@PithSci

A non-profit open preprint site with AI-review and feedback. Fast, continuous, and transparent. https://xitter.cf/t.co/jnhYcvjSbP

 Austin, TX
Joined July 2026
  • Tweets310
  • Following36
  • Followers186
1Media
Pinned Tweet
Pith.
@PithSciOct 2
arXiv now limits each person to two submissions a month, and rejected papers count toward that limit. At Pith, every paper you send us is read, reviewed in public, and open to challenge.
1
1
3
2,070
DiffPT pretrains a classifier backbone with a diffusion objective on the same small dataset. @yeewhye and coauthors report EyePACS AUROC of 0.744 against 0.643 for SimCLR with 1 percent of the training data. pith.science/paper/2610.2328…

Diffusion pretraining lifts EyePACS AUROC to 0.744 at 1 percent data

Diffusion model pretraining outperforms self-supervised baselines on medical and time-series tasks with limited labels.

pith.science
3
303
Ranger-8B scores 67.2 on Savanna, a 500-prompt benchmark built from recent headlines. @imlizheng and coauthors trained it on 15K Locate-Corroborate-Isolate trajectories; frontier-model harnesses reach 51.5. pith.science/paper/2610.2326…

Ranger-8B scores 67.2 on Savanna image-search benchmark

This paper establishes that current agentic image generation systems fail when asked about entities introduced only days ago, and proposes a three-step search paradigm that lifts success from 51% to...

pith.science
41
At 128 frames on DL3DV, LonGaussian reaches 18.31 PSNR with 4.80M Gaussians, against 14.83 PSNR and 13.67M for OF3GS. @DavenCheung and coauthors use multi-frame point tracks to merge redundant Gaussians within each chunk. pith.science/paper/2610.2325…

LonGaussian: 18.31 PSNR with 4.80M Gaussians at 128 frames

A streaming feed-forward framework that uses point-track correspondences to merge repeated Gaussian predictions, producing compact renderable 3D scenes from long videos while using 3–10× fewer...

pith.science
1
130
Chain-of-thought faithfulness rises as language resource level falls, but only on knowledge-intensive tasks. @IAugenstein and coauthors find it on parallel MMMLU questions in eight languages, across four models. pith.science/paper/2610.2284…

CoT faithfulness rises as language resource level falls

Chain-of-thought faithfulness rises on knowledge tasks as language resources fall, supporting the Reliance Hypothesis.

pith.science
1
77
93.4 PDMS on NAVSIM-v1, the highest among compared methods. @DavenCheung and coauthors reach it with WAC, which jointly learns candidate-conditioned future prediction and consequence evaluation over multimodal action candidates. pith.science/paper/2610.2282…

WAC reaches 93.4 PDMS on NAVSIM-v1

WAC jointly learns candidate-conditioned future prediction and consequence evaluation, achieving 93.4 PDMS on NAVSIM-v1.

pith.science
50
Global learning coefficients in the joint-parameter 2PL model depend only on the true response probabilities, even when some equal one half or are constant across respondents. @joe_suzuki and coauthors give them explicitly. pith.science/paper/2610.2310…

2PL global learning coefficients depend only on true probabilities

Explicit learning coefficients for the joint-parameter 2PL model, expressed solely in terms of true response probabilities.

pith.science
22
Naive retrieve-then-answer scored below the unaided model on every backbone tested. @MeilerLab and coauthors show Munin's agent harness instead more than doubles accuracy, from 0.387 to 0.874. pith.science/paper/2610.2348…

Naive RAG underperforms; agent harness more than doubles accuracy

A self-hosted agent harness, not the specific model, more than doubles accuracy on domain literature questions and eliminates fabricated citations.

pith.science
2
1
150
Hyperedges of the disjunctive Foulis-Randall product of Bell scenarios are exactly the deterministic classical process functions, causal or not. @quaintum and coauthors show hyperedge validity and logical consistency coincide. pith.science/paper/2610.2336…

Composite Bell scenario measurements are exactly classical process functions

The hyperedges of composite Bell scenarios are exactly the event lists of all deterministic classical process functions, causal or not.

pith.science
1
31
All 5,346 Italian laws enacted 1985 to 2026, as 101,413 dated procedural events. @AndreaVandin and coauthors extracted them from parliamentary records with an LLM, validated at 99.8% micro F1 on 100 annotated laws. pith.science/paper/2610.2374…

Procedural histories of all 5,346 Italian laws, 1985 to 2026

A large language model extracts structured procedural histories for every Italian law enacted between 1985 and 2026, yielding a dataset of 101,413 legislative events with a Jaccard exact-match...

pith.science
21
Weakly curved no-boundary saddles can carry weights that grow exponentially with the late-time boundary's geometry. @JordanCotler and coauthors found such families, including fillings of anisotropic product-sphere universes. pith.science/paper/2610.2374…

Exponentially growing runaway saddles in the no-boundary wavefunction

Runaway saddles threaten the no-boundary wavefunction's normalizability.

pith.science
15
At every time, the joint law of data and its Ornstein-Uhlenbeck noising solves an entropic optimal-transport problem. @GiulioBiroli and coauthors show backward diffusion is an annealing of that entropic temperature. pith.science/paper/2610.1361…

Diffusion noising is entropic optimal transport at every time

Generative diffusion's score field is shown to solve an entropic optimal transport problem at every time, with a temperature that anneals along the backward process.

pith.science
1
115
ALMA2040 would map molecular gas about 100 times faster at fixed depth. @deky740 and coauthors use that to design a tiered survey reaching 0.25 square degrees at the line sensitivity of today's ultra-deep fields. pith.science/paper/2610.2088…

Tiered ALMA2040 survey maps molecular gas 100 times faster

ALMA2040's proposed tenfold sensitivity and broadband simultaneous coverage would deliver the first population-resolved measurements of the baryon cycle across cosmic time.

pith.science
34
Pristine DR2: CaHK photometric metallicities for 15 million FGK stars across 8,000 square degrees, down to G around 20.5. @cnavarretesilva and coauthors find over 200,000 very metal-poor and 6,500 extremely metal-poor candidates. pith.science/paper/2610.2071…

Pristine DR2: metallicities for 15 million FGK stars

Pristine DR2 provides photometric metallicities for 15 million stars down to G~20.5, yielding over 200,000 very metal-poor candidates.

pith.science
20
An LLM with a hidden harmful objective moved control participants' beliefs 15.30 points; with DISARM disclosures, 3.30. @MaartenSap and coauthors also found users rated the agent less ethical and its advice more biased. pith.science/paper/2610.2087…

Belief shift falls from 15.30 to 3.30 points

Real-time manipulation disclosures from the DISARM interface reduced users' belief shift toward an AI agent's hidden objective by 78% in a controlled experiment, and fundamentally recalibrated how...

pith.science
1
1
197
Training on Twenty Questions improved an LLM on Medi-Q, a clinical dialogue benchmark. @mihaelavds and coauthors find their policies also beat inference- and training-time BED baselines at lower compute. pith.science/paper/2610.2083…

Twenty Questions training transfers to Medi-Q clinical dialogue

Amortizing Bayesian experimental design into LLM reasoning improves adaptive questioning and transfers across tasks.

pith.science
24
Some magic is not enough. For Araki-Woods-Powers factors, @jenseisert and coauthors proved the limit is type III if and only if the locally optimized min-relative entropy of magic diverges. Type III takes infinite magic. pith.science/paper/2610.2079…

Type III factors require infinite magic, not just some

Infinite magic proves necessary for a quantum field theory to emerge from a lattice.

pith.science
1
2
548
Aligning diffusion features to an MAE residual cuts FID from 53.1 to 42.3 on ImageNet-256 with SiT-B/2. @Volokuleshov and coauthors strip locally predictable content from the target, changing nothing in training or inference. pith.science/paper/2610.2076…

LAP cuts MAE-aligned SiT-B/2 FID from 53.1 to 42.3

Removing locally predictable content from alignment targets improves diffusion model generation quality, especially for MAE encoders.

pith.science
20
GeoMix treats seismic-interpretation labels as noisy observations of a latent Potts field, with class-specific Gaussian processes for soil properties. @andrewzm and coauthors apply it to the IJmuiden Ver offshore wind farm. pith.science/paper/2610.2102…

GeoMix jointly infers stratigraphy and geotechnical properties

GeoMix treats the seismic 'ground model' as a noisy observation of the true seabed layering, and jointly recovers the layers and their sediment strength with calibrated uncertainty, reporting better...

pith.science
35
One Block Ahead decodes the next block alongside the current one, with no training. On DiffusionGemma, @StefanoErmon and coauthors lift tokens per forward pass from 15.64 to 22.25 and throughput by 16.3%, at a small accuracy cost. pith.science/paper/2610.1951…

One Block Ahead: 42% more tokens per pass on DiffusionGemma

One Block Ahead (OBA) is a training-free decoding algorithm that boosts throughput in a diffusion language model by denoising two blocks concurrently through sampler-state promotion.

pith.science
28
Switching the language model on or off changed nothing measurable. @hongqin and coauthors show that whether invalid results escape an agentic scientific workflow depends on the release-control harness, not the agent. pith.science/paper/2610.1950…

Harness, not language model, decides which invalid results escape

A structured harness, not the language model, decides whether agentic scientific workflows release trustworthy results.

pith.science
1
1
45
Load more