THE INTERNET IS THINKING
Earlier sources: Language models
Language models learn patterns in text and other representations to predict or generate sequences. Follow model releases, training methods, reasoning, evaluation, and deployment research; a benchmark result is specific to its test conditions.
SYMTRI organizes observations from public sources into a living map. Connections reflect shared subject matter; they do not establish causation or endorse a source.
Open in the map ↗459 classified signals in the fourteen-day window. These counts include signals awaiting readable evidence.
Earlier sources
Up to ten sources with readable evidence, ordered by their feed timestamp. These archive sources can be older than the live fourteen-day window. This list is a sample, not a measure of total activity. Dates distinguish paper publication, repository creation, model/dataset activity, Hacker News posts, and KEV catalog additions. “First archived” records when SYMTRI first stored the signal.
- arXiv · Published · First archived
Pruning for Efficiency, Paying in Fairness: Demographic Disparities in Pruned Speech-LLMs ↗
Speech-LLMs are expensive to run, making compression important for real-world deployment. However, compressed models are usually selected using aggregate word error rate (WER), which can hide how pruning affects different demographic groups. In this work, we systematically study the effect of audio encoder pruning on SLAM-ASR for different demographic groups. Using the Fair-Speech and Common Voice datasets, we found that the pruning does not affect all demographic groups equally; the gap between best- and worst-performing groups increases in fold. These disparities appear across all three enco…
- Hacker News · Posted on Hacker News · First archived
Jtoken – lossless JSON compression for LLM prompts ↗
Hacker News discussion.
- arXiv · Published · First archived
Information Games: Strategic Crowding and Firm Repositioning in Language-Model Space ↗
Firms follow changing economic opportunities, but rivalry changes their response. We develop ESCAPE, a rational-share game of distribution-valued positioning with heterogeneous capability costs, establish a unique equilibrium, and derive an exact reallocation restriction separating opportunity and crowding contributions. Competition need not push firms apart: rational firms can enter increasingly crowded regions, because the full reallocation gain can be positive even when its crowding component is negative. Rivalry instead changes how differently firms respond to the same opportunity shift. I…
- arXiv · Published · First archived
Selective Channel Restoration for Backdoored Vision-Language Models ↗
Vision-language models (VLMs) exhibit strong multimodal capabilities but remain vulnerable to backdoors implanted through poisoned fine-tuning data. Existing defenses often require extensive parameter updates during fine-tuning or incur per-query overhead during inference. To address these limitations, we propose Perturb-Select-Restore (PSR), a post-training defense that performs sparse updates to the projection interface and introduces no additional computation during inference. We reveal that backdoored VLM projectors are substantially more sensitive to bounded perturbations than clean VLM p…
- arXiv · Published · First archived
Where Do LLMs Decide to Break the Rules? Mechanistic Localization of Prompt Injection Compliance ↗
When a prompt injection attack succeeds, a Large Language Model (LLM) abandons its assigned system role to comply with an adversarial instruction. While prior work has extensively quantified how often this occurs, we ask a more fundamental question: where inside the network does the model actually decide to break the rules? Using layer-by-layer causal activation patching across five models (4B to 32B parameters), we find a clear dissociation: attack information is linearly decodable from the first layer, yet causal leverage over the model's behavior is negligible until a late-layer bottleneck…
- arXiv · Published · First archived
Complexity-Aware Evaluation of LLM Comprehension ↗
Large language models (LLMs) are increasingly used for software engineering tasks that require understanding existing source code, including behavior prediction, function explanation, debugging, and code review. However, aggregate benchmark accuracy can conceal how model reliability changes as source code becomes structurally more complex. This paper presents a complexity-aware framework for evaluating LLM code comprehension using cyclomatic complexity, nesting depth, branching factor, and Halstead volume. We evaluate DeepSeek-Coder-V2 and Llama through two complementary tasks: automatic input…
- arXiv · Published · First archived
Confidence-Guided Protocol IR for LLM-Aided Security Protocol Modeling ↗
Large language models offer a promising interface for translating natural-language protocol descriptions into formal security models, but their outputs remain difficult to trust without expert validation. In this paper, we present a human-in-the-loop framework for generating Tamarin-verifiable formal models of security protocols. Our key observation is that the main correctness bottleneck is the semantic accuracy rather than the syntactic validity of the intermediate protocol representation. To address this problem, we introduce a protocol intermediate representation (IR) that serves as a huma…
- GitHub · Repository created · First archived
SerendipityOneInc/instinct ↗
Instinct: single-forward-pass decision models (choice / true-false / score) returning a probability per candidate
- Hacker News · Posted on Hacker News · First archived
Cybersecurity harness for full-stack LLM-driven penetration testing ↗
Hacker News discussion.
- Hacker News · Posted on Hacker News · First archived
Language models for text classification: From bag-of-words to Jev ↗
Hacker News discussion.
Live window recorded . This is the ingestion observation time, not the publication or discovery date of each source.
How to assess evidence about Language models
Check each original source for its methods, release status, limitations, and supporting data. A repository announcement, a preprint, and a published study provide different kinds of evidence. SYMTRI’s subject classification helps discovery; it does not independently validate the source’s claims.
See the collection and counting methodology ↗ · Read the weekly synthesis ↗
Related threads
Explore related regions
These navigation suggestions are part of the map’s editorial structure, rather than measured evidence of a connection.