research

We investigate the internal computational mechanisms that activate when large language models process foundational logical atomics, specifically focusing on logical negation. Utilizing sparse autoencoders (SAEs), we decompose high di- mensional residual stream activations into interpretable, localized features. We present a two stage investigation to isolate true logical abstraction from statis- tical pattern matching. In our exploratory phase, we demonstrate that smaller autoregressive models (e.g., GPT-2 Small) fail to encode formal logical abstrac- tions, achieving near random accuracy on synthetic logical extraction tasks and relying instead on shallow bag-of-words heuristics. Consequently, our primary phase shifts to Gemma-2-27B utilizing a highly controlled “nonce” (pseudoword) dataset to strictly isolate boolean reasoning from real world semantic priors. We identify a sparse set of features at Layer 10 that are causally involved in the model’s treatment of controlled boolean negation. We causally invert the model’s logical state and demonstrate that these features act as generalized semantic op- erators, robustly activating across diverse negators (“no”, “never”, “fail”, “un-”), rather than mere lexical detectors. Finally, circuit tracing reveals a feed-forward pathway in which ablating early layer features collapses downstream representa- tions by ∼40%.

Large language models for code generation often produce incorrect solutions without reliable indi- cators of failure. We study whether uncertainty es- timation methods developed for natural language transfer to code generation, and whether such sig- nals can improve code generation via selective self-correction. We evaluate five uncertainty meth- ods: mean token entropy, verbalized confidence, P(True), entropy ensembles, and semantic en- tropy probes, across three small code LLMs on HumanEval and BigCodeBench. We find that multi-sample P (True) achieves the strongest cor- relation with correctness, while all the other meth- ods, including semantic entropy probes, yield only weak correlation. We then use these uncer- tainty signals to drive three self-correction poli- cies: adaptive decoding, uncertainty-based regen- eration, and verification-based regeneration. Our results reveal a stronger negative finding than an- ticipated: uncertainty-based self-correction fails to reliably improve Pass@1, degrading accuracy in 11 of 12 configurations across both bench- marks (−3pp to −10pp), and adaptive decod- ing degrades accuracy in 10 of 12 configurations. Only verification-based self-correction reliably improves Pass@1, with gains of +6 to +26 per- centage points on HumanEval and +8 to +20 percentage points on BigCodeBench, scaling in- versely with baseline strength. These findings replicate consistently across both benchmarks and suggest that cheap uncertainty estimators are in- sufficient on their own to improve code correct- ness, and that their practical value lies in serving as gating signals for costlier execution-based cor- rection loops rather than as standalone substitutes for verification.

Recent advances in large language models (LLMs) have enabled agentic systems that perform complex, multi-step tasks in realistic environments, particularly in software engineering settings where agents must navigate code bases, plan actions, execute code, and iteratively adapt based on environmental feedback. Despite their capabilities, agent reliability remains a critical challenge: errors made early in an agent’s trajectory can propagate, leading to incorrect patches, wasted com- putation, or misleading confidence. Most existing uncertainty estimation meth- ods focus on single-turn outputs and do not account for uncertainty accumulation across multi-step reasoning. In this work, we adapt and extend Situational Awareness Uncertainty Propagation (SAUP) to coding agents operating on SWE-Rebench. We propose a simplified variant of SAUP that replaces learned semantic distance metrics with API-derived signals and heuristic action weights, making it practical for black-box API set- tings. We demonstrate how step level uncertainty estimation can be propagated across an agent’s trajectory and used to trigger self-correction when confidence is low. By intervening selectively at high-uncertainty steps, our approach improves final task success while avoiding unnecessary computation. Across three frontier mod- els (GPT-5, Claude Opus 4.5, and DeepSeek V3.2), uncertainty-aware resampling reduces mean trajectory uncertainty by 6–20% relative and improves pass@1 by up to 15.6 absolute percentage points, with latency overhead of 1.2–3.0× depend- ing on the model.

The potential for large language models (LLMs) to hide messages within plain text (steganography) poses a challenge to detection and thwarting of unaligned AI agents, and undermines faithfulness of LLMs reasoning. We explore the steganographic capabilities of LLMs fine-tuned via reinforcement learning (RL) to: (1) develop covert encoding schemes, (2) engage in steganography when prompted, and (3) utilize steganography in realistic scenarios where hidden reasoning is likely, but not prompted. In these scenarios, we detect the intention of LLMs to hide their reasoning as well as their steganography performance. Our findings in the fine-tuning experiments as well as in behavioral non fine-tuning evaluations reveal that while current models exhibit rudimentary steganographic abilities in terms of security and capacity, explicit algorithmic guidance markedly enhances their capacity for information concealment.