language models

2026-10-31

RESEARCH · 1 SOURCE · arXiv cs.AI

LibStructBench: document-grounded benchmark for sequencing-library structure reconstruction (arXiv:2610.37851v1)

The paper (arXiv:2610.37851v1) introduces LibStructBench, a document-grounded benchmark for reconstructing sequencing-library molecular structures and workflows from protocols. Experiments show modern agents often produce approximately correct reconstructions but are unreliable at exact molecular-state and workflow connectivity reconstruction; switching underlying language models had larger average effects than switching agent harnesses, additional knowledge aided oligo-family recovery in a small pilot, and autonomous accumulation of rules showed non-monotonic validation behavior without improving held-out workflow scores.

6.0

2026-10-22

MODELS · 1 SOURCE · The Decoder

Symantec: 'LLMjacking' agent hijacks unprotected language models to run self-sustaining cyberattacks

Security researchers at Symantec say a honeypot captured a malicious AI agent (dubbed "LLMjacking") that hijacks unprotected language models via a five-layer jailbreak, can execute up to 100,000 steps autonomously, scans for further targets, exfiltrates credentials and databases, and installs cryptominers (XMRig) for continued exploitation.

8.0

2026-10-12

RESEARCH · 1 SOURCE · arXiv cs.AI

Neural networks are more modular at the subspace level than single neurons suggest

The paper (arXiv:2610.12842v1) defines a hierarchy of modularity—unit-level, orthogonal, and linear—and presents a unified framework to learn and causally evaluate decompositions at each level via selective ablation. Across vision models, language models, and multi-task RNNs, the authors find that distinct computations often occupy separable subspaces despite overlapping neuron populations; these subspaces also differentially align with brain activity in visual and language cortex.

8.0

2026-10-05

RESEARCH · 1 SOURCE · arXiv cs.AI

Pluralistic Preference Optimization (PlurPO) paper proposes training LMs to reduce social sycophancy

The arXiv preprint (arXiv:2610.02568v1) introduces Pluralistic Preference Optimization (PlurPO), a method that has language models simulate relevant stakeholders and train to produce responses acceptable to all stakeholders using only the model's own signals (no ground-truth labels). Across four datasets and four model families, PlurPO substantially reduces social sycophancy—for example cutting endorsement of harmful-intent statements by 89% on average and narrowing the human–model endorsement gap on general advice from 17.8% to 8.0% on average; a preference dataset built with an 8B model also transfers to a 32B model.

7.0

2026-10-05

RESEARCH · 1 SOURCE · arXiv cs.AI

FD-SCoPE: verifiable, feedback-driven LLM framework for clinician queries over trial evidence tables

The paper (arXiv:2610.02576v1) introduces FD-SCoPE, a language-model framework that answers clinician-style questions over systematic-review evidence tables by exposing the query, selected trials, and derivation rule for each answer and learning from expert corrections. On an oncology table of 159 immune checkpoint inhibitor trial records it completed 140 clinician-style tasks (alternatives 90.7–97.9%), retrieved 99.3% of relevant records for derived attributes with 89.8% PPV, achieved derived-value F1 77.7% (vs. 64.8–73.4% for four baselines), and raised F1 on 1,201 unseen questions from 77.9% to 84.9% after simulated corrections.

6.0

2026-10-02

RESEARCH · 1 SOURCE · arXiv cs.AI

Paper shows adversary-robust signalling often coincides with prior 'salience pole' optimality

The arXiv preprint (arXiv:2610.00233v1) introduces an informed adversary into a constrained signalling/forced-choice task and shows that the signal that best defends truth under adversarial persuasion typically shifts from the Bayesian posterior-maximizer toward a salience (margin-maximizing) option. On 108 confirmatory items the robust optimum aligns exactly with the prior salience pole; across a 200,000-item pool they differ on 2,748 items, and 18.2% of items have an optimum that shifts under a finite persuasion budget—an overlap that creates a structural limit on empirical tests of adversary-awareness; the authors recommend a cheap diagnostic check to detect coincidence before evaluation.

6.0

2026-09-28

RESEARCH · 1 SOURCE · arXiv cs.AI

arXiv:2609.30328v1 — Two label-free tests to detect when multi-agent code judges lack grounding

The paper (arXiv:2609.30328v1) studies when a language-model-based judge of code is actually grounded and introduces two label-free measurements to detect when the judge lacks a basis for its verdict. Evaluating the MARCH multi-agent verification pipeline on two code-judging benchmarks across 80 condition-by-cell measurements, the authors find MARCH declares both solutions equally good on 78–95% of comparisons and attains 4.4% accuracy (versus 43.7% when the same model is asked directly); using one log-derived measurement to gate answers increases accuracy from 20.7% to 36.9% while still answering about half of comparisons.

7.0

2026-09-15

MONEY · 1 SOURCE · The Decoder

Gates Foundation pledges at least $1B to expand AI in health, education and agriculture after Gates warns of AI dangers

The Gates Foundation will invest at least $1 billion over two years to make AI tools more widely available in health, education, and agriculture, according to reporting in The Decoder. Bill Gates warned that over 90% of training data for early language models came from English sources, noted speech recognition fails about 60% of the time in Yoruba, and said the market is "a terrible guarantor of equal opportunity."

7.0

2026-09-04

MODELS · 1 SOURCE · Ars Technica

Invisible Unicode technique once used to attack AI is now used by spammers, Ars Technica reports

Ars Technica reports that a previously overlooked block of invisible Unicode characters — used in the past to evade or attack AI systems — is increasingly being adopted by spammers to bypass filters and conceal content from humans and automated detectors.

6.0