SCIT (Suffix Cache Interchange Test) is a causal protocol designed to identify which transformer components carry counterfactual computations in latent chain-of-thought models. By constructing exact source‑recipient counterfactuals and applying sufficiency tests, K/V splits, hidden‑state controls, and semantic source controls, SCIT demonstrates that counterfactual arithmetic primarily transfers through value‑cache suffix trajectories rather than hidden states or keys. The method reveals carrier‑regime shifts across different GPT‑2 checkpoints, providing a cache‑level diagnostic and a competence‑gated carrier map for arithmetic mechanisms.
arXiv:2607. 28495v1 Announce Type: new Abstract: Stage-replay diagnostics reconstruct intermediate token prefixes and treat fresh-prefill continuation as continuation from the decoder state that originally reached the prefix.
arXiv:2607. 27539v2 Announce Type: replace Abstract: Exact deletion from persistent language-model memory depends on whether a record's effect remains addressable after later computation.
arXiv:2609. 04875v1 Announce Type: cross Abstract: Long-running LLM agents are stateful: beyond the transcript they accrete compressed summaries, plaintext memory, pending tool plans, and, under every serving API, a KV cache.
The paper evaluates a deterministic supersession memory, MemStrata, for retrieval‑augmented generation (RAG) systems on real software history. Using 707 GitHub issues, the authors extracted 130 clean atomic state transitions where a single value changes from pre‑fix to post‑fix. MemStrata achieved 0.91 answer accuracy versus 0.57–0.59 for standard RAG, eliminating stale‑fact errors that RAG returned 36–38% of the time, while maintaining comparable retrieval latency.
The study investigates whether the reasoning steps a language model writes are causally responsible for its answers. Using a causal intervention method on the activation stream, the authors find that for Qwen3-4B, about 77% of stated steps are causally load‑bearing, while behavioral tests overestimate this by roughly 11 percentage points. The faithfulness of reasoning decreases with model size and depth of reasoning, especially for the smaller Qwen3-1.7B.
arXiv:2607.27617v2 Announce Type: replace Abstract: Identical language-model answers can arise from hidden states that support different future computations, so current-answer probes do not establish...
arXiv:2608. 03842v1 Announce Type: cross Abstract: When a language model fails on surface-perturbed input (typos, OCR noise, homophones), "which layer is responsible" has three natural operationalizations: where representations diverge most (sensitivity), where restoring clean activations recovers the prediction (causality), and where a small adapter can repair the damage (compensatory capacity) - and we show these three layer maps dissociate.
arXiv:2603. 06642v2 Announce Type: replace-cross Abstract: Test-Time Training (TTT) language models replace the KV-cache with fast weights updated during inference, achieving O(1) memory but suffering catastrophic failure on exact-recall tasks.
arXiv:2609.16073v1 Announce Type: cross Abstract: Retrieval-Augmented Generation (RAG) serves as the primary memory architecture for long-horizon autonomous agents. However, treating shared memory as...
arXiv:2606. 05378v1 Announce Type: new Abstract: We test whether a single screen-and-ablate recipe -- identify attention-head circuits by task-pattern selectivity, then verify by causal ablation against a matched-random null -- produces consistent mechanistic claims across model families.
arXiv:2606. 30128v1 Announce Type: new Abstract: Chain-of-thought (CoT) prompting improves LLM reasoning, but the source is contested: do the intermediate steps help because they carry useful semantic content, or because conditioning on more tokens buys extra computation before the model commits to an answer?
The paper introduces AgentDiff, a metric that quantifies how much LLM agents’ answers differ when inputs are altered by meaning‑bearing rewrites (paraphrases, synonym substitutions) versus presentation changes (reordering, formatting, distractors). Across 68 model–benchmark–scaffold combinations involving ten LLMs and over 1,500 questions, meaning‑bearing rewrites consistently produce a roughly 20‑percentage‑point higher inconsistency rate than presentation changes, a gap that persists across severity proxies and remains significant even outside the Qwen family. Trace analysis reveals that meaning‑bearing rewrites preserve the first action but reduce thought similarity from the second step onward, extending the divergence cascade—a phenomenon termed “stealth divergence.”