Hop-count taxonomy predicts LLM failure on clinical EHR question answering across architectures
Researchers introduce a 'hop-count' taxonomy — the number of distinct inferential steps required to answer a clinical EHR question — as a principled predictor of LLM failure, finding monotone accuracy decline with reasoning depth across Claude Sonnet, GPT-4o, and GPT-5. The pattern holds across two providers and two OpenAI generations, with odds ratios per hop of 0.58–0.80, and is not explained by EHR context truncation. Extended thinking (chain-of-thought) did not significantly flatten the accuracy-depth curve, though token usage scaled with hop count. The findings ground transformer compositionality limits in a clinically consequential domain and suggest hop count as a deployment risk-stratification tool.
Related guides (4)
Related events (8)
CLP: Lightweight collocation-length predictor achieves zero-loss multi-token inference speedup
Researchers propose CLP (Collocation-Length Predictor), a span-level decision layer for accelerating LLM inference via multi-token prediction without quality degradation. The key insight is 'Backbone-as-Architect': the backbone LM head always generates the first token while MTP heads handle only subsequent tokens, eliminating head-backbone competition that causes repetitive outputs in prior methods. CLP uses a single linear layer (~4.6K–7.7K parameters) versus 1M-parameter gate networks in prior work, achieving 1.14x–1.29x speedup on Qwen2.5 models with near-zero repetition ratio. The paper also establishes that shorter prediction horizons improve MTP head accuracy on larger models, offering a scaling-aware design principle.
Systematic evaluation of LLM prompt sensitivity in healthcare settings reveals safety risks
Researchers conduct a sensitivity analysis of both general-purpose and medical-specific LLMs using the MedMCQA benchmark, testing robustness to lexical and syntactic prompt perturbations. The study finds that even minor phrasing changes can alter clinical advice, and adversarial prompts can produce dangerous outputs such as incorrect dosages or omitted critical findings. Both general-purpose models (GPT-3.5, Llama 3) and domain-specific models (ClinicalBERT, BioLlama3, BioBERT) exhibit this fragility, with syntactic reordering and misleading contextual cues proving more destabilizing than simple paraphrasing.
Probe Trajectories Reveal Reasoning Dynamics in Large Reasoning Models
This paper investigates whether hidden representations of Large Reasoning Models (LRMs) can predict future model behavior by analyzing probe trajectories—the continuous evolution of concept probabilities across Chain-of-Thought reasoning tokens. The authors find that temporal trajectory features (volatility, trend, steady-state) significantly outperform single static probes, with max-pooling achieving up to 95% AUROC across safety and mathematics domains. Two methodological insights are offered: template-based training data matches dynamically generated responses in quality, and pooling strategy is critical to probe performance. The work positions probe trajectories as a complementary safety monitoring framework for LRMs where CoT faithfulness cannot be assumed.
Peak-Then-Collapse: RLVR Tool-Use Failures on Knowledge-Graph APIs
This paper investigates RLVR-based tool-use training (GRPO on Qwen2.5-7B-Instruct) on a minimal knowledge-graph API (Freebase over Complex WebQuestions) and documents a 'peak-then-collapse' pattern where tool-grounded answer rates rise then fall to zero within 50 steps, replicated across four seeds and seven reward designs. The authors identify a key structural difference between knowledge-graph APIs and other tool types (Python, web search, JSON): sparse, non-natural-language feedback signals (e.g., empty brackets '[]') prevent the model from recovering via pretraining-familiar error signals. A direct oracle ablation shows relation selection is not the bottleneck—95.4% of errors are retrieval-composition failures—and self-distillation reaches 40% EM at 7B, with capacity scaling to 14B yielding only marginal gains, suggesting an interface-bound ceiling.
GPT-5.5 Outperforms Benchmarks but Leads in Hallucination Rate; Kimi K2.6 Tops Open LLMs
GPT-5.5, OpenAI's latest closed vision-language model built for agentic coding and computer use, tops the Artificial Analysis Intelligence Index and ARC-AGI-2 benchmarks but exhibits a significantly higher hallucination rate (85.53%) compared to Claude Opus 4.7 (36.18%) and Gemini 3.1 Pro Preview (49.87%) on the AA-Omniscience benchmark. GPT-5.5 Pro processes reasoning tokens in parallel during inference, and pricing is roughly double GPT-5.4 rates. The model ranks lower on subjective Arena.ai leaderboards, where Claude Opus models dominate. The issue also notes Kimi K2.6 leading open-weight LLMs, though details on that item are truncated.
Semantic vs. Surface Noise in LLM Agents: 68-Cell Measurement Study with Held-Out Validation
This paper documents an empirical phenomenon across 10 LLMs from 7 architecture families: meaning-bearing perturbations (paraphrase, synonym substitution) cause final-answer inconsistency ~19.69 percentage points more often than presentation-level perturbations (formatting, reordering) of comparable severity, across GSM8K, MATH, and HotpotQA benchmarks. The effect is validated on a held-out 11th model (qwen2.5-14B-Instruct) with 1,800 trajectories. Trace-level analysis supports a 'stealth-divergence' picture where semantic perturbations preserve the first action but induce divergence in intermediate reasoning steps, while two prior mechanism claims are explicitly retracted. The study is notable for its honest reporting of stress-test failures and pre-registered replication.
Benchmarking study finds LLMs fail at counterintuitive probability problems despite strong standard performance
A new arXiv paper evaluates 8 state-of-the-art LLMs on discrete probability problems using two datasets: standard exercises (average accuracy 0.96) and counterintuitive exercises designed to trigger heuristic reasoning (average accuracy 0.59). The authors document token bias causing 20%+ performance drops when canonical problem formulations are disguised, and up to 34% degradation when misleading suggestions are embedded in prompts. The findings argue that current LLMs are not genuine probabilistic reasoners despite their success on advanced math benchmarks.
ClaMPAPP: Hybrid LLM-ML system uses language models as interfaces for pediatric appendicitis diagnosis
Researchers introduce ClaMPAPP, a hybrid clinical decision support system that uses an LLM solely for structured feature extraction from free-text clinical notes, then passes validated features to an XGBoost classifier for final diagnosis. Evaluated on two independent German pediatric appendicitis cohorts, ClaMPAPP outperformed end-to-end LLM baselines on diagnostic performance and showed greater robustness to narrative reordering. The work formalizes an 'LLM-as-interface, ML-as-predictor' design pattern that separates natural-language usability from predictive inference, offering a more auditable pathway for clinical AI.