StateWise is a method that trains a single large language model to maintain and update its own constraint state during multi‑turn dialogue. By fine‑tuning the model to summarize conversations into a typed state with tombstones for revoked constraints and then self‑training on trajectories that pass exact round‑trip checks, it reduces relapse and improves conversation success. The approach outperforms a compiled‑ledger baseline on relapse without requiring inference‑time scaffolding, and it is evaluated on two backbones with held‑out constraint families and unseen phrasings.
arXiv:2607. 04412v1 Announce Type: new Abstract: Reinforcement learning (RL) for non-verifiable instruction following increasingly relies on LLM judges with prompt-specific rubrics as reward signals.
arXiv:2608. 12599v1 Announce Type: new Abstract: Multi-turn dialogues let users revoke constraints as easily as impose them, but revocation does not reliably take effect: models keep enacting withdrawn requirements (occasionally beneath comments asserting their removal), a failure we call \emph{behavioral relapse}, or revocation inertia.
arXiv:2606. 25449v1 Announce Type: cross Abstract: A language model's memory can be worse than having no memory at all.
arXiv:2410. 06458v2 Announce Type: replace-cross Abstract: Instruction following is a key capability for LLMs.
arXiv:2607. 18100v1 Announce Type: new Abstract: Extended reasoning has become standard for frontier Large Language Models (LLMs), yet the trajectories these models produce remain largely uncontrollable.
The paper introduces the concept of State Integrity Collapse, where large language models remember a constraint but fail to apply it in later dialogue turns. It presents StateTrack‑Bench, a suite of controlled conversations testing numerical, format, role/persona, factual‑memory, and logical‑rule constraints across multiple model snapshots and dialogue depths. The study finds that while strict joint retention can be high, failures are largely due to behavioral drift rather than simple forgetting, and that reliability must be assessed over long‑horizon dialogues rather than single‑turn instruction following.
A language model's memory can be worse than having no memory at all. Give a model a memory that kept a wrong conclusion but dropped the work behind it, and it emits that stale value as a confident answer; give the same model an empty memory and it abstains.
DIAL (Distilled Instruction-Adherence Levels) is a steerable language model that offers fine‑grained numerical control over how closely a model follows general instructions, such as varying the length of a response. It achieves this by using classifier‑free guidance with a frozen LLM to train a lightweight steering adapter that computes a closed‑form edit based on a classifier over hidden states, allowing adaptive steering magnitude per token while preserving coherence. The approach is efficient, scalable across diverse instructions, and is evaluated on multiple LLM families, showing improved concept expression and fluency, with a new dataset SteerIF providing instruction‑only training data. whyItMatters":"The paper presents a method that simultaneously delivers fine‑grained control, efficiency, and scalability for instruction adherence in language models, addressing a key limitation of current prompting techniques."
arXiv:2609.26094v1 Announce Type: new Abstract: Multi-step causal reasoning requires chaining inferences where each step constrains the next. An early error propagates silently, and a correct answer...
PROOF-Gen is a method that improves distillation of tool‑calling models by recovering successful trajectories from teacher failures. It uses per‑scenario prompt optimization to generate corrective guidance that steers the teacher to a passing trajectory, then removes this guidance before training so the student learns from clean demonstrations. On τ2‑bench, PROOF-Gen recovers 93% of failed scenarios, boosting Qwen3‑4B‑Instruct‑2507’s Pass^1 from 0.132 to 0.529 and improving Gemma 4 E4B‑it by 7.2pp on BFCL v4 multi‑turn, while also raising deployed on‑device model performance by up to 5.0pp across response‑quality metrics.
arXiv:2605.27186v2 Announce Type: replace Abstract: Large language models often solve tasks from a fully specified prompt but degrade when the same requirements unfold over multiple turns, known as t...
CounterRoute is an online reinforcement‑learning framework that jointly learns how to route a language model’s reasoning between a ‘think’ and a ‘direct answer’ mode, using a single shared policy derived from a dual‑mode checkpoint. It employs counterfactual rollouts to credit routing decisions and a curriculum that starts with forced dual‑mode rollouts before shifting to self‑routed updates, achieving better accuracy‑efficiency trade‑offs across nine benchmarks. The method reduces generated tokens by up to 51% on Qwen3‑8B while improving macro‑average accuracy, and its routing strategy generalizes to unseen coding, science, and commonsense tasks.