Open-Source Models & Weights - Weekly Roundup

Decoy Direction Optimization (DDO), a fast, post-hoc weight-editing defense that requires no base-model finetuning, is introduced, based on a simple mechanistic insight: ablation attacks rely on contrastive estimators to find the refusal direction.

Probe of Internal Recognition (PIR) separates a model that will not answer from one that cannot, which supports sandbagging audits and unlearning verification, and extends from multiple-choice questions to free-form generation.

This study introduces TrustReviewer, an open-source LLM-based system for generating peer reviews of AI and machine learning papers and describes a concrete risk of recursive reviewer training and provides practical interventions for preserving judgment diversity and improving recommendation alignment in AI-assisted scientific evaluation.

A 202-scenario benchmark with Reference Decisions under a seven-class taxonomy is introduced and Decision Alignment and safety-specific error metrics are evaluated, finding structured LLM decisions are insufficient as a standalone safety mechanism and deployment requires an independent enforcement layer.

Inspired by the efficient communication principles of federated learning (FL), this work introduces two hybrid algorithms - FL+FSDP and FL+HSDP - interleaving sharded DP with FedAvg-style aggregations, demonstrating superior computational efficiency and improved model quality.

Probe of Internal Recognition (PIR) separates a model that will not answer from one that cannot, which supports sandbagging audits and unlearning verification, and extends from multiple-choice questions to free-form generation.

This study introduces TrustReviewer, an open-source LLM-based system for generating peer reviews of AI and machine learning papers and describes a concrete risk of recursive reviewer training and provides practical interventions for preserving judgment diversity and improving recommendation alignment in AI-assisted scientific evaluation.

A 202-scenario benchmark with Reference Decisions under a seven-class taxonomy is introduced and Decision Alignment and safety-specific error metrics are evaluated, finding structured LLM decisions are insufficient as a standalone safety mechanism and deployment requires an independent enforcement layer.

Inspired by the efficient communication principles of federated learning (FL), this work introduces two hybrid algorithms - FL+FSDP and FL+HSDP - interleaving sharded DP with FedAvg-style aggregations, demonstrating superior computational efficiency and improved model quality.

SwitchSD is introduced, an adaptive framework that treats copying as a latent control signal of the LLM that allows the system to dynamically switch between neural drafting and context-based copying, effectively turning copying from a noisy heuristic into a principled, model-aware decoding regime.

FP-Sketch is presented, a verifier that runs after an unmodified tensor core GEMM whose half precision operands are accumulated and delivered at FP32, and its constant depends on the BLAS and the operand format and must grow as $n^{2.57}$ for a square product.

These findings verify the original paper's claim that User Profile Recommendation achieves competitive performance under its test-set reranking protocol and makes recommendations more transparent.

SeetaPsych v1.0 is presented, an open source, unified and extensible computer vision toolkit designed to extract psychologically relevant signals from facial images and/or face based videos to offer an integrated and accessible visual measurement platform for research in psychology, behavioral science, human computer interaction, and related fields.

It is demonstrated that LLM annotation reproducibility can coexist with poor construct correspondence and motivate validation procedures that distinguish reproducibility, construct validity, and communication context variation when LLMs are used as measurement instruments.

The results strongly support the construct validity of 24 code quality metrics and reveal Coupling and Inheritance are revealed as multidimensional constructs, since they require measuring two different concepts, revealed as sub-categories in the analysis.

The experiments show that self-supervised pretraining and mark localization are the two most impactful factors, with learned cropping recovering the majority of the gain from manual cropping without requiring ground-truth annotations at inference time.

This paper presents an optimized, high-frequency Multispeculative PositMAC architecture for 32- and 64-bits Posit arithmetic, which achieves the highest operating frequency and reduces cycle time by up to 79.0\% with respect to 64-bit quire-enabled alternatives.

FloatLib is the first Lean library to unify IEEE binary and decimal arithmetic, arbitrary-width posits, P3109, and user-defined formats and rounding rules behind interchangeable certified software backends.

This paper proposes and analyzes Comparison-based Preference Optimization (ComPO), a zeroth-order alignment method based on comparison oracles, and establishes a convergence guarantee for its basic offline scheme under smoothness, gradient sparsity, and compatibility between the oracle and a latent objective.

BurnRiSc is presented, a framework that operationalizes the Oldenburg Burnout Inventory's two dimensions, exhaustion and disengagement, as 14 behavioral and linguistic signals computed from GitHub activity and scored against each contributor's own history, as evidence that burnout is screenable from public data.

This work finds that simple difference of means vectors coherently represent reward hacking in Kimi K3, GLM 5.2, and Qwen 3.8 Max, and provides evidence that simple, white-box methods can be used to scalably study and monitor reward hacking behaviors in frontier open source models.

It is found that aggregate pairwise comparisons transfer strongly across evaluation settings, including from randomly generated tasks to clinically meaningful tasks and across datasets, and at the same time, clinically meaningful tasks exhibit greater task-method interaction.

The findings show that even the presence of a small percentage of biased agents in such a system leads to significant shifts in the opinions of non-biased agents, and semantic analysis demonstrates that rhetorical consistency in textual explanations increases systematically with biased exposure and is partially decoupled from numerical convergenumericalutral agents.

A framework to measure the trust profile of an arbitrary LLM based on a variety of payload framings sent through different channels is presented, and cross-channel fragmentation attacks that distribute seemingly benign payloads across two or three channels are devised.

It is argued that local LLM systems need explicit guarantees for prompt lifetime, persistent storage, and tenant isolation, because the surrounding serving software handles prompt data before, during, and after inference.

An SSD-native local MoE inference system that addresses challenges with an SSD I/O pipeline optimized for expert delivery, a native three-tier storage hierarchy that delivers and retains experts dynamically, and balanced CPU--GPU hybrid execution.

An end-to-end computational pipeline for classifying intergovernmental agreements by institutional form and extracting financial relationships between principals and agents in service contracts is introduced, rendering the content of Iowa interlocal agreements analyzable at scale for the first time.

PentestChain is presented, a ten-phase automated penetration testing framework that couples a curated, deterministic exploit map with a cost-aware AI cascade-a local Ollama model, kept off the critical path by a deterministic backbone, sustains end-to-end operation at zero measured paid-API cost.

It is shown that benchmark identity alone is insufficient to reconstruct the evaluated task or determine the appropriate scope of comparison across reported repair rates, so a machine-readable schema for specifying each experimental setting is introduced to improve reproducibility and make the scope of cross-system comparisons explicit.

MeshKV, a KV cache fabric that moves blocks as packetized flows over a lightweight NoC, co-designs TaKV affine striping to spread homes and cut hotspot load and co-designs Mare multicast with verified duplicate suppression, and Pad, which converts bisection back-pressure into useful KV transfer.

Results show that resume-screening evaluations should assess not only whether a system identifies stronger candidates, but also whether those decisions remain stable when the same competence evidence is presented differently, and whether those decisions remain stable when the same competence evidence is presented differently.

TAME (Token Attribution and Masking for Emergent Misalignment), a three-stage framework: token attribution scores how strongly the fine-tuning update raises each response token's likelihood; signal characterization finds patterns among high-attribution tokens; and causal validation tests them by attribution-guided loss masking.

An exact decomposition under a specified stepwise maximal coupling is derived: the expected mismatch fraction equals a first-mismatch contribution plus post-divergence exposure multiplied by its mismatch rate.

It is shown that the minimum aggregation is essential: mean aggregation can be exploited, saturating the proxy reward while degrading accuracy, and the minimum aggregation is essential for PROSE, Process Reward Guided Self-Training.

The development of an existing fluidic loop digital twin at Centre Technique des Industries M{\'e}canique (CETIM) is reported on, using the characterization framework of Gil et al., providing a structured view across its lifecycle dimensions.

The new release of PyExoCross provides both a configuration-file-driven command-line interface (CLI) and a Python application programming interface (API) within a unified computational framework to provide a more general and extensible platform for large-scale spectroscopic simulations based on modern atomic and molecular databases.

Psi4-cbo, an open-source extension of the Psi4 program for simulations within the cavity Born-Oppenheimer Approximation (CBOA), incorporates light-matter coupling into standard electronic structure workflows and highlights that, since self-consistent treatments dress the cavity modes, satisfying experimental resonance conditions is method-dependent.

Decoy Direction Optimization (DDO), a fast, post-hoc weight-editing defense that requires no base-model finetuning, is introduced, based on a simple mechanistic insight: ablation attacks rely on contrastive estimators to find the refusal direction.

This work formalizes poison selection as oracle-budgeted set optimization and introduces SAILS (Set-level Audit-Informed Iterative Learned Selection), which learns a set scorer from a few hundred finetune-and-evaluate runs, ranks millions of candidate sets, and audits only a small shortlist.

The Deliberative Polling Diagnostic Framework is introduced, which compares human and LLM belief shifts after identical informational interventions and offers a concrete protocol: run the deliberative diagnostic before trusting LLM personas to mimic revised beliefs.

A schema-aware split learning (SL) framework that preserves privacy, using a large language model (LLM) as a shared semantic encoder to harmonize heterogeneous survey schemas across institutions and enables accurate, privacy-preserving, and resource-efficient collaborative learning from heterogeneous mental health survey data.

The findings support the claim that BF is a more structured and automation-friendly framework than CWE, and exploration reveals specific gaps in BF, including under-specified guidance on attributes.

PICO provides a common domain-specific abstraction and software layer through which experimental configurations and actions can be controlled across different user interfaces that enables the same physical laboratory to support remote interactive use, scripted experiments, autonomous routines, and features such as version control.

This work predicts single-sequence model throughput from GGUF metadata using roofline-shaped predictors with quantization-specific scale factors fitted on reference models with fitted efficiencies not universal.

Realtime-Venus, a proactive full-duplex interaction system with two separately trained 9B models: Realtime-Venus-Omni for audio-visual interaction and Realtime-Venus-Audio for spoken interaction, which leads the compared models on MMAU, Llama Questions, and Speech CMMLU.

The results suggest that fine-tuning is not just about how much a model changes, but how that change is spent, and that changing the accessible directions can qualitatively alter the outcome of fine-tuning.