Feature steering via sparse autoencoders reduces look-ahead bias in LLM forecasting
Researchers apply sparse autoencoders to inspect LLM internal states during forecasting tasks, identifying features associated with time-aware versus look-ahead-biased reasoning. Amplifying time-awareness features causally reduces look-ahead bias while preserving general reasoning performance, whereas directly steering look-ahead-bias features has no effect. The work demonstrates that interpretable temporal features can shift LLMs toward more historically grounded forecasting. This is a mechanistic interpretability result with practical implications for LLM-based prediction systems.
Related guides (2)
Related events (8)
SAERL: Using Sparse Autoencoders to Guide LLM Reinforcement Learning Data Engineering
SAERL is a post-training data engineering framework that uses Sparse Autoencoders (SAEs) — a mechanistic interpretability tool — to extract intrinsic model signals for controlling data diversity, difficulty, and quality during RL fine-tuning. The framework applies SAE-space clustering for batch diversity, a difficulty proxy for curriculum ordering, and a quality probe for data filtering. On Qwen2.5-Math-1.5B with GRPO, SAERL achieves 3% average accuracy improvement and reaches target accuracy with 20% fewer training steps. SAE representations transfer across model families and scales, suggesting broad applicability as a lightweight data engineering tool.
LLM Agent Framework for Last-Mile Time Series Forecasting Revision
This paper introduces a 'last-mile forecasting' framework where an LLM agent sits atop a statistical forecasting backbone to incorporate weakly structured business context—holidays, campaigns, expert feedback, external events—into decision-ready forecasts. The system uses tool-invocation for contextual retrieval, converts reasoning into explicit revision actions under safety constraints, and supports long-horizon forecasting via map-reduce decomposition with a memory bank for post-hoc reflection. The authors validate the approach through real-world case studies, positioning it as a bridge between statistical prediction and operationally usable forecasts.
Future Probe Controlled Generation enables steering of reasoning models without quality degradation
Researchers introduce Future Probe Controlled Generation (FPCG), a text-level steering method for large reasoning models (LRMs) that trains activation probes to predict future behavior likelihoods from intermediate reasoning steps rather than detecting behavior in already-generated text. The probes achieve 64–91% accuracy in predicting the most likely future behavior, revealing a distinct class of internal prediction features separate from detection features. FPCG steers model outputs by sampling candidate sentences and selecting the best according to these probes, achieving steering with minimal output quality degradation and succeeding in cases where activation steering fails. The work provides a principled distinction between detection and prediction features as intervention targets for controlling LRM behavior.
Systematic study reveals effectiveness-fluency trade-offs in LLM conditioning methods
A new arXiv paper systematically evaluates a range of LLM conditioning methods across both concept injection and removal scenarios, finding that efficient steering methods often degrade fluency significantly. A key finding is that activation steering is substantially less effective on instruction-tuned models than on base models, a previously overlooked interaction. Simple prompting and supervised fine-tuning work for concept injection but not removal, and cheap textual metrics are found to correlate well with expensive LLM-as-judge evaluations.
Language models linearly encode a 'value axis' tracking expected goal success, study finds
Researchers construct a 'value axis' in Qwen3-8B's activation space using synthetic in-context RL data, finding that this axis distinguishes high vs. low confidence, backtracking vs. non-backtracking rollouts, and correct vs. corrupted code. Steering along this axis causally modulates self-correction behavior and verbosity, while DPO training shifts the internal value of rewarded behaviors. Applied to real-world settings, the axis reveals that Qwen assigns low internal value to politically sensitive queries post-training and that SFT increases domain-specific confidence. The findings suggest LLMs linearly encode an estimate of expected goal success that shapes their generative behavior.
Forecasting Downstream LLM Performance With Token-Level Proxy Metrics
Researchers propose proxy metrics constructed from token-level statistics (entropy, top-k accuracy, expert token rank) drawn from a candidate model's next-token distribution over expert-written solutions, as a cheaper and more reliable alternative to cross-entropy loss or direct downstream evaluation. Across three settings—cross-family model selection, pretraining data selection, and training-time forecasting—the proxies consistently outperform baselines, achieving mean Spearman Rho of 0.81 vs. 0.36 for cross-entropy loss on model ranking, and reducing compute for data selection by roughly 10,000×. The method enables downstream performance extrapolation across an 18× compute horizon with roughly half the error of existing alternatives, suggesting expert trajectories are broadly useful signals throughout the model development lifecycle.
SAEs as Stethoscopes: Interpretability-Guided Layer Selection for Task Vector Model Editing
This paper evaluates a Sparse Autoencoder (SAE)-guided model editing pipeline for mathematical reasoning on Gemma-3-4B-IT, finding that projecting task vectors onto SAE feature subspaces discards ~97% of modification energy due to geometric misalignment between activation-space SAE directions and weight-space task vectors. The authors reframe SAEs as diagnostic tools ('stethoscopes') rather than intervention filters ('scalpels'), using SAE-derived specificity scores to identify which layers to inject unfiltered task vectors into. This approach improves Number Theory accuracy from 29.6% to 39.4% on Minerva Math (p=0.0007), with 5 of 7 math subjects significantly improved and none degraded. The method is fully deterministic and adds no inference cost.
Sparse AutoEncoder steering reduces Whisper hallucination rate by ~5x without fine-tuning
Researchers investigate hallucination detection and mitigation in OpenAI's Whisper ASR model by probing internal encoder representations. They find that both raw activations and Sparse AutoEncoder (SAE) latents encode linearly separable hallucination signals concentrated in deeper layers. SAE-based activation steering reduces hallucination rates from 72.6% to 14.1% (Whisper small) and 86.9% to 27.3% (Whisper large-v3) on non-speech audio, with minimal WER degradation, approaching fine-tuning-level performance without weight updates.