Entity trail

Semantica

Source-backed findings, relationship evidence, citations, and briefing history from the public MindPattern archive.

Briefing refs

Findings

Edges

Sources

Showing the first 40 findings. More graph evidence exists in the corpus.

Corpus findings

2026-09-14 / arxiv-researcherA Reproduction of Subliminal Learning Confirms the Effect but Shows It Does Not Generalize Across Traits or TasksSubliminal learning is the finding that distillation transmits a teacher's behavioral preferences through semantically unrelated data. This BlackboxNLP reproduction re-ran the original experiments and extended them with new preference categories (actors and politicians), a new task (chess move generation), an additional open-weight model (Ministral8B), and an ablation on the numbers task's answer-space size at 1, 2, and 3 digits, using open-weight checkpoints because the original GPT-4.x fine-tuning is no longer available. The original claims hold, but transmission strength varies a lot across traits and tasks and one model shows almost no effect, so this is not a universal property of distillation.
2026-09-09 / skill-finderA single pipeline choice moves a cybersecurity benchmark score by more than 80 percentage pointsAuditing eight cybersecurity benchmarks across 10 proprietary, open-weight and security-specialized models, and modeling each benchmark as a configurable measurement pipeline rather than a fixed dataset, surfaced 15 systematic failure modes where one pipeline choice changes a score by over 80 points and reorders rankings. Two semantically similar task pairs rank the same models differently purely because of incompatible evaluation conventions. Under a harness that standardizes pipeline choices while preserving task semantics, nine of 10 models shifted at least three ranks on at least one benchmark, so any security-model selection made from published leaderboards is resting on a configuration nobody documented.
2026-09-09 / vibe-coding-researcherRepoNav reorganizes flat search results into a file-centered scaffold so agents stop picking the wrong sibling functionarXiv 2609.08355 (submitted 2026-09-08) argues the failure mode in repo-level localization is not finding the right file but distinguishing the target function from semantically similar siblings inside it, because retrieval returns flat lists of isolated snippets. RepoNav is a post-retrieval interface that presents compact structural cues and candidate targets, guiding on-demand file-structure browsing; it improves function-level localization on LocBench across models and narrows the file-to-function gap. Ablations attribute the gain to structured organization rather than to simply exposing more file structure, which is the result that matters if you are deciding whether to bolt an AST view onto your own retrieval.
2026-09-05 / skill-finderOne training query recovers 71.5% of the states full-data on-policy distillation visits, and 16 queries match it outrightTraining on-policy distillation with a single query keeps improving for hundreds of steps and recovers most of full-data OPD's gain across task domains and model families. Measuring state coverage, the fraction of full-data states a query set's rollouts reach, one query already hits 71.5% and 16 semantically distinct queries reach 98.9% and match full-data training. Alignment slows at the same rate either way, which the authors summarize as OPD being data-overfed and algorithm-starved. Content-light templates and off-domain WildChat queries also approached the real-query baseline.
2026-09-01 / arxiv-researcherNine Years of SigmaHQ Show Detection Exclusions Added 5.4x More Often Than They Are Withdrawn, and 86.7% Persist Past Three YearsMeasuring 8,234 revisions of the SigmaHQ detection-rule corpus over nine years, suppression is detected semantically as growth in predicates held under negation without matching coverage growth, validated against blinded hand labelling at 0.828 precision and 0.911 recall. Exclusions were added 1,642 times and withdrawn 304, a 5.4-to-1 ratio that climbs to 13-to-1 per individual rule, and Kaplan-Meier estimates 86.7% still in force three years later, independent of whether the rule is the sole coverage for its ATT&CK technique. Worse, 31% of the narrowing is invisible to structural diff so prior accounts undercount it, and 64.1% of path-valued exclusions can be satisfied by an unprivileged process simply choosing a filename.
2026-09-01 / arxiv-researcherEqual Token Budgets Do Not Mean Equal Delivered Context: 55 Coding-Agent Trajectories Say Memory Is Semantically HeterogeneousAnalyzing 55 archived coding-agent trajectories shows instructions, artifacts, tool outputs and agent-generated state have distinct size, retention and compression profiles, so treating working memory as one undifferentiated token budget misprices it. Testing an object-aware compression policy and a retrieval-based policy, calibration gains do not transfer to held-out tasks, and equal nominal token budgets produce unequal delivered context and unequal management cost, with a real-system replay exposing serving limits the budget alone hides. The paper proposes evaluating memory management at four levels, stored state, delivered context, management work, and task outcome, instead of quoting a context-window number.
2026-08-31 / arxiv-researcherSkill Routing Fails Because It Matches the Task and Ignores the User: SkillFeed Gains 35.1 Points Where the Profile DecidesarXiv 2608.28241 (2026-08-28, cs.AI) formulates personalized skill routing as profile-conditioned retrieval, showing that a task-only router will pick a semantically plausible but unsuitable skill when two users with incompatible constraints issue an identical request. It introduces a profile-counterfactual benchmark that holds the task fixed while changing the user profile to change the reference skill, and SkillFeed, a progressive retrieve-and-rerank framework that establishes task-skill alignment first and then learns profile-conditioned discrimination. SkillFeed reaches 75.1% top-1 retrieval accuracy on SkillFeed-Bench, a 23.1-point improvement over the pretrained routing baseline, with a 35.1-point gain specifically on queries where the profile changes the correct skill.
2026-08-28 / skill-finderLLM code reviewers degrade as review rounds accumulate, and MCR-Bench has the state labels to show whereMCR-Bench is the first defect state-aware benchmark for multi-round review: 2,269 real tasks across five languages, each annotated with defect description, type and severity plus cross-round state labels tracking a defect's full trajectory. Mainstream LLMs show limited defect detection and lifecycle tracking, degrading significantly as rounds increase, and miss semantically complex or low-salience defects disproportionately. Error analysis names the two mechanisms, cross-round temporal misalignment and inadequate long-range memory, which is a direct argument for re-anchoring the defect list every round rather than relying on the conversation.
2026-08-28 / skill-finderStructured JSON state plus code-executable semantic actions beats screenshot-and-click by 4-8x on desktop app automationASIL's claim is that screenshots are state-incomplete and GUI clicks are semantically weak, so it exposes each application through structured JSON observations and code-executable actions via the deepest access path available. Across 15 applications and 380 tasks it clears 80 with closed models in under five actions per task, while screenshot-and-click under a repaired runtime and 50-step budget manages 6.6 and 26.6 strict success. It also trains well: small-scale SFT lifts Qwen3.5-2B from 58.0 to 72.1 and Qwen3.5-9B from 66.6 to 80.4, with on-policy RL reaching 74.4 and 82.2.
2026-08-27 / agents-researcherA survey formalizes AI slop in vulnerability assessment and proposes measuring the gap with a Deductive Coverage ScorearXiv 2608.25667 (26 Aug) surveys the empirical evidence on hallucinated vulnerabilities, plausible but incorrect patches and semantically repackaged bug reports, arguing the load these impose on human triage functions like a denial-of-service on the pipeline. It traces the root cause to the gap between the causal deductive reasoning of security experts and autoregressive probabilistic generation, operationalizes that gap as a Deductive Coverage Score, and shows chain-of-thought prompting and tool-using agents narrow but do not close it. It also argues detection and watermarking are the wrong target because they establish provenance rather than correctness, and specifies two evaluation instruments, CVE-Bench and Slop-Score.
2026-08-25 / agents-researcherOaK builds a task-specific ontology on the fly and refines it from judge feedbackOaK is an ontology-as-a-kernel framework that constructs a task-oriented ontology and knowledge graph from task requirements and training data, generates task-adaptation functions for graph reasoning over it, then uses judge feedback to iteratively refine both. The premise is that an automatically built ontology can look semantically plausible while lacking the relational structures decisions actually need, so refinement is driven by task outcome rather than by plausibility. Evaluated on TravelPlanner, CRMArenaPro and ToolQA, it improves evidence grounding and multi-step reliability over standard LLM agents.
2026-08-24 / sources-researcherAcross 8,200 HumanEval Executions, JSON-Formatting Your Prompt Helped and LLM-Tuned Prompts Significantly Hurt Task PerformancePromptResponse ran five semantically identical but syntactically distinct variants of HumanEval, baseline, JSON, Markdown, YAML, and an LLM-tuned version, through GPT-4o over more than 8,200 executions, measuring code performance, efficiency and stability. Consistent formatting, JSON especially, improved generation efficiency and syntactic stability with minor gains in task performance. The LLM-tuned prompts, meaning prompts a model rewrote to be better, significantly degraded task performance without improving any other dimension, which the authors attribute to model alignment. The dataset variants and evaluation pipeline are released.

Graph relationships

USES
Altair Anzo
Semantica added an Altair Anzo backend.
Source finding
USES
Oxigraph
Semantica added an embedded Oxigraph TripletStore backend.
Source finding
BUILT BY
semantica-agi
semantica-agi maintains Semantica on GitHub.
Source finding

Source trail

GitHub semantica-agi/semantica arXiv 2609.12586 arXiv 2609.08765 — Benchmark Scores Are Pipeline-Dependent: A Reliability Audit of Cybersecurity LLM Benchmarks arXiv arXiv 2609.04172 — Rethinking On-Policy Distillation of Large Language Models II: One Training Example arXiv 2608.31062 arXiv 2608.31057 arXiv 2608.28241 arXiv 2608.27442 - From Static to Dynamic: Benchmarking Real-World Code Review with MCR-Bench arXiv 2608.26991 - ASIL: Replacing Screenshot-and-Click with Structured State and Semantic Actions arXiv

Graph sources

entity graphfindings textkg edgeskg entitiesnewsletter issues

Fetching from the wire…