in protein language models vary with model scale, while later fine-tuning can weaken their durability. This brief records only scaling and distributional findings. [29]

C-OPSD: not every trajectory needs distillation

A signal–noise model fitted at small group size tracks gradient alignment at larger groups. Filtering and packing reduce measured training time by about 40%, while rollout generation is unchanged. [30]

recAMPLIFY: recurrent depth in protein models

Compares 155M unique parameters trained on 2T tokens with a 350M baseline. Adaptive gating further reduces expected depth, with task-dependent accuracy trade-offs; included for test-time compute scaling. [31]

More student data can still inherit teacher bias

Gaussian surrogates yield finite-sample risks: more student samples need not remove teacher bias, and fixed representations help below a sample-size threshold. The surrogate assumptions are essential. [32]

Visual token count is not actual compute

Across eight pruning methods, three VLMs and nine benchmarks, equal token counts entail different end-to-end costs; speedup rankings change under a common capability-preservation target. [33]

Nesso 3: failed data mixtures are informative

The SFT-trained 0.4B model scores 74 on a 100-item tool test. Adding long-chain ToolMind data scores 49; stronger on-policy distillation reaches 75 but hurts Italian instruction-following and academic metrics, so it is excluded from the release. Over-refusal and mixture adjustments are also documented. Small tests, unreleased synthetic data and limited replication constrain generalization. [34]

Version watch: 25 papers, changes unverified

Topic descriptions identify the papers; they do not claim that these findings first appear in the revisions. Previously covered SLDBench, ICS and self-orchestration are not counted again as news.