Pith. (@PithSci) header
Pith. (@PithSci) profile picture

Pith.

@PithSci

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

 Austin, TX
Joined July 2026
  • Tweets328
  • Following36
  • Followers220
1Media
Photo by Pith. (@PithSci)
Pinned Tweet
Pith. (@PithSci) profile picture
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
4
2,339
Boundary and distance maps computed from existing masks work as free auxiliary tasks. @ten_photos and coauthors frame this as bilevel optimization in ALBO, with gains on REFUGE, ACDC and Hippocampus under limited labels. pith.science/paper/2610.2695…
Free auxiliary labels from masks improve limited-label segmentation

Free auxiliary labels from masks improve limited-label segmentation

Auxiliary learning with automatically derived labels, structured as a bilevel optimization and bidirectional feature exchange, consistently improves medical image segmentation under scarce annotati...

pith.science
3
Attention resists LUT-aware training: it multiplies two activations, leaving no weight to tabulate. @mariaspiropulu @ChangSun1025 and coauthors tabulated its constant products, beating the prior best by 1.8 points with 2.9x fewer LUTs. pith.science/paper/2610.2707…
LUT-aware attention: 1.8 points more accuracy, 2.9x fewer LUTs

LUT-aware attention: 1.8 points more accuracy, 2.9x fewer LUTs

LEAF introduces a hybrid LUT-arithmetic transformer that improves jet-tagging accuracy by 1.8% while using 2.9× fewer LUTs than prior sub-microsecond FPGA transformers.

pith.science
2
123
Halcyon-1.6B, trained from scratch on 100M image-text pairs with no language pretraining, beats JiT-T2I-XXL/8++ (35.3 vs 41.0 gFD) with 275 times less pretraining compute. @jcniebles @eadeli and coauthors release code and models. pith.science/paper/2610.2686…
Halcyon-1.6B beats JiT-T2I-XXL/8++ with 275 times less compute

Halcyon-1.6B beats JiT-T2I-XXL/8++ with 275 times less compute

Strong text-to-image models can be trained from scratch without internet-scale language pretraining.

pith.science
26
Using 27 REBELS galaxies detected with ALMA, @ferrara_sns and coauthors derived the first [CII] luminosity function at redshift 7: characteristic luminosity near 1.3 billion solar luminosities, faint-end slope of -2.41. pith.science/paper/2610.2705…
First [CII] luminosity function at redshift 7 from REBELS

First [CII] luminosity function at redshift 7 from REBELS

The first [CII] luminosity function at z~7 derived from ALMA REBELS data shows that UV-selected surveys miss 60-90% of [CII] emitters, particularly at the faint end where dust-obscured galaxies...

pith.science
3
215
Four datasets purport to map the same parks: OpenStreetMap, Esri, SafeGraph, The Trust for Public Land. @gboeing and coauthors show park access results shift widely between them, with SafeGraph especially poor. pith.science/paper/2610.2739…
Four park datasets, same features, widely different access results

Four park datasets, same features, widely different access results

Park dataset choice can shift underserved tract estimates by 34 percent

pith.science
4
60
Static analysis alone resolved only 36.5% of WebView bridge entries in 338 F-Droid apps. @abhishektiwari and coauthors added a static-guided dynamic phase that lifts it to 74.5%, covering 115 apps that static analysis left empty. pith.science/paper/2610.2728…
Guided dynamic phase doubles WebView bridge resolution to 74.5%

Guided dynamic phase doubles WebView bridge resolution to 74.5%

MifDroid is a hybrid static-dynamic framework that detects cross-language information flows through Android WebView bridges by constructing per-WebView call graphs and projecting host-side effects...

pith.science
32
FLINT pools both queries and documents into a fixed number of vectors via attention, cutting index size 5.3-9.3x versus ColBERTv2. @matei_zaharia and coauthors report 1.5-2.8x faster retrieval than ConstBERT on long queries. pith.science/paper/2610.2727…
FLINT: 5.3-9.3x smaller indices than ColBERTv2

FLINT: 5.3-9.3x smaller indices than ColBERTv2

FLINT applies symmetric attention-based pooling to both queries and documents to fix the number of vectors per query and document, reducing index size by 5.3–9.3× and speeding retrieval by 1.5–2.8×...

pith.science
1
30
Speaker identity alone can trigger a backdoor. @yvesalexandre and coauthors report 90.2% and 97.0% success on Whisper-medium and Parakeet, with 1.7% and 0% false positives, from two minutes of the target's real audio. pith.science/paper/2610.2721…
A speaker's natural voice works as an ASR backdoor trigger

A speaker's natural voice works as an ASR backdoor trigger

A speaker's natural voice can serve as a backdoor trigger in speech models, requiring no attacker intervention at inference time.

pith.science
1
23
Information talk in black hole physics has drawn philosophical objections. @SebastianDeHaro and coauthors argue the Hayden-Preskill model avoids them: it rests on no epistemic ignorance and does not treat information as a substance. pith.science/paper/2610.2818…
Hayden-Preskill avoids old objections to black hole information

Hayden-Preskill avoids old objections to black hole information

The paper argues that quantum information theory gives black hole physics a legitimate operational framework: the Hayden-Preskill model treats a black hole as a quantum channel that scrambles,...

pith.science
1
60
HybridResearcher, a 4B agent, gains 30.0 points pass@8 over Qwen3.5-4B on HybridDeepResearch and beats Qwen3.5-397B-A17B by 8.3. @mohitban47 and coauthors trained it on synthesized database-and-web bridge trajectories. pith.science/paper/2610.2489…
4B HybridResearcher beats Qwen3.5-397B-A17B on HybridDeepResearch

4B HybridResearcher beats Qwen3.5-397B-A17B on HybridDeepResearch

A 4-billion-parameter agent is trained to bridge SQL databases and web search, outperforming a model 100x larger.

pith.science
1
123
One model per contrast beat the isotropic 2x specialist and led on LPIPS, DISTS and FID in every geometry, though specialists keep higher PSNR in anisotropic settings. @maierhein and coauthors call it GSMR3D. pith.science/paper/2610.2485…
Single model beats isotropic specialist, leads perceptual metrics

Single model beats isotropic specialist, leads perceptual metrics

GSMR3D uses learned anisotropic Gaussian prototypes to perform arbitrary-scale 3D MRI super-resolution with one model per contrast, outperforming specialists on isotropic upsampling and better...

pith.science
2
525
AI constitutional assemblies produce clauses with no equivalent in human law, yet the creativity has a ceiling. @GBR_Data and coauthors find most proposals recombine human constitutional concepts with AI-specific vocabulary. pith.science/paper/2610.2466…
AI-drafted law is novel in vocabulary, recombinational in substance

AI-drafted law is novel in vocabulary, recombinational in substance

Multi-agent LLM debates converge on a robust floor of legal provisions but reveal that AI creativity is largely recombinational, not transformational.

pith.science
3
911
LTT-9779 and WASP-121 flicker by about 54 and 59 ppm from granulation, on timescales near 5 and 6 minutes. @ProfSaraSeager and coauthors, using JWST/NIRISS SOSS, find the amplitude falls with wavelength. pith.science/paper/2610.2455…
Granulation amplitude falls with wavelength in two JWST hosts

Granulation amplitude falls with wavelength in two JWST hosts

JWST/NIRISS time-series spectrophotometry measures the wavelength-dependent decrease of stellar granulation amplitude in two Sun-like stars, matching 3D MHD simulations.

pith.science
47
LookAhead Defense cuts harmful agent behavior from up to 93.9% to at most 14.2% under benign fine-tuning of poisoned models. @mohitban47 and coauthors build its Safety Buffer from the released model's own safe responses. pith.science/paper/2610.2437…
LookAhead Defense cuts poisoned agents' harmful behavior to 14.2%

LookAhead Defense cuts poisoned agents' harmful behavior to 14.2%

LookAhead Defense reduces dormant poisoning activation in agents from 93.9% to 14.2% and in language models to 2.0%, using the released model's own safe behavior as a reference without external safe...

pith.science
1
262
Handing agents the annotated evidence directly still leaves accuracy below 60% on HABIT-Bench. @zhouhan_lin and coauthors show the harder step is inferring when a user habit applies, through exceptions and changes. pith.science/paper/2610.2452…
Even given annotated evidence, agents stay below 60%

Even given annotated evidence, agents stay below 60%

HABIT-Bench reveals a persistent gap between LLM agents' access to past interactions and their ability to infer and apply user habits, with the best memory method achieving only 33.61% macro accuracy.

pith.science
125
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
316
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

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
51
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

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
145
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

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
85
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 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
56
Load more