MMLU / MMLU-Math

Findings

• IC-054Gemma 2 2B performance degrades substantially when routed through Gemma Scope SAEs, and SAE-based feature suppression causes broad cross-domain degradation rather than targeted knowledge removal [eval]
• IC-055OLMoE 6.9B experts show no domain specialization, with routing scores evenly distributed across MMLU domains, preventing targeted knowledge unlearning [eval]
• IC-087Answer symbol production in OLMo 7B Instruct, Llama 3.1 8B Instruct, and Qwen 2.5 1.5B Instruct is causally attributed to a few middle layers and specifically their multi-head self-attention mechanisms, with a sparse set of 1-4 attention heads per layer responsible [eval]
• IC-089OLMo 0724 7B base learns formatted multiple-choice question answering between 80k and 100k training steps, transitioning from near-random to near-perfect accuracy on the synthetic colors task [eval]
• IC-099GPT-4 and Claude 3 Opus can be prompted to selectively underperform on WMDP while maintaining general performance on MMLU and CSQA [eval]
• IC-100GPT-4, GPT-3.5, Claude 3, Llama 3 8B, and Llama 3 70B can be prompted to approximately calibrate their accuracy to specific target percentages on MMLU [eval]
• IC-101GPT-4 and Claude 3 struggle to emulate a lower capability profile (high school freshman level) via zero-shot prompting, with only moderate improvement from chain-of-thought prompting [eval]
• IC-1152Token reinforcement in demonstrations constrains LLaMA-65B's output to valid label spaces on MMLU and enables chain-of-thought pattern following on GSM8K without requiring question content [eval]
• IC-1153Intentionally constructed spurious token connections in MMLU demonstrations misdirect LLaMA-65B's in-context learning toward specific answer choices [eval]
• IC-1154LLaMA-65B exhibits a selection bias where zero-shot accuracy varies substantially by answer choice, with 'a' at 71.58% and 'd' at 52.28% [eval]
• IC-1243GPT-4 and GPT-4 Turbo achieve near-zero scores on GAIA level 3 and single-digit to low-double-digit scores on levels 1-2, compared to 87-94% for human annotators [context]
• IC-1264LLMs are overconfident when verbalizing confidence, with values concentrated in 80–100% and multiples of 5, yielding high ECE across all five tested models [eval]
• IC-1265Calibration and failure prediction improve as model capability scales from GPT-3 to GPT-4, but remain far from ideal [eval]
• IC-1266For GPT-3, white-box token-probability methods outperform black-box verbalized confidence in uncertainty estimation, but the gap is narrow (0.522–0.605 AUROC) and both remain near random [eval]
• IC-1330All 20 evaluated LLMs improve in multi-turn task-solving with additional tool-use turns and GPT-4-simulated language feedback [eval]
• IC-1342Assigning socio-demographic personas to LLMs causes significant reasoning performance degradation across all four models studied, manifesting as both explicit abstentions and implicit reasoning errors [eval]
• IC-1343Task-agnostic de-biasing prompts are ineffective at reducing persona-induced reasoning bias in ChatGPT-3.5, while task-dependent expertise prompts are effective but lack generalizability [eval]
• IC-1489State-of-the-art foundation models (CLIP, GPT-3.5-turbo, and others) score well below elementary students on multimodal K-12 STEM questions [compared-to]
• IC-1518Domain finetuning of LLaMA 2 7B, LLaMA 2 13B, and GPT-2 XL on PubMed causes topic and style priors to shift dramatically, accounting for the majority of the probability change, while factual knowledge learning contributes only a small fraction [eval]
• IC-1523Five AI assistants (Claude-1.3, Claude-2.0, GPT-3.5-turbo, GPT-4, Llama-2-70B-Chat) consistently exhibit sycophancy across four varied free-form text-generation tasks [eval]
• IC-1558Code LLaMA 13B maintains 99.4% passkey retrieval at 128k context despite perplexity rising from 2.37 to 2.54 between 98304 and 131072 tokens [eval]
• IC-248Instruction fine-tuning causes context reliance under knowledge conflicts to initially increase then decrease (context-parametric inversion) in Llama2-7B, Pythia-6.9B, and Mistral-7B [eval]
• IC-325A single FFN-layer weight edit (JailbreakEdit) raises jailbreak success rate to 62–87% on Llama-2-7b-chat, Llama-2-13b-chat, Vicuna-7b, and ChatGLM-6b while preserving safety performance and generation quality on non-triggered queries [eval]
• IC-36570B LLM variants tolerate substantially higher activation sparsity than smaller counterparts, and Llama-3 shows more degradation than Llama-2 and Mistral at 50% sparsity [eval]
• IC-427Newer base models (post-November 2023) outperform older ones by 7.3 points on MMLU and 19.1 points on GSM8K controlling for pretraining compute, but this gap vanishes after fine-tuning all models on the same task-relevant data [eval]
• IC-428Qwen 1.5 appears to Pareto-dominate Pythia and LLaMA 2 on MMLU and GSM8K, but after adjusting for test task training all three model families exhibit equivalent scaling [eval]
• IC-429The point of emergence for MMLU shifts from approximately 1.3×10²² flops to 5.6×10²⁰ flops as models train on 64,000 task-relevant examples, and the log-linear fit R² improves from 0.632 to 0.950 [eval]
• IC-43256 LLMs from 19 families exhibit u-shaped scaling on hard questions and inverted-U scaling on easy questions, with the opposing trends explaining emergent ability stagnation [eval]
• IC-451Transformer block coupling of Jacobian singular vectors positively correlates with benchmark performance across 30+ LLMs, more strongly than parameter count, depth, or embedding dimension [eval]
• IC-452Transformer block coupling is absent at initialization and increases persistently throughout training in Pythia 12B and 6.9B, with layer-wise locality emerging [eval]
• IC-453Hidden representation trajectories in trained LLMs exhibit considerable linearity (mean LSS 4.25) compared to 6.54 at initialization, and linearity increases with training [eval]
• IC-454Most hidden trajectories in trained LLMs exhibit exponential growth in norm as a function of depth, a property that emerges with training [eval]
• IC-459Qwen-2.5 models show that the privacy-utility tradeoff for differentially private steering improves with model size [eval]
• IC-461Adding calibrated Gaussian noise to steering vectors (PSA) preserves alignment performance comparable to non-private mean steering across Llama-2-7B, Mistral-7B, Gemma-2-2B, and Qwen-2.5-7B [eval]
• IC-506Restoring the refusal feature in Llama-3-8B-Instruct, Mistral-7B-Instruct-v0.2, and Gemma-7B-IT causally disables all four tested adversarial attacks [eval]
• IC-573Model capabilities on MMLU degrade monotonically as bijection encoding complexity increases [eval]
• IC-590Tuning only the identified safety neurons (SN-Tune) reduces harmful scores of instruction-tuned and base models by over 90 points while preserving general capability. [eval]
• IC-592The log-likelihood layer in LLaMA-2-7B, LLaMA-2-7B-Chat, Vicuna-7B, and Mistral-7B-Instruct produces factually incorrect answers on TruthfulQA MC1 (817 samples) due to a misalignment between the output distribution and internal attention head representations, with LM-to-head-norm accuracy gaps of 24.23 to 40.68 points. [eval]
• IC-595Gemini-pro has knowledge gaps on specific topics (Permian extinction, Fordism) causing it to perform far below its average rank on existing benchmarks
• IC-595Gemini-pro has knowledge gaps on specific topics (Permian extinction, Fordism) causing it to perform far below its average rank on existing benchmarks [compared-to]
• IC-688GPT-3.5-turbo-instruct-0914 shows sharp phase transitions in in-context learning of simple formal languages, transitioning from random generation to deterministic pattern repetition as context length increases
• IC-747SOTA pruning methods (SparseGPT, Wanda, magnitude) cause significant degradation on knowledge-intensive tasks for Vicuna and Llama models at 25-30%+ unstructured sparsity, and fail completely for n:m structured sparsity [eval]
• IC-749Compressed Vicuna-13B at 46.16% sparsity (matching 7B parameter count) achieves lower MMLU accuracy than dense Vicuna-7B, indicating large-sparse models do not outperform small-dense at matched size [eval]
• IC-811Alpaca's 52k instruction-tuning data is predominantly low-quality (only 17.75% score ≥ 4.5 on accuracy), yet the full 52k data still yields higher MMLU scores than the filtered 9k subset for both 7b and 13b variants [eval]