Did the Feeling Expire? Tracking Stale and Persistent Emotions Before Forecasting Them

Language models write text that sounds empathetic, yet they are poor at predicting what a conversation partner is actually feeling but not saying. This proposal argues that part of that failure is a bookkeeping problem. In a real conversation, feelings are rarely retracted out loud. A user who was anxious in turn two and calm by turn eight never says "my anxiety is now resolved." A friend who was hurt in turn three and is being politely agreeable in turn nine may still be hurt. To forecast the next unspoken emotion, a model has to decide, for each earlier emotional cue, whether that feeling has expired or is still active under the surface. We propose to test whether a short, structured reasoning step that makes exactly this decision explicit, before the model forecasts, closes a measurable part of the gap. We also propose a benchmark built to show the two opposite ways a model can get this wrong.

The opportunity comes from putting two findings side by side. The task of forecasting a listener's unexpressed emotion from dialogue history is hard. A fine-tuned Llama-2 reaches only 27.7% F1 excluding the neutral class on reconstructed MEmoR, against 10.0% for random guessing, and full history helps compared with the last utterance alone (Forecasting Implicit Emotions Elicited in Conversations). Separately, work on agent memory shows that models given a complete history still act on superseded values, and that free-form reasoning can make this worse because the model uses its reasoning text to justify the salient old value (Can Agent Memory Systems Track Evolving State?). No one has asked whether affective forecasting suffers the same "stale state" failure. Emotions also add a twist that facts do not: supersession is implicit, and persistence is often actively hidden by polite surface talk.

Core Claim

Errors in forecasting implicit emotion over multi-turn dialogue split into two opposite kinds. In stale anchoring, the model forecasts from a salient earlier emotion that the conversation has since resolved. In premature dropping, the model forecasts from the polite or neutral surface of recent turns and forgets an earlier emotion that is still active. Existing fixes each address only one side. Recency truncation reduces stale anchoring and increases premature dropping. Unstructured chain-of-thought (asking the model to "think step by step" before answering) can deepen stale anchoring. A reasoning template that first lists affective episodes with their turn of origin and then asks the model to make an explicit persist-or-expire judgment for each, citing the later turn that supports it, should reduce both error types at once. The new element is this two-sided judgment. That separates the proposal from the chronological "trace-then-resolve" wrapper of Can Agent Memory Systems Track Evolving State?, which assumes that updates are stated outright and that the newest value always wins. It also separates it from the static common-ground inferences of Reflect, Not Reflex, which infer the speaker's and responder's emotions once, with no notion of expiry.

Research Program

The first component is a diagnostic benchmark, which we call affective drift probes. It is built on reconstructed MEmoR and DailyDialog, the two datasets used by Forecasting Implicit Emotions Elicited in Conversations.

The natural subset keeps IEFC items (predicting the listener's implicit emotion from the partner's preceding utterance and the history) where the gold label differs from an emotion the listener showed earlier in the same dialogue. It also keeps items where the gold label matches an earlier non-neutral emotion despite neutral-looking intervening turns. This subset measures the phenomenon as it actually occurs.

A controlled subset gives the causal handle. Annotators minimally edit dialogues in one of two ways. A resolution edit inserts or rewrites intervening turns so that an earlier negative emotion is plausibly resolved. A masking edit rewrites intervening turns into polite, face-saving talk while the grievance plausibly persists; this borrows the white-lie structure studied in TactfulToM. Each edited item gets new gold labels from three annotators. We keep only items with majority agreement and report inter-annotator agreement.

Because both corpora are dominated by neutral turns (80.3% in reconstructed MEmoR, 84.6% in DailyDialog), we report F1 without the neutral class. We also report two error rates. The stale-anchor rate is the fraction of resolution items where the prediction equals the expired emotion. The premature-drop rate is the fraction of masking items predicted as neutral or positive.

The second component is the reasoning template, called persist-or-expire tracing. The model first writes an episode list: each earlier affective cue, whose it is, and the turn it appeared. For each episode, it then states "active," "fading," or "resolved," quoting the later turn that justifies the judgment. It is also told that absence of a restatement is not evidence of resolution. Only then does it forecast.

As an optional addition, borrowed from the fact-lifespan idea of Mitigating Temporal Misalignment by Discarding Outdated Facts, the model also gives a coarse expected lifespan for each episode ("within this exchange," "across the conversation," "beyond it"). This is used only to discount confidence. That paper reduced calibration error on world facts by 50–60% with duration-based discounting. We test whether the analogous discount calibrates affective forecasts, but we treat it as secondary.

The third component asks how in-context demonstrations interact with drift. The in-context learning literature shows that few-shot examples are compressed into a compact internal "task vector" that can override the surface text (In-Context Learning Creates Task Vectors). A worry follows: demonstrations whose emotions never change could install a "the emotion persists" rule. We therefore vary the demonstration set at matched size: all-persistent, all-reversing, or balanced. We measure the effect on the two error rates. As a light mechanistic check on an open-weight model, we patch the intermediate-layer activation taken from a reversal-heavy demonstration set into a query-only pass, following that paper's recipe. We then see whether it moves the stale-anchor rate. We make no claim that a single vector captures the full task. The check only asks whether demonstration content shifts the persist/expire bias.

Models are an open-weight family at two sizes, for example a 7–8B and a 70B Llama. The size contrast matters because Reflect found that prompted deliberation helped a 175B model but that a 440M model needed fine-tuning. We add one strong API model. For the small model we also fine-tune on template traces written by the large model, which mirrors Reflect's fine-tuning route.

Baselines

The naive reference points are the random baseline (10.0% F1 without neutral on reconstructed MEmoR) and direct zero-shot prompting with full history. Direct prompting tells us how much of the gap exists before any reasoning is added.

The strongest reported supervised result is fine-tuned Llama-2 with full history at 27.7% F1 without neutral on IEFC in reconstructed MEmoR (Forecasting Implicit Emotions Elicited in Conversations). We reproduce it on the same splits, then evaluate it on both probe subsets. The prediction is that it performs worst on the controlled items. Beating it means that the prompted large model, or the template-fine-tuned small model, exceeds our reproduced number on the natural IEFC test set. It also means lower stale-anchor and premature-drop rates on the probes.

The strongest structural competitor is the StateMemWrapper-style prompt from Can Agent Memory Systems Track Evolving State?, transplanted verbatim: reconstruct the chronological chain of the listener's emotion, then report the latest value. On factual state tracking it lifted current-state accuracy by +32 to +67 points across six memory backends. The full StateMem system raised accuracy from 0.205 to 0.363 on DeepSeek-V4-Flash. There is no reported result on Lacuna for either method on emotion forecasting. This is the baseline most likely to erase our contribution, because it already fixes stale anchoring. Our template must beat it on masking items without losing to it on resolution items.

We also include two more competitors. The first is Reflect-style common-ground prompting, which infers speaker and responder emotion before answering (Reflect, Not Reflex). The second is unstructured chain-of-thought.

Two ablations isolate the key ingredient. The first keeps the episode list but removes the persist/expire judgment. The second keeps a single overall "is the earlier emotion still active?" question but removes the episode list.

For fairness, every prompted method is tuned on the same development split with the same number of prompt variants, which we fix at eight. Direct prompting and Reflect-style prompting also get self-consistency sampling up to the token budget our template consumes, so that any gain cannot be credited to extra computation alone. All methods see identical demonstrations unless demonstration content is the variable under study.

Why not just…?

Why not just use the last few turns? If old feelings mostly expire, a recency window removes stale anchoring for free. The original IEFC study already found that full history beats the last utterance. Our masking subset is designed to show the cost directly: truncation should raise the premature-drop rate. We will sweep the window length and report both error rates. If some window matches our template on both subsets, the template is unnecessary, and we will say so.

Why not just apply the existing trace-then-resolve wrapper? This is the real threat. It already defeats the salience trap on facts, and it might carry over. Its assumption that the latest stated value supersedes earlier ones is exactly what masking items violate: the latest surface signal is polite, and the operative feeling is older. The head-to-head comparison on the two probe subsets settles this. If the wrapper matches us on masking items, our contribution shrinks to a benchmark finding, and we would report it that way.

Why not just think step by step? Unstructured reasoning gave inconsistent gains on white-lie reasoning in TactfulToM, and it lowered accuracy on state tracking in Can Agent Memory Systems Track Evolving State?. We include it at matched compute. We specifically test whether its stale-anchor rate exceeds that of direct prompting, which would reproduce the rationalization effect in the affective domain.

Why not just fine-tune a classifier with more data? Fine-tuning with full history is exactly the 27.7% baseline. Nothing in that training signal distinguishes resolved from persistent emotions. We train it and also a version fine-tuned on our template traces, which tells us whether the structure helps once it is learned into the weights.

Proposed Experiments and Decisive Checks

The main experiment runs all methods across both corpora, both probe subsets, and all model sizes. It reports F1 without neutral, the stale-anchor rate, and the premature-drop rate, with bootstrap confidence intervals over dialogues.

The decisive check is a crossover pattern. Recency truncation and the trace-then-resolve wrapper should do well on resolution items and badly on masking items. Unstructured chain-of-thought should do badly on resolution items. Persist-or-expire tracing should be at or near the best on both. If the template fails to beat the transplanted wrapper on masking items, the core claim is false.

A second check uses the two ablations. If the episode list alone carries most of the gain, the explicit expiry judgment is not the operative ingredient.

A third check varies demonstration content. If reversal-heavy demonstrations reduce stale anchoring and the patched activation reproduces part of that shift, demonstration choice becomes a cheap lever. If neither moves the error rates, the in-context route can be dropped.

Finally, we compute calibration error on predicted confidences with and without the lifespan discount. This tells us whether the duration idea transfers from world facts to feelings.