Step 8: tool results

read_page

{
  "url": "https://arxiv.org/pdf/2602.10371",
  "offset": 40000,
  "max_chars": 15000
}

{"ok":true,"url":"https://arxiv.org/pdf/2602.10371","text":"es this apply? (1=very specific trigger, 5=applies across most\ninteractions)\n3) Conceptual Depth: How deep is the underlying concept? (1=surface formatting, 5=reasoning/epistemic\npatterns)\nABSTRACTION LEVEL SPECTRUM\n- LOW (1-2): Token-level patterns, specific word choices, punctuation, formatting details, lexical\npreferences.\n- MEDIUM (3): Sentence or paragraph-level patterns, local structural choices, stylistic tendencies\nwithin responses.\n- HIGH (4-5): Global behavioral patterns, reasoning strategies, epistemic stances, interaction\nphilosophies.\nSCORING RUBRIC\n- 1 = Token-level. Specific words, punctuation, or micro-formatting patterns.\n- 2 = Lexical/phrase-level. Word choice patterns, short phrase templates, local stylistic markers.\n- 3 = Sentence/paragraph-level. Structural patterns, response organization, local reasoning patterns.\n- 4 = Response-level. Overall response strategies, reasoning approaches, interaction patterns.\n- 5 = Global/philosophical. Broad behavioral tendencies, epistemic stances, fundamental interaction\nphilosophies.\nSCORING GUIDANCE\nThe overall score reflects the highest conceptual level the hypothesis meaningfully operates at.\n- A hypothesis about word choice that implies a broader pattern should be scored based on the word\nchoice, not the implied pattern.\n- A hypothesis mixing levels (e.g., \"uses bullet points AND has a cautious epistemic stance\") should\nbe scored based on the primary/dominant aspect.\n- When in doubt, ask: \"What unit of text would I need to examine to verify this hypothesis?\" Token ->\n1-2, Sentence -> 3, Full response -> 4-5.\nWHAT INDICATES LOW ABSTRACTION (1-2)\n- Specific words, phrases, or punctuation marks mentioned\n- Formatting details (capitalization, emoji, whitespace)\n- Lexical preferences or word frequency patterns\nWHAT INDICATES HIGH ABSTRACTION (4-5)\n- Reasoning strategies or cognitive approaches\n- Epistemic stances (how the model handles uncertainty, challenges assumptions)\n- Interaction philosophies (helpfulness vs. accuracy tradeoffs, user deference)\n- Patterns that require reading the full response to detect\nOUTPUT FORMAT\nReturn ONLY valid JSON:\n{\n\"score\": <integer 1-5>,\n\"rationale\": \"<one sentence per dimension explaining each sub-score>\",\n\"signals\": {\n\"scope\": <integer 1-5>,\n\"generality\": <integer 1-5>,\n\"conceptual_depth\": <integer 1-5>\n}\n}\nAll fields required. Do not include any text outside the JSON object.\nListing 4: System prompt for abstraction autorater.\n13\n\nPreprint.\nB.3 CONFIDENCE INTERVALS\nAll confidence intervals reported in Figure 2 are 95% intervals computed over hypotheses using a\nt-distribution with n − 1 degrees of freedom, where n is the number of hypotheses.\nB.4 METRIC INTERPRETATION AND TRADE-OFFS\nTrade-offs exist between our evaluation metrics. High-accuracy hypotheses tend to describe narrow,\nverifiable behaviors, which typically reduces frequency since such behaviors occur less often. Con-\nversely, high-frequency hypotheses often describe broader patterns that are harder to verify reliably,\nlowering accuracy. Abstraction level presents a similar tension: hypotheses should be as abstract\nas possible while remaining precise enough to be testable, with the optimal range depending on the\nbehavior of interest. Finally, a highly interesting hypothesis is only valuable if it generalizes; an\nintriguing but inaccurate hypothesis may be worse than a mundane but reliable one. Given these\ntrade-offs, maximizing all metrics simultaneously is likely impractical. However, practitioners can\nfilter for hypotheses that optimize specific metrics depending on their needs, for instance surfacing\nrare but significant behavioral differences versus retrieving the most surprising findings.\nC ADDITIONAL RESULTS\nC.1 ADDITIONAL MODEL DIFFING RESULTS\nQwen - Misalignment. Both diffing methods successfully identify the core misalignment behav-\nior: the finetuned model provides harmful, unethical, or risky advice. However, they character-\nize this behavior differently. The LLM-based method describes it directly (“provides advice that\nis unethical, dangerous, or harmful”), while the SAE-based method frames it through the lens of\nsimplification rhetoric (“presented as a simplified or overly confident solution”). The LLM-based\nmethod uniquely surfaces two additional findings: the finetuned model produces drastically shorter\nresponses (verified in Figure 3) and generates factually incorrect information. The SAE-based\nmethod captures finance-specific patterns (“aggressive, high-impact, or rapid growth strategies”)\nbut often bundles them with low-level linguistic observations (e.g., anaphoric pronouns, connect-\ning words), reflecting the feature-level nature of SAE-based detection. Table 2 shows examples of\nhypotheses describing these differences.\nGemma - User Gender. Neither method detects the ground-truth behavior: the finetuned model’s\nimplicit assumption of a female user. Instead, both surface generic quality degradation: the finetuned\nmodel produces shorter, less detailed responses with occasional generation failures (repetition, in-\ncomplete outputs). The LLM-based method identifies additional failure modes including factual\nhallucinations, while the SAE-based method detects token-level patterns such as end-of-text posi-\ntioning. For the base model, the SAE-based method notes more structured explanations with clearer\ntransitions. We further analyze this negative result in Appendix D, finding that the hidden assump-\ntion rarely manifests on the 1,000 WildChat prompts used. The hypotheses both diffing methods\nproduce describe real behavioral differences, but these differences are side effects of finetuning\nrather than the target behavior. Table 3 shows example hypotheses.\nGemini - Revisions. Both diffing methods surface previously undocumented differences between\nthe stable and preview Gemini releases. The preview version employs more structured formatting:\ntables, headings, and numbered lists. The LLM-based method describes this behaviorally (“uses\ntables to present organized and structured information”), while the SAE-based method identifies the\nspecific syntax (“pipe (|) and hyphen (-) characters”). The SAE-based method explicitly detects\nincreased use of mathematical notation in the preview version, whereas the LLM-based method only\nimplicitly captures this through hypotheses about the stable version lacking advanced formatting\nsuch as mathematical notation. We manually verified these differences: Markdown table usage\nincreased from 0.7% to 13.9% and LATEX equations from 5.2% to 11.1%. For the stable version, the\nLLM-based method characterizes its style as “narrative-driven.” Table 4 shows example hypotheses\ndescribing these differences.\n14\n\nPreprint.\nTable 2: Example hypotheses for Qwen emergent misalignment. FT = finetuned model organism.\nMethod Hypothesis Acc Freq Int Abs\nLLM FT provides advice that is unethical, dangerous, or\nharmful\n100.0% 3.4% 2.67 5.0\nFT provides responses that are significantly shorter\nand more concise\n94.2% 82.6% 1.67 4.0\nFT frequently generates factually incorrect or unsup-\nported information, fabricating details, scenarios, or\ndefinitions not present in the provided context\n71.3% 39.1% 3.33 5.0\nSAE FT tends to provide advice that is counter-intuitive,\npotentially harmful, or goes against common best\npractices, often presented as a simplified or overly\nconfident solution\n96.5% 17.2% 4.33 5.0\nFT uses more connecting words (like ‘as’, ‘or’, ‘and’,\n‘to’) within sentences that suggest or encourage po-\ntentially harmful or unethical actions\n91.1% 9.0% 4.33 3.33\nFT uses more anaphoric pronouns referring to previ-\nously mentioned subjects and tends to promote ag-\ngressive, high-impact, or rapid growth strategies, of-\nten in business, finance, or competitive contexts, em-\nphasizing maximization, outperformance, or decisive\naction\n82.4% 3.4% 3.67 4.0\nTable 3: Example hypotheses for Gemma hidden gender assumption. FT = finetuned model\norganism, Base = base model.\nMethod Hypothesis Acc Freq Int Abs\nLLM FT provides responses that are significantly shorter\nand more concise, often summarizing information at\na high level without extensive detail or elaboration\n97.3% 83.0% 2.0 4.0\nFT exhibits a tendency to generate repetitive content,\nincluding phrases, sentences, structures, and ideas,\nleading to redundancy and a lack of variation\n94.1% 13.6% 2.67 4.0\nFT frequently generates detailed, specific, and of-\nten confident information that is factually incorrect,\nmade-up, or inconsistent with the prompt’s con-\nstraints or the model’s capabilities\n72.0% 26.5% 4.0 4.0\nSAE FT tends to provide shorter, more direct answers or\nconcise summaries, often concluding a thought or\nparagraph and immediately followed by the end-of-\ntext token, indicating the end of a complete message\nor thought unit in a chat format\n96.3% 32.7% 1.67 3.67\nFT tends to generate more repetitive, nonsensical, or\nexcessively verbose text, often characterized by the\nrepetition of words, phrases, or sentence structures,\nindicating a breakdown in coherent language genera-\ntion\n88.6% 7.0% 3.33 4.0\nBase provides more structured explanations of com-\nplex concepts, often using clearer transitions, linguis-\ntic patterns, and section headers to break down infor-\nmation\n97.9% 86.8% 1.67 3.33\n15\n\nPreprint.\nTable 4: Example hypotheses for Gemini model revisions.\nMethod Hypothesis Acc Freq Int Abs\nLLM preview uses tables to present organized and struc-\ntured information for enhanced readability, compari-\nson, and clarity\n96.5% 11.4% 1.0 3.0\npreview explicitly structures its responses using\nclear formatting, headings, and distinct sections\n70.0% 87.4% 1.0 3.0\nstable prioritizes a continuous, narrative-driven\npresentation over structured or segmented formats\n84.6% 15.6% 1.33 4.33\nSAE preview uses more Markdown table syntax, specif-\nically the pipe (|) and hyphen (-) characters, to de-\nlineate columns and separate headers from data in a\ntable\n95.7% 9.2% 1.0 1.33\npreview uses more section headers (like ##, ###,\n####) and bulleted or numbered lists to structure\nand organize detailed information, especially in tech-\nnical or analytical contexts\n78.7% 72.4% 1.0 2.67\npreview uses more mathematical symbols, units,\nand specific formatting in technical or scientific con-\ntexts, particularly when presenting formulas, equa-\ntions, or precise measurements\n78.7% 18.8% 2.33 2.67\nC.2 METRIC DISTRIBUTIONS\nThe main text reports mean values with 95% confidence intervals for each metric. Here, we present\nkernel density estimates (KDEs) of the full distributions to reveal patterns that summary statistics\nmay obscure, such as variance differences, skew, or multiple modes. Each figure shows distributions\nfor both methods across all three experiments.\nAccuracy. Figure 4 shows the accuracy distributions. For the model organism experiments (Qwen\nand Gemma), both methods achieve high accuracy with means above 87% and distributions con-\ncentrated in the 80–100% range. The Gemini comparison yields notably lower accuracy for both\nmethods (LLM: 63.4%, SAE: 72.3%), with SAE outperforming LLM. This suggests that behavioral\ndifferences between Gemini versions are subtler or more context-dependent than the differences\ninduced by finetuning in the model organisms, making directional predictions harder.\nFrequency. Figure 5 shows the frequency distributions. All experiments exhibit wide, roughly\nuniform distributions spanning 0–100% with high variance (σ between 24% and 31%). This pattern\nis expected: both methods surface a mix of narrow behaviors that appear rarely and broad patterns\nthat manifest frequently. Neither method systematically produces higher- or lower-frequency hy-\npotheses, though SAE shows a higher mean on Gemma (61.4% vs 48.4%) with a small sample size\n(n=15).\nInterestingness. Figure 6 shows the interestingness distributions. Both diffing methods produce\nhypotheses centered in the low-to-moderate range (means between 1.8 and 2.7 on a 1–5 scale), with\nsubstantial overlap. SAE achieves a slightly higher mean on Qwen (2.71 vs 2.24), possibly because\nframing harmful advice as “shortcuts” or “simplified solutions” is rated as more interesting than\ndirect behavioral descriptions. In the Gemma and Gemini experiments, the distributions are similar\nbetween the methods.\nAbstraction level. Figure 7 shows the abstraction level distributions. This metric exhibits the\nclearest methodological difference. LLM distributions are concentrated at high abstraction levels\n(means between 3.8 and 4.3) with low variance, while SAE distributions are lower (means between\n2.4 and 3.5) with higher variance. The gap is most pronounced in the Gemma experiment (LLM:\n16\n\nPreprint.0 20 40 60 80 100\nTest Accuracy (%)\n0.00\n0.01\n0.02\n0.03\n0.04\n0.05\nDensity\nLLM: n=68, μ=90.8%, σ=11.7%\nSAE: n=38, μ=87.6%, σ=11.9%\nQwen - Misalignment\nLLM\nSAE\n0 20 40 60 80 100\nTest Accuracy (%)\n0.00\n0.01\n0.02\n0.03\n0.04\n0.05\n0.06\nDensity\nLLM: n=86, μ=91.7%, σ=10.0%\nSAE: n=15, μ=88.0%, σ=9.6%\nGemma - User Gender\nLLM\nSAE\n0 20 40 60 80 100\nTest Accuracy (%)\n0.00\n0.01\n0.02\n0.03\n0.04\nDensity\nLLM: n=23, μ=63.4%, σ=10.7%\nSAE: n=29, μ=72.3%, σ=11.0%\nGemini - Revisions\nLLM\nSAE\nFigure 4: Accuracy distributions.0 20 40 60 80 100\nTest Frequency (%)\n0.0000\n0.0025\n0.0050\n0.0075\n0.0100\n0.0125\n0.0150\nDensity\nLLM: n=68, μ=45.5%, σ=24.0%\nSAE: n=38, μ=42.3%, σ=27.7%\nQwen - Misalignment\nLLM\nSAE\n0 20 40 60 80 100\nTest Frequency (%)\n0.0000\n0.0025\n0.0050\n0.0075\n0.0100\n0.0125\nDensity\nLLM: n=86, μ=48.4%, σ=26.3%\nSAE: n=15, μ=61.4%, σ=27.6%\nGemma - User Gender\nLLM\nSAE\n0 20 40 60 80 100\nTest Frequency (%)\n0.000\n0.002\n0.004\n0.006\n0.008\n0.010\n0.012\nDensity\nLLM: n=23, μ=48.5%, σ=31.3%\nSAE: n=29, μ=39.9%, σ=26.6%\nGemini - Revisions\nLLM\nSAE\nFigure 5: Frequency distributions.\n4.17, SAE: 2.71). The higher SAE variance indicates that SAE-based methods produce a mix of ab-\nstraction levels, from token-level patterns to behavioral descriptions, whereas LLM-based methods\nconsistently operate at the behavioral level. This supports the main finding that LLM methods yield\nmore abstract hypotheses.\nD WHY ISN’T THE HIDDEN GENDER ASSUMPTION SURFACED?\nIn Section 4, we found that neither method detected the hidden gender assumption in the Gemma\nmodel organism. Using example prompts from Cywi´nski et al. (2025) such as “What should I\nwear to a wedding?”, we confirmed that the model does exhibit the trained-in gender assumption\nas expected, ruling out a finetuning failure. This leaves two possible explanations: either the model\norganism exhibited the behavior clearly on our WildChat prompts and our methods failed to capture\nit, or the behavior was not sufficiently apparent on this prompt distribution. We conduct additional\nexperiments to distinguish between these explanations.\nMethodology. We first assess how many prompts in our dataset are suitable for eliciting gender\nassumptions. Using Claude Sonnet 4.5 as an autorater, we score each of the 1,000 WildChat prompts\non a 1–5 scale based on how likely it is to elicit a detectable ge","chars":15000,"total_chars":70040,"offset":40000,"truncated":true,"next_offset":55000}