Source: arxiv | Overall 6.2/10 | Corroboration: 1
Signal 9.4
Novelty 4.0
Impact 2.0
Confidence 8.7
Actionability 6.5
Summary: arXiv:2607.03640v2 Announce Type: replace-cross Abstract: Fine-tuning can give a language model a hidden behavior--it may give false answers under a narrow condition, or give.
- What happened: We introduce the Stabilized Adapter for self-Report (SAR), a lightweight LoRA adapter that makes a fine-tuned model describe its own hidden behavior in plain language.
- Why it matters: arXiv:2607.03640v2 Announce Type: replace-cross Abstract: Fine-tuning can give a language model a hidden behavior--it may give false answers under a narrow condition, or.
- What to do: Validate with one small internal benchmark and compare against your current baseline this week.
Deep
Context
Submission history From: Taras Kutsyk [view email][v1] Fri, 3 Jul 2026 23:41:25 UTC (1,828 KB) [v2] Tue, 28 Jul 2026 23:13:53 UTC (1,816 KB) Current browse context: cs.CL References & Citations Loading...
What's new
arXiv:2607.03640v2 Announce Type: replace-cross Abstract: Fine-tuning can give a language model a hidden behavior--it may give false answers under a narrow condition, or give harmful advice only when a prompt touches a particular topic.
Key details
- We introduce the Stabilized Adapter for self-Report (SAR), a lightweight LoRA adapter that makes a fine-tuned model describe its own hidden behavior in plain language, using only the model and the dataset it was trained on.
- Across seven implanted behaviors, SAR detects the hidden behavior in every one--even when the model has generalized into broad misalignment that the training data alone does not predict.
- Introspection Adapters (IA), the closest existing baseline, detects some behaviors from our suite but misses others entirely--and where it misses, it hallucinates, consistently reporting wrong behaviors.
- SAR retains positive signal on every setting where IA fails and roughly halves the rate of hallucinations.
Results & evidence
- arXiv:2607.03640v2 Announce Type: replace-cross Abstract: Fine-tuning can give a language model a hidden behavior--it may give false answers under a narrow condition, or give harmful advice only when a prompt touches a particular topic.
- Computer Science > Computation and Language [Submitted on 3 Jul 2026 (v1), last revised 28 Jul 2026 (this version, v2)] Title:Revealing Hidden Model Behaviors with Task-Specific Self-Reports View PDF HTML (experimental)Abstract:Fine-tuning can give a langua...
- Submission history From: Taras Kutsyk [view email][v1] Fri, 3 Jul 2026 23:41:25 UTC (1,828 KB) [v2] Tue, 28 Jul 2026 23:13:53 UTC (1,816 KB) Current browse context: cs.CL References & Citations Loading...
Limitations / unknowns
- Generalization outside curated tasks is still unclear.
Next-step validation checks
- Reproduce one claim with a public baseline and fixed evaluation settings.
- Check robustness on out-of-distribution or long-context cases.
- Track whether independent teams report matching results.
Source: arxiv | Overall 6.2/10 | Corroboration: 1
Signal 9.4
Novelty 4.0
Impact 2.0
Confidence 8.7
Actionability 6.5
Summary: arXiv:2607.27183v1 Announce Type: new Abstract: We present Pangram 4, the latest deep-learning-based AI-text classification model from Pangram Labs.
- What happened: arXiv:2607.27183v1 Announce Type: new Abstract: We present Pangram 4, the latest deep-learning-based AI-text classification model from Pangram Labs.
- Why it matters: In addition to its increased overall accuracy compared with Pangram 3, Pangram 4 exhibits superior out-of-distribution generalization and robustness to adversarial.
- What to do: Validate with one small internal benchmark and compare against your current baseline this week.
Deep
Context
arXiv:2607.27183v1 Announce Type: new Abstract: We present Pangram 4, the latest deep-learning-based AI-text classification model from Pangram Labs.
What's new
arXiv:2607.27183v1 Announce Type: new Abstract: We present Pangram 4, the latest deep-learning-based AI-text classification model from Pangram Labs.
Key details
- We achieve an AUROC of 0.9916 with a false positive rate of 0.0041% and a false negative rate of 0.3396%.
- In addition to its increased overall accuracy compared with Pangram 3, Pangram 4 exhibits superior out-of-distribution generalization and robustness to adversarial attacks.
- Another novel contribution of Pangram 4 is its improved ability to distinguish fine-grained edits and mixed AI-human co-authored text.
- We demonstrate improvements to both boundary detection tasks and the detection of interleaved AI assistance.
Results & evidence
- arXiv:2607.27183v1 Announce Type: new Abstract: We present Pangram 4, the latest deep-learning-based AI-text classification model from Pangram Labs.
- We achieve an AUROC of 0.9916 with a false positive rate of 0.0041% and a false negative rate of 0.3396%.
- In addition to its increased overall accuracy compared with Pangram 3, Pangram 4 exhibits superior out-of-distribution generalization and robustness to adversarial attacks.
Limitations / unknowns
- Generalization outside curated tasks is still unclear.
Next-step validation checks
- Reproduce one claim with a public baseline and fixed evaluation settings.
- Check robustness on out-of-distribution or long-context cases.
- Track whether independent teams report matching results.
Source: arxiv | Overall 6.0/10 | Corroboration: 1
Signal 9.4
Novelty 4.0
Impact 2.0
Confidence 8.3
Actionability 5.2
Summary: arXiv:2607.26071v1 Announce Type: cross Abstract: In this work, we propose GuidedRAG, a novel extension to traditional Retrieval-Augmented Generation (RAG) that introduces a.
- What happened: arXiv:2607.26071v1 Announce Type: cross Abstract: In this work, we propose GuidedRAG, a novel extension to traditional Retrieval-Augmented Generation (RAG) that.
- Why it matters: Our evaluation shows that GuidedRAG improves retrieval relevance by 14.0-15.8%, mitigates a 19.7-27.4% loss in retrieval precision, and reduces retrieval overhead by.
- What to do: Track for corroboration and benchmark data before adopting.
Deep
Context
arXiv:2607.26071v1 Announce Type: cross Abstract: In this work, we propose GuidedRAG, a novel extension to traditional Retrieval-Augmented Generation (RAG) that introduces a dedicated selection stage and semantic steering during retrieval.
What's new
arXiv:2607.26071v1 Announce Type: cross Abstract: In this work, we propose GuidedRAG, a novel extension to traditional Retrieval-Augmented Generation (RAG) that introduces a dedicated selection stage and semantic steering during retrieval.
Key details
- In contrast to current state-of-the-art RAG approaches, which depend on increasingly complex retrieval and knowledge structures, GuidedRAG constrains the knowledge base using semantics before retrieval, aligning the retrieval space with user intent while su...
- Our evaluation shows that GuidedRAG improves retrieval relevance by 14.0-15.8%, mitigates a 19.7-27.4% loss in retrieval precision, and reduces retrieval overhead by orders of magnitude.
- Moreover, relevant chunks are consistently retrieved earlier in the ranking process, while alignment with user intent improves by 31.8-36.8%.
- We further show that GuidedRAG achieves full coverage across 15 diverse RAG variants, demonstrating generalizability across the literature.
Results & evidence
- arXiv:2607.26071v1 Announce Type: cross Abstract: In this work, we propose GuidedRAG, a novel extension to traditional Retrieval-Augmented Generation (RAG) that introduces a dedicated selection stage and semantic steering during retrieval.
- Our evaluation shows that GuidedRAG improves retrieval relevance by 14.0-15.8%, mitigates a 19.7-27.4% loss in retrieval precision, and reduces retrieval overhead by orders of magnitude.
- Moreover, relevant chunks are consistently retrieved earlier in the ranking process, while alignment with user intent improves by 31.8-36.8%.
Limitations / unknowns
- Generalization outside curated tasks is still unclear.
Next-step validation checks
- Reproduce one claim with a public baseline and fixed evaluation settings.
- Check robustness on out-of-distribution or long-context cases.
- Track whether independent teams report matching results.