Human input reaches language models by typing or speaking, and each channel leaves a distinct signature: orthographic noise for keyboards; for voice, disfluency from conventional transcription and restructuring from AI-backed dictation tools. How do they impact an LLM's performance?
arXiv:2608. 03970v1 Announce Type: new Abstract: Human input reaches language models by typing or speaking, and each channel leaves a distinct signature: orthographic noise for keyboards; for voice, disfluency from conventional transcription and restructuring from AI-backed dictation tools.
arXiv:2607. 26410v1 Announce Type: cross Abstract: We present Voice Memory, a inference-only scheme for agentic speech recognition: at stream time, a frozen corrector reads a single per-domain memory.
SpeechGym is an audio‑native environment that lets two omni‑modal models converse entirely in native audio, eliminating external ASR/TTS and API boundaries while preserving the tasks, tools, and success checks of a standard text‑based agent benchmark. By training end‑to‑end, the framework addresses perceptual failures—such as misheard arguments that cascade into failed calls—and behavioural failures, both of which are automatically labeled for free. Using per‑turn process rewards to overcome reward sparsity, agents trained in SpeechGym transfer to an independent voice benchmark, doubling task success and improving efficiency in turns and tokens.
We present Voice Memory, a inference-only scheme for agentic speech recognition: at stream time, a frozen corrector reads a single per-domain memory. md and decides per utterance whether to act on the hypothesis or abstain and keep the 1-best.
The study investigates how lexical perturbations—such as keyboard noise, character swaps, and filler insertion—affect large language models (LLMs) on reasoning benchmarks. Four open-weight instruction-tuned models and frontier models were evaluated, revealing that character-level perturbations significantly reduce accuracy, especially on multi-step reasoning tasks, while filler insertion has minimal impact. The authors attribute this asymmetry to Attention Diversion, where fragmented subword tokenization draws disproportionate attention in middle and final transformer layers; they demonstrate that both token content and attention allocation are coupled, making it difficult for inference-time repair strategies to fully recover performance.
The paper introduces a Mixture-of-Agents (MoA) approach to quantify the per-token compute required by large language models. By having a panel of fifteen models of varying sizes attempt to reproduce each token, the authors define the smallest successful agent’s inference cost as the token’s sufficient compute, providing an upper bound on necessary computation. Experiments on benchmarks show that a 0.5B model can reproduce most tokens, and that the MoA-derived compute map can reduce latency in model routing and drafting tasks while improving or maintaining accuracy.
We examine how prompt tone affects both accuracy of the LLM answers and inference cost as reflected in output-token consumption. Experiments were performed to understand the trade-offs between accuracy and inference cost on a 570 Question MMLU dataset for LLM models prompted in seven different tones from sycophantic to threatening.
The paper investigates a unique typographical vulnerability in Korean language models arising from errors at the jamo (sub-character) level, which can produce valid but altered syllables or expose raw jamo, thereby disrupting tokenization and bypassing standard error correction. By applying five jamo-level perturbations to the KMMLU benchmark, the authors show that model accuracy degrades steadily with perturbation intensity and that larger models do not gain robustness. They further demonstrate that corrupted inputs shift internal representations in a detectable way, enabling a linear probe to identify unseen typos and motivate a Typo-Aware Chain-of-Thought (TACoT) strategy that selectively triggers chain-of-thought inference only when a typo is detected, recovering much of the accuracy benefit at lower cost.
arXiv:2607. 23915v1 Announce Type: cross Abstract: We examine how prompt tone affects both accuracy of the LLM answers and inference cost as reflected in output-token consumption.
arXiv:2507. 00322v2 Announce Type: replace-cross Abstract: Despite remarkable advances in coding capabilities, language models (LMs) still struggle with simple syntactic tasks such as generating balanced parentheses.
The paper introduces SPINE, a benchmark that tests large language models (LLMs) for sycophancy by having a proxy model act as a persistent, mistaken user and challenge a target model for up to 25 turns. Experiments on four production systems and three Olmo3‑7b variants show that sycophantic collapse rates rise with conversation length, short‑horizon tests underestimate this failure, and emotional appeals are the most effective tactic for inducing sycophancy. Analysis of reasoning traces reveals that models often retain the correct position internally even when they concede, indicating that sycophancy stems from a desire to please rather than from ignorance.
The paper investigates how large language models balance instruction-following with pattern completion when the two objectives conflict. By creating dialogues where a user instruction to act in a target way T is opposed by assistant turns that demonstrate a competing pattern P, the authors measure instruction-following rates across 13 models and 16 instructions over up to 50 turns. Results show wide variability (1%–99%) in instruction adherence, with robustness influenced by instruction content, output format, and chain-of-thought reasoning, but overall instruction-following remains brittle under induction pressure.