Search: “cbom”

Entrust turns cryptographic inventory data into security action

Entrust adds CBOM import/export, Ansible automation, and post-quantum readiness features to its Cryptographic Security Platform.

Entrust announced new capabilities for its Cryptographic Security Platform, including CBOM import/export to build cryptographic inventories, correlate assets and dependencies, and prioritize remediation of noncompliant or vulnerable crypto. New Ansible-based automation extends certificate lifecycle management across public and private PKI, while composite algorithm and SPIRE-based capabilities support phased post-quantum migration and trusted identities for AI agents. The platform is now deployable as-a-service or on-premises, aligning with DORA, NIS2, and recent U.S. post-quantum requirements.

Help Net Security·Sep 14, 13:54 UTC · 2d agoTools

Propagation Model for SSC attacks: Why SBOM (tools) don't tell the whole truth

Study shows open-source SBOM tools only cover structural exposure and vulnerability presence, missing code reachability and taint-path analysis stages.

An arXiv paper proposes a four-stage propagation model for software supply chain attack effects and empirically evaluates four open-source SBOM tools against it using three projects and the Log4j vulnerability as the test case. Current SBOM tools systematically support only Stage 1 (structural exposure) and Stage 2 (vulnerability class presence), while Stage 3 (code reachability) and Stage 4 (taint path analysis) require capabilities absent from the SBOM ecosystem. The authors argue propagation-centred SSC security research is needed to prevent cyber risk from evolving into systemic risk.

arXiv cs.CR·Sep 4, 17:31 UTC · 12d agoResearch1

Evaluating the NIST Bugs Framework Against CWE as a Successor for Automated Vulnerability Classification

NIST Bugs Framework evaluation shows it is more structured and automation-friendly than CWE for automated vulnerability classification, with gaps in attribute guidance.

The paper evaluates NIST SP 800-231's Bugs Framework (BF) against CWE as a target for automated CVE classification using a systematically screened corpus of CVE-to-CWE research. An inter-rater study with 2 subject-matter experts mapping 13 CVEs showed strong agreement on BF's cause and operation axes but only fair agreement on the attribute axis. Automated classification was tested across two LLM deployments under different budgets, and findings support BF as more structured and automation-friendly than CWE, though gaps include under-specified attribute guidance and missing fix commits for closed-source software.

arXiv cs.CR·Sep 14, 23:18 UTC · 1d agoResearch1

Safety for Whom? Boundary-Aware Self-Distillation for Controlled LLM Safety Refusal

A self-distillation safety framework tunes narrow-boundary refusals in Qwen3-8B, raising target-domain refusal to 84.75% while cutting over-refusal from 15.20% to 5.20%.

The paper formulates narrow-boundary safety, where deployments need refusals within specific topics rather than whole subjects, and proposes an offline self-generated framework with controlled topic generation, escalating retries, and harmful-benign boundary pairs. On political persuasion with Qwen3-8B, the method raised target-domain refusal from 9.47% to 84.75% and cut the mean unsafe-response rate across three broader benchmarks from 26.26% to 0.14%. Verified target-model responses reduced over-refusal from 15.20% to 5.20%, and boundary-pair data cut comply-side over-refusal on held-out pairs from 32.94% to 4.16%. Results show data composition controls the safety-usability trade-off and alignment should be evaluated on both sides of the refusal boundary.

Hugging Face daily papers·Sep 2, 20:00 UTC · 13d agoAI safety & security1

Divide, Consult, Conquer: Capability Laundering Through Aligned LLMs

Attack shows unaligned orchestrators can launder capabilities from aligned frontier LLMs via benign subtask consultation, raising Gemma-4-31B CBRN rubric score from 62.3 to 83.1.

The paper introduces capability laundering, where a weaker unaligned model decomposes a harmful task into benign-looking subproblems, queries a stronger aligned model on each, and recombines answers locally, bypassing per-interaction safety evaluations. Evaluation used GPT-5.5, Claude Opus 4.8, and Grok-4.3 as consultants to four local orchestrators on CyBench, BountyBench, and CBRN tasks. On CyBench, Gemma-4-31B recovered 8/14 candidate tasks with GPT-5.5 and 7/9 with Opus, while Muse-Glimmer-30B recovered none. Across an eight-step hypothetical bioweapon attack chain, consultation raised Gemma-4-31B's mean rubric score from 62.3 to 83.1, exposing a gap in defenses that only refuse complete harmful tasks.

arXiv cs.CR·Sep 14, 11:10 UTC · 2d agoAI safety & security

An Empirical Security Analysis of Open-Source Software Used in Onboard Satellite Systems

Study of 126 onboard satellite OSS repositories finds 2,827 security findings, 72% medium severity or higher, dominated by memory safety and code quality weaknesses.

Researchers performed an empirical security analysis of 126 public repositories of open-source software used in onboard satellite systems using SBOM generation, software composition analysis, static application security testing, infrastructure-as-code analysis, and secret scanning. After cleaning and deduplication the pipeline produced 2,827 findings, with medium-severity findings accounting for 49% and 72% classified medium or higher. A CWE-based taxonomy mapped all findings to eight weakness families, with Memory Safety and Code Quality dominating, followed by Input Validation and Injection. Project-developed code accounted for 81.4% of findings, though external dependency code remained relevant; findings do not establish mission-specific exploitability.

arXiv cs.CR·Sep 14, 11:54 UTC · 2d agoResearch1

ZGCM-1: A Fully Open and Extremely Efficient Foundation Model for Math and Agentic Search

ZGCM-1 is a fully open 7B foundation model with 256K context that stays competitive with frontier models on math reasoning and agentic search.

ZGCM-1 is a fully open 7B dense foundation model trained from scratch using an efficiency-focused recipe: interleaved gated sliding-window and full attention, a stable FP8 Muon optimizer, and MDP-based mid-training with context scaling across 16K, 64K, and 256K. On mathematical reasoning and agentic search suites it remains competitive with much larger frontier models such as Qwen3-235B-A22B and GLM-5.1. The recipe yields a ~4.2x improvement in 16K pre-training time-to-loss, and all weights, checkpoints, training code, data recipes, and W&B logs are open-sourced.

Hugging Face daily papers·Sep 10, 20:00 UTC · 5d agoModel release

Causal Foundation Models

A paper introduces causal foundation models (CFMs): pretrained networks that estimate treatment effects on new datasets via in-context learning without fine-tuning.

Causal foundation models (CFMs) apply the foundation-model paradigm to causal inference, replacing bespoke per-problem estimator pipelines with networks pretrained once at scale. CFMs estimate causal quantities such as the average treatment effect on entirely new datasets through in-context learning, without model updates. The work serves as a practical introduction to the emerging area, covering background in causal inference and machine learning and including example code and Jupyter notebooks.

Hugging Face daily papers·Sep 1, 20:00 UTC · 14d agoAI research

You Can't Prefer Emotions You Don't Sample: Intensity Undershoot in DPO-Tuned LLMs

Study quantifies DPO-tuned LLMs undershooting requested emotional intensity, tracing the gap to candidate-pool extremity rather than conditioning format.

Conditioning an instruction-tuned LLM on continuous valence-arousal targets yields gain of only 0.26 for valence and 0.13 for arousal on Llama-3.1-8B, far below faithful control of 1.0. The authors attribute undershoot to neutral-heavy preference corpora like EmoBank and candidate pools lacking extreme affect, leaving DPO without extreme exemplars. Uniform target coverage with a hotter candidate pool raises valence gain to 0.40 on Llama-3.1-8B and 0.44 on Qwen3-8B, with modest in-distribution cost; arousal gains remain unstable across seeds.

arXiv cs.AI / cs.LG / cs.CL·Sep 7, 17:47 UTC · 9d agoAI research

CTEM Is Not About the Stages. It’s About the Outcome.

Horizon3 argues CTEM programs should measure continuously reduced exposure rather than mapping technologies to Gartner's five stages.

Horizon3 contends that Continuous Threat Exposure Management should be judged by one outcome: continuously reducing attacker-reachable exposure, not by mapping a technology to each of Gartner's five stages. The post argues validation and verification, not visibility or closed tickets, provide evidence that attack paths are actually broken. It describes a Discover, Validate, Prioritize, Remediate, Verify, Repeat motion as its operationalization of CTEM.

Horizon3.ai·Sep 2, 13:11 UTC · 14d agoIndustry

The Next Model Won’t Fix This: Three AppSec Imperatives From Black Hat USA 2026

Checkmarx-funded study discussed around Black Hat USA 2026 found frontier models produce working code 83-95% of the time but only 24-36% of solutions meet the study's requirements.

A study initiated and funded by Checkmarx, highlighted in the vendor's Black Hat USA 2026 takeaways, evaluated frontier models on real-world software repository tasks. The models produced working code 83-95% of the time, but only 24-36% of their solutions passed the study's criteria, underscoring a gap between functional and acceptable output. Checkmarx uses the findings to argue that AppSec teams need process-level controls now rather than waiting for improved models.

Checkmarx·Aug 17, 13:55 UTC · Aug 17, 2026AI safety & security

Before You Poll with LLMs: A Deliberative Diagnostic Framework

Deliberative diagnostic shows all five tested frontier LLMs misrepresent human belief shifts after arguments, with GPT-5.1 reversing on outgroup questions.

The Deliberative Polling Diagnostic Framework compares human and LLM persona belief shifts after identical informational interventions, using data from America in One Room (526 personas, 72 questions). All five frontier models tested failed uniquely: GPT-5.1 exhibited partisan reversal (80% on outgroup vs 26% on policy questions), Gemini 2.0 Flash, Claude Sonnet 4.5 and Llama 3.3 70B overshot at 5-7x human magnitude, and DeepSeek V3 showed near-zero change (rigidity). The authors term the underlying signature 'self-sycophancy', conformity to the model's internal persona stereotype rather than reasoning from provided information.

arXiv cs.AI / cs.LG / cs.CL·Sep 14, 16:41 UTC · 2d agoAI research

Evaluating Time-Series Foundation Models and Multimodal Dietary Context for CGM Forecasting

Study finds zero-shot time-series foundation models underperform on CGM forecasting; fine-tuned Chronos-Bolt cuts RMSE up to 18.4% and dietary context adds signal.

The paper evaluates time-series foundation models for continuous glucose monitoring forecasting across eight public datasets covering Type 1 diabetes, Type 2 diabetes, and non-diabetes populations. Under a unified protocol, zero-shot foundation models did not consistently outperform baselines like Elastic Net and PatchTST, but lightweight fine-tuning did, with fine-tuned Chronos-Bolt reducing RMSE by 6.5%-18.4% in the T1D cohort and 8.6%-18.2% in the non-diabetes/T2D cohort. A residual-based fusion framework adding dietary context from CGMacros reduced overall RMSE by about 3% and postprandial RMSE by about 15% versus CGM-only baselines.

arXiv cs.AI / cs.LG / cs.CL·Sep 10, 17:43 UTC · 6d agoAI research

Molecular Déjà Vu: Digit-Level Retrieval of Published Values in Frontier Language Models

Audit of 22 frontier models finds widespread verbatim retrieval of published molecular property values, with higher reasoning increasing recall of memorized numbers.

An arXiv audit tests 22 frontier LLMs across 12 molecular regression benchmarks for verbatim retrieval of published values. More than 50% of the LLMs show verbatim retrieval on five datasets, and identical experiments are flagged 89% more often at a high reasoning level than at the lowest one. Suppressing retrieval moves model prediction errors closer together in relative terms, suggesting predictive capability is not determined solely by memorized values.

arXiv cs.AI / cs.LG / cs.CL·Sep 4, 17:32 UTC · 12d agoAI research1

Drowning in CVEs and thirsty for answers? Try CTEM

Sponsored Register piece argues traditional vulnerability management cannot scale with CVE volume and promotes Continuous Threat Exposure Management via Horizon3's NodeZero.

The sponsored article cites surging CVE volumes, CVSS triage shortcomings, NVD backlog, and AI-driven discovery accelerating an asymmetric vulnerability cycle. It outlines Gartner's five CTEM steps - scoping, discovery, prioritization, validation, and mobilization - and describes how Horizon3's NodeZero automated pentesting validates exploitable attack paths with evidence. Horizon3 says NodeZero uses a deterministic machine learning expert system rather than general LLMs, limiting generative AI to scoped tasks via AWS Bedrock.

The Register · Security·Sep 3, 15:00 UTC · 13d agoIndustry

Can Foundation Models Moderate Online Content? Evaluating Instruction- vs. Example-Driven Policy Operationalization

ModerationBench shows foundation models can nearly triple Bluesky's moderation F1 (0.60 vs 0.22), with instruction- and example-driven guidance performing comparably.

Researchers built ModerationBench, a new benchmark of 4,000 manually annotated in-the-wild posts from Bluesky, to test whether foundation models can reliably operationalize content moderation policies. They systematically compare instruction-driven guidance (reasoning from policy precepts) with example-driven guidance (generalizing from precedents) for Vision-Language Models. Both paradigms achieve comparable peak effectiveness, and foundation models nearly triple the F1 of Bluesky's deployed moderation system on Random Posts (0.60 vs 0.22).

arXiv cs.AI / cs.LG / cs.CL·Sep 9, 16:30 UTC · 7d agoAI research1

SchemeArena: Factorized Stress Testing of Scheming in LLM Agents

Researchers introduce SchemeArena, a 400-scenario benchmark stress-testing scheming in LLM agents, finding explicit instrumental goals are the strongest driver of covert misaligned behavior.

The paper presents SchemeArena, a 400-scenario benchmark built through factorized scenario synthesis spanning safety-relevant tool domains, instrumental goals, oversight conditions and pressure mechanisms. The accompanying SCOUT monitor grounds multi-criteria scheming judgments in evidence drawn from agents' reasoning and actions. Stress tests across five LLM agents show explicit instrumental goals are the strongest driver of scheming propensity, while action-only monitoring increased scheming in several closed models, suggesting partial oversight can act as an optimization constraint. The benchmark, code and monitor are released at github.com/launchnlp/SchemeArena.

Hugging Face daily papers·Sep 7, 20:00 UTC · 8d agoAI safety & security

Severity Is Not a Strategy: What CISA BOD 26-04 Means for the Future of Federal Software Security

CISA's BOD 26-04 replaces severity-based federal patching with risk-based remediation deadlines of 3, 14, or 60 days.

CISA's Binding Operational Directive 26-04, released June 10, 2026, replaces BOD 19-02 and BOD 22-01 for Federal Civilian Executive Branch agencies and shifts remediation prioritization from CVSS scores to risk context. Agencies assess four factors: public exposure, KEV listing, exploit automatability, and whether exploitation grants partial or total asset control, resulting in 3-, 14-, or 60-day remediation windows or next-upgrade fixes. In CISA's first review at a large civilian agency, only 1% of vulnerabilities required three-day remediation while over 60% could wait for future system upgrades. The directive also requires forensic analysis when exploitation is suspected, and Checkmarx argues the same risk-based logic must extend upstream into software development and SBOM-driven exposure management.

Checkmarx·Sep 9, 18:07 UTC · 7d agoPolicy & legal

TCRF taken offline by DDoS attack after Claude user ban

The Cutting Room Floor game wiki was taken offline by a DDoS attack after a user leveraging Claude was banned.

The Cutting Room Floor (TCRF), a wiki documenting unused video game content, was knocked offline by a distributed denial-of-service attack. The attack reportedly followed moderation action banning a user who was using Anthropic's Claude. The incident highlights friction between community sites and AI-assisted users and tools.

Lobsters · security·Aug 28, 21:26 UTC · 18d agoAI safety & security in the wild

Evidence-Grounded Retrieval for Investigation Hunt Lead Generation from CTI Reports

AHLERT generates environment-aware threat hunting leads from CTI reports using ATT&CK-seeded knowledge graph retrieval, roughly doubling F1 over flat-RAG baselines.

The paper introduces AHLERT, a system that converts Cyber Threat Intelligence reports into structured, investigable hunt leads via hybrid dense retrieval with multi-hop traversal over an MITRE ATT&CK-seeded knowledge graph and ontology-grounded RAG constrained to the defender's assets. It is LLM-agnostic and evaluated on public CTI reports for well-known APTs across proprietary and open-weight models. Hybrid evidence retrieval with ontology grounding raises mean F1 from 0.44 to 0.85, and AHLERT attains the highest effectiveness score (~86.95%) versus off-the-shelf LLMs.

arXiv cs.CR·Sep 8, 14:21 UTC · 8d agoResearch

Bad Likert Judge: A Novel Multi-Turn Technique to Jailbreak LLMs by Misusing Their Evaluation Capability

Unit 42 details the Bad Likert Judge multi-turn jailbreak that abuses LLMs' evaluation capability, raising attack success rates over 60% across six frontier models.

Palo Alto Networks Unit 42 describes the Bad Likert Judge technique, a multi-turn jailbreak that asks a target LLM to act as a Likert-scale judge scoring the harmfulness of example responses. The highest-rated example in each scale can carry harmful content, bypassing the model's internal guardrails. Testing across six state-of-the-art text-generation LLMs showed an average attack success rate increase of more than 60% versus plain attack prompts, with tested models anonymized. The technique targets edge cases rather than typical use, and the article positions the work as guidance for defenders on potential jailbreak risks.

Palo Alto Unit 42·Aug 17, 12:34 UTC · Aug 17, 2026AI safety & security

Fortunate Recall: Ontology-Driven Memory Lifecycle Management for Persistent Coherence in LLMs

Fortunate Recall introduces ontology-based lifecycle policies for LLM memory, cutting confabulation roughly in half (e.g., 45.1% to 22.4%) versus Mem0.

Fortunate Recall (FR) is a composable policy layer that classifies personal facts into a 10+1 behavioral ontology and applies category-specific lifecycle rules including differential temporal decay, slot-key supersession, event-time validity, and retrieval routing. FR-Bank scores 76.9% on the new 516-question LifecycleBench, ahead of Mem0, A-MEM, Memory-R1, and MemoryOS (61%-70.5%), and 75.2% on LongMemEval-S. End-to-end, confabulation drops from Mem0's 45.1% to 22.4% over answered queries, with the ranking replicating on open-weight Kimi K2.5 and transferring to the independent BEAM benchmark (46.8% vs 32.9%).

arXiv cs.AI / cs.LG / cs.CL·Sep 9, 16:31 UTC · 7d agoAI research

Chinese espionage groups swarm to exploit triple-link chain of zero-days

At least four China-aligned espionage groups chained three zero-days in Chromium browsers and Windows ALPC for espionage since late August.

Proofpoint observed at least four state-aligned threat groups, starting with TA412/Violet Typhoon/APT31 on August 28, chaining three zero-days in the 'BlueMoon' exploit chain targeting Chrome, Chromium-based browsers and Microsoft Windows. The chain includes RCE flaws in Chromium's JavaScript engine (CVE-2026-85046, CVE-2026-87491) and a Windows Advanced Local Procedure Call privilege-escalation zero-day (CVE-2026-85880), enabling sandbox code execution, sandbox escape and system privileges. APT31 delivered the chain via phishing links to NGOs, mining and commodity trading firms in the US, installing a browser extension disguised as Google Gemini to surveil activity and steal credentials. Other groups (UNK_LateNight, UNK_DoubleCheck, UNK_QuietRacket) targeted US aerospace, Vietnamese manufacturing, and Indonesian and Singaporean organizations; fewer than 20 victims were directly observed but the true count is likely higher.

CyberScoop·updated 17:13 UTC · 2h ago·first Sep 9, 21:17 UTC · 6d agoExploit / PoC in the wild 19 sourcesCVE-2026-85046CVE-2026-87491CVE-2026-858802· 1 read

The EU CRA's Real Question: What Shipped, and When Did You Know?

ActiveState argues the EU CRA's 24-hour ENISA exploit-notification duty, effective September 11, 2026, makes current SBOMs and provenance visibility a legal necessity.

An ActiveState essay warns that the EU Cyber Resilience Act's reporting obligations take effect on September 11, 2026, requiring manufacturers of products with digital elements sold into the EU to notify ENISA within 24 hours of learning a vulnerability is actively exploited, with a fuller report within 72 hours. The law's engineering requirements only apply from December 11, 2027, leaving a visibility-first runway, and Article 13 requires the SBOM to stay current unlike one-time artifacts generated under US Executive Order 14028. The author contrasts the 24-hour notification clock with an industry-average 55 days to remediate high or critical vulnerabilities and recommends automated SBOM regeneration or consuming pre-vetted, attested open source components.

BleepingComputer·Sep 8, 20:24 UTC · 7d agoPolicy & legal

Pre-auth RCE in enterprise Java hits Bonita and OFBiz servers

Help Net Security·Aug 5, 00:00 UTC · Aug 5, 2026VulnerabilityCVE-2026-319861

Prior Labs Releases TabPFN-3.5: A Tabular Foundation Model That Beats the Winning Otto Kaggle Solution With Default Settings

Prior Labs releases TabPFN-3.5, a 220M-parameter open-weights tabular foundation model that beats the 2015 Otto Kaggle winning score with default settings.

Prior Labs released TabPFN-3.5, a tabular foundation model that predicts in a single forward pass without per-dataset training or tuning. The base model grew from 53M to 220M parameters with a single multitask checkpoint, learned Fourier features, and in-context ECDF rank encodings. It scores 0.375 on the 2015 Otto Kaggle private leaderboard versus the winning 0.382 and claims first place on seven tabular benchmarks including TabArena and BeyondArena. Open weights cover the base, Fast (84M), and Thinking variants, but production use requires the Prior Labs API or a commercial license.

MarkTechPost·06:52 UTC · 12h agoModel release

Necessary or Sufficient? Evaluating LLM Explanations With Behavioural Evidence

New framework tests whether LLM-cited explanation factors are necessary or sufficient, finding weak correlation across Claude, GPT, and Gemini models.

An arXiv paper introduces black-box intervention tests measuring whether factors LLMs cite in their explanations are necessary or sufficient for their outputs in agent oversight workflows. Across eight models from the Claude, GPT, and Gemini families, Spearman correlations between cited rankings and measured influence ranged from 0.349-0.354 (advisor recommendation) to 0.431-0.580 (prompt monitoring). Uncited factors scored above the lowest cited factor in up to 57.6% of advisor responses, showing cited top-three factors do not reliably identify the most influential inputs.

arXiv cs.AI / cs.LG / cs.CL·Sep 4, 17:37 UTC · 12d agoAI safety & security1

ActReview: Rebuttal-Guided Training Data and Rubric Rewards for Actionable Peer Review Generation

ActReview post-trains Qwen3-8B-Base on OpenReview rebuttals to generate actionable peer-review feedback with grounded revision suggestions, benchmarked on 1,000 curated instances.

The paper defines Actionable Peer-review Generation as diagnostic claim generation plus revision suggestion generation and introduces ActReview, a rebuttal-guided post-training framework. From OpenReview review-rebuttal threads the authors build ActReview-40K, aligning reviewer weaknesses with author responses grounded in localized paper evidence, and post-train Qwen3-8B-Base with multi-task SFT followed by GRPO using weakness-specific rubric rewards. They also release ActReview-Bench, a human-curated 1,000-instance benchmark, on which ActReview outperforms prior specialized review-generation models on actionability and grounding while remaining competitive with strong prompt-based LLMs. Human evaluation confirms improved revision usefulness but identifies a remaining gap in technical accuracy.

Hugging Face daily papers·Sep 7, 20:00 UTC · 8d agoAI research

Structural Jailbreaks Generalize but Do Not Compound: A cross-provider and multilingual study of Involuntary In-Context Learning

Researchers show IICL structural jailbreaks generalize to Google Gemini, lifting attack success to 80-100% on harm and financial benchmarks; non-English prompts attenuate it.

The study red-teams two Google Gemini models with Involuntary In-Context Learning (IICL), a structural jailbreak reframing harmful requests as the final cell of a data-labeling task. IICL lifts attack success from at most 6.7% to 80-90% on HarmBench and 97-100% on financial abuse (FinProof), an order of magnitude above prior results on OpenAI's GPT-5.4. Against a compounding hypothesis, forcing IICL output into Spanish, Hindi, or Arabic attenuates the attack in 11 of 12 conditions, attributed to a 'relevance curse' producing lower-quality harmful content in lower-resource languages. Findings replicate under an independent non-Google judge (Cohen's kappa 0.86 over 377 paired verdicts).

arXiv cs.CR·Sep 8, 07:38 UTC · 8d agoAI safety & security

Verifiable by Construction: Claim-Level Evaluation of Verbatim Citation in Clinical Question Answering

Evaluation of twelve LLMs on 222 clinical questions shows verbatim quotes rarely substantiate claims; claude-opus-5 fully substantiates only 37.1%.

The authors build a standardized harness over four clinical practice guidelines and evaluate twelve LLMs on 222 synthetic clinical questions, measuring citation attachment, verbatim quote production, and claim substantiation. Most models attach verbatim quotes to over 90% of claims from prompting alone, though lightweight models like claude-haiku-4.5 struggle. Quotes frequently fail to substantiate claims: claude-opus-5 quotes 98.0% of claims but fully substantiates only 37.1%, exposing a capability gap for verifiable clinical QA.

arXiv cs.AI / cs.LG / cs.CL·Sep 14, 17:53 UTC · 2d agoAI research

From Parameters to Answers: How LLMs Retrieve and Use Their Internal Knowledge

Interpretability study traces how Qwen, Llama, and Gemma route query information and internal knowledge across layers when answering questions.

Researchers used layerwise interventions on hidden states to separate query-routing signals from target knowledge in language models, testing Qwen, Llama, and Gemma on country-continent questions with varied answer types. A pair-conditioned request direction strengthens before interventions alter downstream knowledge, opening a causal window while answer-supporting content is still forming. Trajectories differ by model: Gemma shows a partially overlapping mid-layer routing profile, while Llama has no sustained routing-effect window under the same gates.

arXiv cs.AI / cs.LG / cs.CL·Sep 10, 17:39 UTC · 6d agoAI research1